Upload 65 files

app.py

@@ -0,0 +1,362 @@
+import os, argparse
+import sys
+import gradio as gr
+# from scripts.gradio.i2v_test_application import Image2Video
+sys.path.insert(1, os.path.join(sys.path[0], 'lvdm'))
+import spaces
+
+
+import os
+import time
+from omegaconf import OmegaConf
+import torch
+from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z
+from utils.utils import instantiate_from_config
+from huggingface_hub import hf_hub_download
+from einops import repeat
+import torchvision.transforms as transforms
+from pytorch_lightning import seed_everything
+from einops import rearrange
+import cv2
+
+import torch
+print("cuda available:", torch.cuda.is_available())
+
+
+from huggingface_hub import snapshot_download
+import os
+
+
+
+def download_model():
+    REPO_ID = 'fbnnb/tc_1024'
+    filename_list = ['tc1024_4k.ckpt']
+    tar_dir = './checkpoints/tooncrafter_1024_interp_sketch/'
+    if not os.path.exists(tar_dir):
+        os.makedirs(tar_dir)
+    for filename in filename_list:
+        local_file = os.path.join(tar_dir, filename)
+        if not os.path.exists(local_file):
+            hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir=tar_dir, local_dir_use_symlinks=False)
+    print("downloaded")
+    
+
+def get_latent_z_with_hidden_states(model, videos):
+    b, c, t, h, w = videos.shape
+    x = rearrange(videos, 'b c t h w -> (b t) c h w')
+    encoder_posterior, hidden_states = model.first_stage_model.encode(x, return_hidden_states=True)
+
+    hidden_states_first_last = []
+    ### use only the first and last hidden states
+    for hid in hidden_states:
+        hid = rearrange(hid, '(b t) c h w -> b c t h w', t=t)
+        hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2)
+        hidden_states_first_last.append(hid_new)
+
+    z = model.get_first_stage_encoding(encoder_posterior).detach()
+    z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t)
+    return z, hidden_states_first_last
+
+
+
+def extract_frames(video_path):
+    # 動画ファイルを読み込む
+    cap = cv2.VideoCapture(video_path)
+    
+    frame_list = [] 
+    frame_num = 0
+    
+    while True:
+        # フレームを読み込む
+        ret, frame = cap.read()
+        if not ret:
+            break
+        
+        # フレームをリストに追加
+        frame_list.append(frame)
+        frame_num += 1
+
+    print("load video length:", len(frame_list))
+    # 動画ファイルを閉じる
+    cap.release()
+    
+    return frame_list
+
+
+resolution = '576_1024'
+resolution = (576, 1024)
+download_model()
+print("after download model")
+result_dir = "./results/"
+if not os.path.exists(result_dir):
+    os.mkdir(result_dir)
+
+#ToonCrafterModel
+ckpt_path='checkpoints/tooncrafter_1024_interp_sketch/tc1024_4k.ckpt'
+# ckpt_path="/group/40005/gzhiwang/tc1024_4k.ckpt"
+config_file='configs/inference_1024_v1.0.yaml'
+config = OmegaConf.load(config_file)
+model_config = config.pop("model", OmegaConf.create())
+model_config['params']['unet_config']['params']['use_checkpoint']=False  
+
+model = instantiate_from_config(model_config)
+assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!"
+# model = load_model_checkpoint(model, ckpt_path)
+state = torch.load(ckpt_path, map_location='cpu')
+if 'state_dict' in state:
+    state = state['state_dict']
+if 'module' in state:
+    state = state['module']
+
+missing, unexpected = model.load_state_dict(state, strict=False)
+print("missing:", missing)
+print("unexpected:", unexpected)
+model.eval()
+
+# cn_model.load_state_dict(load_state_dict(cn_ckpt_path, location='cpu'))  
+# cn_model.eval()
+
+# model.control_model = cn_model    
+# model_list.append(model)
+
+save_fps = 8
+print("resolution:", resolution)
+print("init done.")
+
+def transpose_if_needed(tensor):
+    h = tensor.shape[-2]
+    w = tensor.shape[-1]
+    if h > w:
+        tensor = tensor.permute(0, 2, 1)
+    return tensor
+
+def untranspose(tensor):
+    ndim = tensor.ndim
+    return tensor.transpose(ndim-1, ndim-2)
+
+@spaces.GPU(duration=200)
+def get_image(image, sketch, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123, control_scale=0.6):
+    print("enter fn")
+    # control_frames = extract_frames(frame_guides)
+    print("extract frames")
+    seed_everything(seed)
+    transform = transforms.Compose([
+        transforms.Resize(min(resolution)),
+        transforms.CenterCrop(resolution),
+        ])
+    print("before empty cache")
+    torch.cuda.empty_cache()
+    print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time())))
+    start = time.time()
+    gpu_id=0
+    if steps > 60:
+        steps = 60
+
+    global model 
+    # model = model_list[gpu_id]
+    model = model.cuda()
+    
+    batch_size=1
+    channels = model.model.diffusion_model.out_channels
+    frames = model.temporal_length
+    h, w = resolution[0] // 8, resolution[1] // 8
+    noise_shape = [batch_size, channels, frames, h, w]
+
+    # text cond
+    transposed = False 
+    with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.float16):
+        text_emb = model.get_learned_conditioning([prompt])
+        print("before control")
+        #control cond
+        # if frame_guides is not None:
+        #     cn_videos = []
+        #     for frame in control_frames:
+        #         frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+        #         frame = cv2.bitwise_not(frame)
+        #         cn_tensor = torch.from_numpy(frame).unsqueeze(2).permute(2, 0, 1).float().to(model.device)
+                
+        #         #cn_tensor = (cn_tensor / 255. - 0.5) * 2
+        #         cn_tensor = ( cn_tensor/255.0 )
+        #         cn_tensor = transpose_if_needed(cn_tensor)
+        #         cn_tensor_resized = transform(cn_tensor) #3,h,w
+
+        #         cn_video = cn_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw
+        #         cn_videos.append(cn_video)
+            
+        #     cn_videos = torch.cat(cn_videos, dim=2)
+        #     if cn_videos.shape[2] > frames:
+        #         idxs = []
+        #         for i in range(frames):
+        #             index = int((i + 0.5) * cn_videos.shape[2] / frames)
+        #             idxs.append(min(index, cn_videos.shape[2] - 1))
+        #         cn_videos = cn_videos[:, :, idxs, :, :]
+        #         print("cn_videos.shape after slicing", cn_videos.shape)
+        #     model_list = []
+        #     for model in model_list:
+        #         model.control_scale = control_scale
+        #         model_list.append(model)
+            
+        # else:
+        cn_videos = None
+
+        print("image cond")
+
+        # img cond
+        img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
+        input_h, input_w = img_tensor.shape[1:]
+        img_tensor = (img_tensor / 255. - 0.5) * 2
+        img_tensor = transpose_if_needed(img_tensor)
+        
+        image_tensor_resized = transform(img_tensor) #3,h,w
+        videos = image_tensor_resized.unsqueeze(0).unsqueeze(2) # bc1hw
+        print("get latent z")
+        # z = get_latent_z(model, videos) #bc,1,hw
+        videos = repeat(videos, 'b c t h w -> b c (repeat t) h w', repeat=frames//2)
+
+        if sketch is not None:
+            img_tensor2 = torch.from_numpy(sketch).permute(2, 0, 1).float().to(model.device)
+            img_tensor2 = (img_tensor2 / 255. - 0.5) * 2
+            img_tensor2 = transpose_if_needed(img_tensor2)
+            image_tensor_resized2 = transform(img_tensor2) #3,h,w
+            videos2 = image_tensor_resized2.unsqueeze(0).unsqueeze(2) # bchw
+            videos2 = repeat(videos2, 'b c t h w -> b c (repeat t) h w', repeat=frames//2)
+            
+            videos = torch.cat([videos, videos2], dim=2)
+        else:
+            videos = torch.cat([videos, videos], dim=2)
+            
+        z, hs = get_latent_z_with_hidden_states(model, videos)
+
+        img_tensor_repeat = torch.zeros_like(z)
+
+        img_tensor_repeat[:,:,:1,:,:] = z[:,:,:1,:,:]
+        img_tensor_repeat[:,:,-1:,:,:] = z[:,:,-1:,:,:]
+
+        print("image embedder")
+        cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc
+        img_emb = model.image_proj_model(cond_images)
+
+        imtext_cond = torch.cat([text_emb, img_emb], dim=1)
+
+        fs = torch.tensor([fs], dtype=torch.long, device=model.device)
+        # print("cn videos:",cn_videos.shape, "img emb:", img_emb.shape)
+        cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat], "control_cond": cn_videos}
+
+        print("before sample loop")
+        ## inference
+        batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale, hs=hs)
+
+        ## remove the last frame
+        # if image2 is None:
+        batch_samples = batch_samples[:,:,:,:-1,...]
+        ## b,samples,c,t,h,w
+        prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt
+        prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str
+        prompt_str=prompt_str[:40]
+        if len(prompt_str) == 0:
+            prompt_str = 'empty_prompt'
+
+    global result_dir
+    global save_fps
+    if input_h > input_w:
+        batch_samples = untranspose(batch_samples)
+        
+    save_videos(batch_samples, result_dir, filenames=[prompt_str], fps=save_fps)
+    print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds")
+    model = model.cpu()
+    saved_result_dir = os.path.join(result_dir, f"{prompt_str}.mp4")
+    print("result saved to:", saved_result_dir)
+    return saved_result_dir
+
+
+    # @spaces.GPU
+
+    
+
+# i2v_examples_interp_1024 = [
+#     ['prompts/1024_interp/frame_000000.jpg', 'prompts/1024_interp/frame_000041.jpg', 'a cat is eating', 50, 7.5, 1.0, 10, 123]
+# ]
+
+i2v_examples_interp_1024 = [
+    ['prompts/1024_interp/hall_first.jpg', 'prompts/1024_interp/hall_sketch.jpg', 
+     'At the start, a still image of a wooden hallway with arched arches, doors, and various furniture. The scene then transitions to an animated version of the hallway, showcasing more details like a bookshelf and a window.', 
+     50, 7.5, 1.0, 10, 123]
+]
+
+
+
+
+def dynamicrafter_demo(result_dir='./tmp/', res=1024):
+    if res == 1024:
+        resolution = '576_1024'
+        css = """#input_img {max-width: 1024px !important} #output_vid {max-width: 1024px; max-height:576px}"""
+    elif res == 512:
+        resolution = '320_512'
+        css = """#input_img {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px} #input_img2 {max-width: 512px !important} #output_vid {max-width: 512px; max-height: 320px}"""
+    elif res == 256:
+        resolution = '256_256'
+        css = """#input_img {max-width: 256px !important} #output_vid {max-width: 256px; max-height: 256px}"""
+    else:
+        raise NotImplementedError(f"Unsupported resolution: {res}")
+    # image2video = Image2Video(result_dir, resolution=resolution)
+    with gr.Blocks(analytics_enabled=False, css=css) as dynamicrafter_iface:
+
+
+
+        with gr.Tab(label='ToonCrafter_576x1024'):
+            with gr.Column():
+                with gr.Row():
+                    with gr.Column():
+                        with gr.Row():
+                            i2v_input_image = gr.Image(label="Input Image1",elem_id="input_img")
+                            # frame_guides = gr.Video(label="Input Guidance",elem_id="input_guidance", autoplay=True,show_share_button=True)
+                        with gr.Row():
+                            i2v_input_text = gr.Text(label='Prompts')
+                        with gr.Row():
+                            i2v_seed = gr.Slider(label='Random Seed', minimum=0, maximum=50000, step=1, value=123)
+                            i2v_eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='ETA', value=1.0, elem_id="i2v_eta")
+                            i2v_cfg_scale = gr.Slider(minimum=1.0, maximum=15.0, step=0.5, label='CFG Scale', value=7.5, elem_id="i2v_cfg_scale")
+                        with gr.Row():
+                            i2v_steps = gr.Slider(minimum=1, maximum=60, step=1, elem_id="i2v_steps", label="Sampling steps", value=50)
+                            i2v_motion = gr.Slider(minimum=5, maximum=30, step=1, elem_id="i2v_motion", label="FPS", value=10)
+                            control_scale = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, elem_id="i2v_ctrl_scale", label="control_scale", value=0.6)
+                        i2v_end_btn = gr.Button("Generate")
+                    with gr.Column():
+                        with gr.Row():
+                            i2v_input_sketch = gr.Image(label="Input Image2",elem_id="input_img2")
+                        with gr.Row():
+                            i2v_output_video = gr.Video(label="Generated Video",elem_id="output_vid",autoplay=True,show_share_button=True)
+
+                gr.Examples(examples=i2v_examples_interp_1024,
+                            inputs=[i2v_input_image, i2v_input_sketch, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, control_scale],
+                            outputs=[i2v_output_video],
+                            fn = get_image,
+                            cache_examples=False,
+                )
+            i2v_end_btn.click(inputs=[i2v_input_image, i2v_input_sketch, i2v_input_text, i2v_steps, i2v_cfg_scale, i2v_eta, i2v_motion, i2v_seed, control_scale],
+                            outputs=[i2v_output_video],
+                            fn = get_image
+            )
+
+
+    return dynamicrafter_iface
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    return parser
+    
+
+if __name__ == "__main__":
+    parser = get_parser()
+    args = parser.parse_args()
+
+    result_dir = os.path.join('./', 'results')
+    dynamicrafter_iface = dynamicrafter_demo(result_dir)
+    dynamicrafter_iface.queue(max_size=12)
+    print("launching...")
+    # dynamicrafter_iface.launch(max_threads=1, share=True)
+    
+    dynamicrafter_iface.launch(server_name='0.0.0.0', server_port=12345)
+    # dynamicrafter_iface.launch()
+    # print("launched...")

configs/inference_1024_v1.0.yaml

@@ -0,0 +1,103 @@
+model:
+  target: lvdm.models.ddpm3d.LatentVisualDiffusion
+  params:
+    rescale_betas_zero_snr: True
+    parameterization: "v"
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: False
+    conditioning_key: hybrid
+    image_size: [72, 128]
+    channels: 4
+    scale_by_std: False
+    scale_factor: 0.18215
+    use_ema: False
+    uncond_type: 'empty_seq'
+    use_dynamic_rescale: true
+    base_scale: 0.7
+    fps_condition_type: 'fps'
+    perframe_ae: True
+    loop_video: true
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        dropout: 0.1
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: True
+        temporal_conv: True
+        temporal_attention: True
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: False
+        temporal_length: 16
+        addition_attention: true
+        image_cross_attention: true
+        default_fs: 24
+        fs_condition: true
+
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL_Dualref
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: "penultimate"
+
+    img_cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2
+      params:
+        freeze: true
+    
+    image_proj_stage_config:
+      target: lvdm.modules.encoders.resampler.Resampler
+      params:
+        dim: 1024
+        depth: 4
+        dim_head: 64
+        heads: 12
+        num_queries: 16
+        embedding_dim: 1280
+        output_dim: 1024
+        ff_mult: 4
+        video_length: 16

configs/inference_512_v1.0.yaml

@@ -0,0 +1,103 @@
+model:
+  target: lvdm.models.ddpm3d.LatentVisualDiffusion
+  params:
+    rescale_betas_zero_snr: True
+    parameterization: "v"
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: False
+    conditioning_key: hybrid
+    image_size: [40, 64]
+    channels: 4
+    scale_by_std: False
+    scale_factor: 0.18215
+    use_ema: False
+    uncond_type: 'empty_seq'
+    use_dynamic_rescale: true
+    base_scale: 0.7
+    fps_condition_type: 'fps'
+    perframe_ae: True
+    loop_video: true
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        dropout: 0.1
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: True
+        temporal_conv: True
+        temporal_attention: True
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: False
+        temporal_length: 16
+        addition_attention: true
+        image_cross_attention: true
+        default_fs: 24
+        fs_condition: true
+
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL_Dualref
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: "penultimate"
+
+    img_cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2
+      params:
+        freeze: true
+    
+    image_proj_stage_config:
+      target: lvdm.modules.encoders.resampler.Resampler
+      params:
+        dim: 1024
+        depth: 4
+        dim_head: 64
+        heads: 12
+        num_queries: 16
+        embedding_dim: 1280
+        output_dim: 1024
+        ff_mult: 4
+        video_length: 16

configs/training_1024_v1.0/config.yaml

@@ -0,0 +1,166 @@
+model:
+  pretrained_checkpoint: checkpoints/dynamicrafter_1024_v1/model.ckpt
+  base_learning_rate: 1.0e-05
+  scale_lr: False
+  target: lvdm.models.ddpm3d.LatentVisualDiffusion
+  params:
+    rescale_betas_zero_snr: True
+    parameterization: "v"
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: False
+    image_proj_model_trainable: True
+    conditioning_key: hybrid
+    image_size: [72, 128]
+    channels: 4
+    scale_by_std: False
+    scale_factor: 0.18215
+    use_ema: False
+    uncond_prob: 0.05
+    uncond_type: 'empty_seq'
+    rand_cond_frame: true
+    use_dynamic_rescale: true
+    base_scale: 0.3
+    fps_condition_type: 'fps'
+    perframe_ae: True
+
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        dropout: 0.1
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: True
+        temporal_conv: True
+        temporal_attention: True
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: False
+        temporal_length: 16
+        addition_attention: true
+        image_cross_attention: true
+        default_fs: 10
+        fs_condition: true
+
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: "penultimate"
+
+    img_cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2
+      params:
+        freeze: true
+    
+    image_proj_stage_config:
+      target: lvdm.modules.encoders.resampler.Resampler
+      params:
+        dim: 1024
+        depth: 4
+        dim_head: 64
+        heads: 12
+        num_queries: 16
+        embedding_dim: 1280
+        output_dim: 1024
+        ff_mult: 4
+        video_length: 16
+
+data:
+  target: utils_data.DataModuleFromConfig
+  params:
+    batch_size: 1
+    num_workers: 12
+    wrap: false
+    train:
+      target: lvdm.data.webvid.WebVid
+      params:
+        data_dir: <WebVid10M DATA>
+        meta_path: <.csv FILE>
+        video_length: 16
+        frame_stride: 6
+        load_raw_resolution: true
+        resolution: [576, 1024]
+        spatial_transform: resize_center_crop
+        random_fs: true  ## if true, we uniformly sample fs with max_fs=frame_stride (above)
+
+lightning:
+  precision: 16
+  # strategy: deepspeed_stage_2
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2
+    max_steps: 100000
+    # logger
+    log_every_n_steps: 50
+    # val
+    val_check_interval: 0.5
+    gradient_clip_algorithm: 'norm'
+    gradient_clip_val: 0.5
+  callbacks:
+    model_checkpoint:
+      target: pytorch_lightning.callbacks.ModelCheckpoint
+      params:
+        every_n_train_steps: 9000 #1000
+        filename: "{epoch}-{step}"
+        save_weights_only: True
+    metrics_over_trainsteps_checkpoint:
+      target: pytorch_lightning.callbacks.ModelCheckpoint
+      params:
+        filename: '{epoch}-{step}'
+        save_weights_only: True
+        every_n_train_steps: 10000 #20000 # 3s/step*2w=
+    batch_logger:
+      target: callbacks.ImageLogger
+      params:
+        batch_frequency: 500
+        to_local: False
+        max_images: 8
+        log_images_kwargs:
+          ddim_steps: 50
+          unconditional_guidance_scale: 7.5
+          timestep_spacing: uniform_trailing
+          guidance_rescale: 0.7

configs/training_1024_v1.0/run.sh

@@ -0,0 +1,37 @@
+# NCCL configuration
+# export NCCL_DEBUG=INFO
+# export NCCL_IB_DISABLE=0
+# export NCCL_IB_GID_INDEX=3
+# export NCCL_NET_GDR_LEVEL=3
+# export NCCL_TOPO_FILE=/tmp/topo.txt
+
+# args
+name="training_1024_v1.0"
+config_file=configs/${name}/config.yaml
+
+# save root dir for logs, checkpoints, tensorboard record, etc.
+save_root="<YOUR_SAVE_ROOT_DIR>"
+
+mkdir -p $save_root/$name
+
+## run
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \
+--nproc_per_node=$HOST_GPU_NUM --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \
+./main/trainer.py \
+--base $config_file \
+--train \
+--name $name \
+--logdir $save_root \
+--devices $HOST_GPU_NUM \
+lightning.trainer.num_nodes=1
+
+## debugging
+# CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch \
+# --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \
+# ./main/trainer.py \
+# --base $config_file \
+# --train \
+# --name $name \
+# --logdir $save_root \
+# --devices 4 \
+# lightning.trainer.num_nodes=1

configs/training_512_v1.0/config.yaml

@@ -0,0 +1,166 @@
+model:
+  pretrained_checkpoint: checkpoints/dynamicrafter_512_v1/model.ckpt
+  base_learning_rate: 1.0e-05
+  scale_lr: False
+  target: lvdm.models.ddpm3d.LatentVisualDiffusion
+  params:
+    rescale_betas_zero_snr: True
+    parameterization: "v"
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: False
+    image_proj_model_trainable: True
+    conditioning_key: hybrid
+    image_size: [40, 64]
+    channels: 4
+    scale_by_std: False
+    scale_factor: 0.18215
+    use_ema: False
+    uncond_prob: 0.05
+    uncond_type: 'empty_seq'
+    rand_cond_frame: true
+    use_dynamic_rescale: true
+    base_scale: 0.7
+    fps_condition_type: 'fps'
+    perframe_ae: True
+
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        dropout: 0.1
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: True
+        temporal_conv: True
+        temporal_attention: True
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: False
+        temporal_length: 16
+        addition_attention: true
+        image_cross_attention: true
+        default_fs: 10
+        fs_condition: true
+
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: "penultimate"
+
+    img_cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2
+      params:
+        freeze: true
+    
+    image_proj_stage_config:
+      target: lvdm.modules.encoders.resampler.Resampler
+      params:
+        dim: 1024
+        depth: 4
+        dim_head: 64
+        heads: 12
+        num_queries: 16
+        embedding_dim: 1280
+        output_dim: 1024
+        ff_mult: 4
+        video_length: 16
+
+data:
+  target: utils_data.DataModuleFromConfig
+  params:
+    batch_size: 2
+    num_workers: 12
+    wrap: false
+    train:
+      target: lvdm.data.webvid.WebVid
+      params:
+        data_dir: <WebVid10M DATA>
+        meta_path: <.csv FILE>
+        video_length: 16
+        frame_stride: 6
+        load_raw_resolution: true
+        resolution: [320, 512]
+        spatial_transform: resize_center_crop
+        random_fs: true  ## if true, we uniformly sample fs with max_fs=frame_stride (above)
+
+lightning:
+  precision: 16
+  # strategy: deepspeed_stage_2
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2
+    max_steps: 100000
+    # logger
+    log_every_n_steps: 50
+    # val
+    val_check_interval: 0.5
+    gradient_clip_algorithm: 'norm'
+    gradient_clip_val: 0.5
+  callbacks:
+    model_checkpoint:
+      target: pytorch_lightning.callbacks.ModelCheckpoint
+      params:
+        every_n_train_steps: 9000 #1000
+        filename: "{epoch}-{step}"
+        save_weights_only: True
+    metrics_over_trainsteps_checkpoint:
+      target: pytorch_lightning.callbacks.ModelCheckpoint
+      params:
+        filename: '{epoch}-{step}'
+        save_weights_only: True
+        every_n_train_steps: 10000 #20000 # 3s/step*2w=
+    batch_logger:
+      target: callbacks.ImageLogger
+      params:
+        batch_frequency: 500
+        to_local: False
+        max_images: 8
+        log_images_kwargs:
+          ddim_steps: 50
+          unconditional_guidance_scale: 7.5
+          timestep_spacing: uniform_trailing
+          guidance_rescale: 0.7

configs/training_512_v1.0/run.sh

@@ -0,0 +1,37 @@
+# NCCL configuration
+# export NCCL_DEBUG=INFO
+# export NCCL_IB_DISABLE=0
+# export NCCL_IB_GID_INDEX=3
+# export NCCL_NET_GDR_LEVEL=3
+# export NCCL_TOPO_FILE=/tmp/topo.txt
+
+# args
+name="training_512_v1.0"
+config_file=configs/${name}/config.yaml
+
+# save root dir for logs, checkpoints, tensorboard record, etc.
+save_root="<YOUR_SAVE_ROOT_DIR>"
+
+mkdir -p $save_root/$name
+
+## run
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \
+--nproc_per_node=$HOST_GPU_NUM --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \
+./main/trainer.py \
+--base $config_file \
+--train \
+--name $name \
+--logdir $save_root \
+--devices $HOST_GPU_NUM \
+lightning.trainer.num_nodes=1
+
+## debugging
+# CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch \
+# --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12352 --node_rank=0 \
+# ./main/trainer.py \
+# --base $config_file \
+# --train \
+# --name $name \
+# --logdir $save_root \
+# --devices 4 \
+# lightning.trainer.num_nodes=1

lvdm/basics.py

@@ -0,0 +1,100 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+import torch.nn as nn
+from utils.utils import instantiate_from_config
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def nonlinearity(type='silu'):
+    if type == 'silu':
+        return nn.SiLU()
+    elif type == 'leaky_relu':
+        return nn.LeakyReLU()
+
+
+class GroupNormSpecific(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+
+def normalization(channels, num_groups=32):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNormSpecific(num_groups, channels)
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}

lvdm/common.py

@@ -0,0 +1,94 @@
+import math
+from inspect import isfunction
+import torch
+from torch import nn
+import torch.distributed as dist
+
+
+def gather_data(data, return_np=True):
+    ''' gather data from multiple processes to one list '''
+    data_list = [torch.zeros_like(data) for _ in range(dist.get_world_size())]
+    dist.all_gather(data_list, data)  # gather not supported with NCCL
+    if return_np:
+        data_list = [data.cpu().numpy() for data in data_list]
+    return data_list
+
+def autocast(f):
+    def do_autocast(*args, **kwargs):
+        with torch.cuda.amp.autocast(enabled=True,
+                                     dtype=torch.get_autocast_gpu_dtype(),
+                                     cache_enabled=torch.is_autocast_cache_enabled()):
+            return f(*args, **kwargs)
+    return do_autocast
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+def exists(val):
+    return val is not None
+
+def identity(*args, **kwargs):
+    return nn.Identity()
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+def shape_to_str(x):
+    shape_str = "x".join([str(x) for x in x.shape])
+    return shape_str
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+ckpt = torch.utils.checkpoint.checkpoint
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        return ckpt(func, *inputs, use_reentrant=False)
+    else:
+        return func(*inputs)

lvdm/data/base.py

@@ -0,0 +1,23 @@
+from abc import abstractmethod
+from torch.utils.data import IterableDataset
+
+
+class Txt2ImgIterableBaseDataset(IterableDataset):
+    '''
+    Define an interface to make the IterableDatasets for text2img data chainable
+    '''
+    def __init__(self, num_records=0, valid_ids=None, size=256):
+        super().__init__()
+        self.num_records = num_records
+        self.valid_ids = valid_ids
+        self.sample_ids = valid_ids
+        self.size = size
+
+        print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
+
+    def __len__(self):
+        return self.num_records
+
+    @abstractmethod
+    def __iter__(self):
+        pass

lvdm/data/webvid.py

@@ -0,0 +1,202 @@
+import os
+import random
+from tqdm import tqdm
+import pandas as pd
+from decord import VideoReader, cpu
+
+import torch
+from torch.utils.data import Dataset
+from torch.utils.data import DataLoader
+from torchvision import transforms
+
+
+class WebVid(Dataset):
+    """
+    WebVid Dataset.
+    Assumes webvid data is structured as follows.
+    Webvid/
+        videos/
+            000001_000050/      ($page_dir)
+                1.mp4           (videoid.mp4)
+                ...
+                5000.mp4
+            ...
+    """
+    def __init__(self,
+                 meta_path,
+                 data_dir,
+                 subsample=None,
+                 video_length=16,
+                 resolution=[256, 512],
+                 frame_stride=1,
+                 frame_stride_min=1,
+                 spatial_transform=None,
+                 crop_resolution=None,
+                 fps_max=None,
+                 load_raw_resolution=False,
+                 fixed_fps=None,
+                 random_fs=False,
+                 ):
+        self.meta_path = meta_path
+        self.data_dir = data_dir
+        self.subsample = subsample
+        self.video_length = video_length
+        self.resolution = [resolution, resolution] if isinstance(resolution, int) else resolution
+        self.fps_max = fps_max
+        self.frame_stride = frame_stride
+        self.frame_stride_min = frame_stride_min
+        self.fixed_fps = fixed_fps
+        self.load_raw_resolution = load_raw_resolution
+        self.random_fs = random_fs
+        self._load_metadata()
+        if spatial_transform is not None:
+            if spatial_transform == "random_crop":
+                self.spatial_transform = transforms.RandomCrop(crop_resolution)
+            elif spatial_transform == "center_crop":
+                self.spatial_transform = transforms.Compose([
+                    transforms.CenterCrop(resolution),
+                    ])            
+            elif spatial_transform == "resize_center_crop":
+                # assert(self.resolution[0] == self.resolution[1])
+                self.spatial_transform = transforms.Compose([
+                    transforms.Resize(min(self.resolution)),
+                    transforms.CenterCrop(self.resolution),
+                    ])
+            elif spatial_transform == "resize":
+                self.spatial_transform = transforms.Resize(self.resolution)
+            else:
+                raise NotImplementedError
+        else:
+            self.spatial_transform = None
+                
+    def _load_metadata(self):
+        metadata = pd.read_csv(self.meta_path)
+        print(f'>>> {len(metadata)} data samples loaded.')
+        if self.subsample is not None:
+            metadata = metadata.sample(self.subsample, random_state=0)
+   
+        metadata['caption'] = metadata['name']
+        del metadata['name']
+        self.metadata = metadata
+        self.metadata.dropna(inplace=True)
+
+    def _get_video_path(self, sample):
+        rel_video_fp = os.path.join(sample['page_dir'], str(sample['videoid']) + '.mp4')
+        full_video_fp = os.path.join(self.data_dir, 'videos', rel_video_fp)
+        return full_video_fp
+    
+    def __getitem__(self, index):
+        if self.random_fs:
+            frame_stride = random.randint(self.frame_stride_min, self.frame_stride)
+        else:
+            frame_stride = self.frame_stride
+
+        ## get frames until success
+        while True:
+            index = index % len(self.metadata)
+            sample = self.metadata.iloc[index]
+            video_path = self._get_video_path(sample)
+            ## video_path should be in the format of "....../WebVid/videos/$page_dir/$videoid.mp4"
+            caption = sample['caption']
+
+            try:
+                if self.load_raw_resolution:
+                    video_reader = VideoReader(video_path, ctx=cpu(0))
+                else:
+                    video_reader = VideoReader(video_path, ctx=cpu(0), width=530, height=300)
+                if len(video_reader) < self.video_length:
+                    print(f"video length ({len(video_reader)}) is smaller than target length({self.video_length})")
+                    index += 1
+                    continue
+                else:
+                    pass
+            except:
+                index += 1
+                print(f"Load video failed! path = {video_path}")
+                continue
+            
+            fps_ori = video_reader.get_avg_fps()
+            if self.fixed_fps is not None:
+                frame_stride = int(frame_stride * (1.0 * fps_ori / self.fixed_fps))
+
+            ## to avoid extreme cases when fixed_fps is used
+            frame_stride = max(frame_stride, 1)
+            
+            ## get valid range (adapting case by case)
+            required_frame_num = frame_stride * (self.video_length-1) + 1
+            frame_num = len(video_reader)
+            if frame_num < required_frame_num:
+                ## drop extra samples if fixed fps is required
+                if self.fixed_fps is not None and frame_num < required_frame_num * 0.5:
+                    index += 1
+                    continue
+                else:
+                    frame_stride = frame_num // self.video_length
+                    required_frame_num = frame_stride * (self.video_length-1) + 1
+
+            ## select a random clip
+            random_range = frame_num - required_frame_num
+            start_idx = random.randint(0, random_range) if random_range > 0 else 0
+
+            ## calculate frame indices
+            frame_indices = [start_idx + frame_stride*i for i in range(self.video_length)]
+            try:
+                frames = video_reader.get_batch(frame_indices)
+                break
+            except:
+                print(f"Get frames failed! path = {video_path}; [max_ind vs frame_total:{max(frame_indices)} / {frame_num}]")
+                index += 1
+                continue
+        
+        ## process data
+        assert(frames.shape[0] == self.video_length),f'{len(frames)}, self.video_length={self.video_length}'
+        frames = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float() # [t,h,w,c] -> [c,t,h,w]
+        
+        if self.spatial_transform is not None:
+            frames = self.spatial_transform(frames)
+        
+        if self.resolution is not None:
+            assert (frames.shape[2], frames.shape[3]) == (self.resolution[0], self.resolution[1]), f'frames={frames.shape}, self.resolution={self.resolution}'
+        
+        ## turn frames tensors to [-1,1]
+        frames = (frames / 255 - 0.5) * 2
+        fps_clip = fps_ori // frame_stride
+        if self.fps_max is not None and fps_clip > self.fps_max:
+            fps_clip = self.fps_max
+
+        data = {'video': frames, 'caption': caption, 'path': video_path, 'fps': fps_clip, 'frame_stride': frame_stride}
+        return data
+    
+    def __len__(self):
+        return len(self.metadata)
+
+
+if __name__== "__main__":
+    meta_path = "" ## path to the meta file
+    data_dir = "" ## path to the data directory
+    save_dir = "" ## path to the save directory
+    dataset = WebVid(meta_path,
+                 data_dir,
+                 subsample=None,
+                 video_length=16,
+                 resolution=[256,448],
+                 frame_stride=4,
+                 spatial_transform="resize_center_crop",
+                 crop_resolution=None,
+                 fps_max=None,
+                 load_raw_resolution=True
+                 )
+    dataloader = DataLoader(dataset,
+                    batch_size=1,
+                    num_workers=0,
+                    shuffle=False)
+
+    
+    import sys
+    sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
+    from utils.save_video import tensor_to_mp4
+    for i, batch in tqdm(enumerate(dataloader), desc="Data Batch"):
+        video = batch['video']
+        name = batch['path'][0].split('videos/')[-1].replace('/','_')
+        tensor_to_mp4(video, save_dir+'/'+name, fps=8)
+

lvdm/distributions.py

@@ -0,0 +1,95 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self, noise=None):
+        if noise is None:
+            noise = torch.randn(self.mean.shape)
+        
+        x = self.mean + self.std * noise.to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

lvdm/ema.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
+                             else torch.tensor(-1,dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                #remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.','')
+                self.m_name2s_name.update({name:s_name})
+                self.register_buffer(s_name,p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self,model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

lvdm/models/autoencoder.py

@@ -0,0 +1,275 @@
+import os
+from contextlib import contextmanager
+import torch
+import numpy as np
+from einops import rearrange
+import torch.nn.functional as F
+import pytorch_lightning as pl
+from lvdm.modules.networks.ae_modules import Encoder, Decoder
+from lvdm.distributions import DiagonalGaussianDistribution
+from utils.utils import instantiate_from_config
+
+TIMESTEPS=16
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 test=False,
+                 logdir=None,
+                 input_dim=4,
+                 test_args=None,
+                 additional_decode_keys=None,
+                 use_checkpoint=False,
+                 diff_boost_factor=3.0,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        self.input_dim = input_dim
+        self.test = test
+        self.test_args = test_args
+        self.logdir = logdir
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        if self.test:
+            self.init_test()
+    
+    def init_test(self,):
+        self.test = True
+        save_dir = os.path.join(self.logdir, "test")
+        if 'ckpt' in self.test_args:
+            ckpt_name = os.path.basename(self.test_args.ckpt).split('.ckpt')[0] + f'_epoch{self._cur_epoch}'
+            self.root = os.path.join(save_dir, ckpt_name)
+        else:
+            self.root = save_dir
+        if 'test_subdir' in self.test_args:
+            self.root = os.path.join(save_dir, self.test_args.test_subdir)
+
+        self.root_zs = os.path.join(self.root, "zs")
+        self.root_dec = os.path.join(self.root, "reconstructions")
+        self.root_inputs = os.path.join(self.root, "inputs")
+        os.makedirs(self.root, exist_ok=True)
+
+        if self.test_args.save_z:
+            os.makedirs(self.root_zs, exist_ok=True)
+        if self.test_args.save_reconstruction:
+            os.makedirs(self.root_dec, exist_ok=True)
+        if self.test_args.save_input:
+            os.makedirs(self.root_inputs, exist_ok=True)
+        assert(self.test_args is not None)
+        self.test_maximum = getattr(self.test_args, 'test_maximum', None) 
+        self.count = 0
+        self.eval_metrics = {}
+        self.decodes = []
+        self.save_decode_samples = 2048
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")
+        try:
+            self._cur_epoch = sd['epoch']
+            sd = sd["state_dict"]
+        except:
+            self._cur_epoch = 'null'
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        # self.load_state_dict(sd, strict=True)
+        print(f"Restored from {path}")
+
+    def encode(self, x, return_hidden_states=False, **kwargs):
+        if return_hidden_states:
+            h, hidden = self.encoder(x, return_hidden_states)
+            moments = self.quant_conv(h)
+            posterior = DiagonalGaussianDistribution(moments)
+            return posterior, hidden
+        else:
+            h = self.encoder(x)
+            moments = self.quant_conv(h)
+            posterior = DiagonalGaussianDistribution(moments)
+            return posterior
+
+    def decode(self, z, **kwargs):
+        if len(kwargs) == 0: ## use the original decoder in AutoencoderKL
+            z = self.post_quant_conv(z)
+        dec = self.decoder(z, **kwargs)  ##change for SVD decoder by adding **kwargs
+        return dec
+
+    def forward(self, input, sample_posterior=True, **additional_decode_kwargs):
+        input_tuple = (input, )
+        forward_temp = partial(self._forward, sample_posterior=sample_posterior, **additional_decode_kwargs)
+        return checkpoint(forward_temp, input_tuple, self.parameters(), self.use_checkpoint)
+        
+
+    def _forward(self, input, sample_posterior=True, **additional_decode_kwargs):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, **additional_decode_kwargs)
+        ## print(input.shape, dec.shape) torch.Size([16, 3, 256, 256]) torch.Size([16, 3, 256, 256])
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if x.dim() == 5 and self.input_dim == 4:
+            b,c,t,h,w = x.shape
+            self.b = b
+            self.t = t 
+            x = rearrange(x, 'b c t h w -> (b t) c h w')
+
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+    
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+from lvdm.models.autoencoder_dualref import VideoDecoder
+class AutoencoderKL_Dualref(AutoencoderKL):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 test=False,
+                 logdir=None,
+                 input_dim=4,
+                 test_args=None,
+                 additional_decode_keys=None,
+                 use_checkpoint=False,
+                 diff_boost_factor=3.0,
+                 ):
+        super().__init__(ddconfig, lossconfig, embed_dim, ckpt_path, ignore_keys, image_key, colorize_nlabels, monitor, test, logdir, input_dim, test_args, additional_decode_keys, use_checkpoint, diff_boost_factor)
+        self.decoder = VideoDecoder(**ddconfig)
+
+    def _forward(self, input, sample_posterior=True, **additional_decode_kwargs):
+        posterior, hidden_states = self.encode(input, return_hidden_states=True)
+
+        hidden_states_first_last = []
+        ### use only the first and last hidden states
+        for hid in hidden_states:
+            hid = rearrange(hid, '(b t) c h w -> b c t h w', t=TIMESTEPS)
+            hid_new = torch.cat([hid[:, :, 0:1], hid[:, :, -1:]], dim=2)
+            hidden_states_first_last.append(hid_new)
+
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, ref_context=hidden_states_first_last, **additional_decode_kwargs)
+        ## print(input.shape, dec.shape) torch.Size([16, 3, 256, 256]) torch.Size([16, 3, 256, 256])
+        return dec, posterior

lvdm/models/autoencoder_dualref.py

@@ -0,0 +1,1178 @@
+#### https://github.com/Stability-AI/generative-models
+from einops import rearrange, repeat
+import logging
+from typing import Any, Callable, Optional, Iterable, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+from packaging import version
+logpy = logging.getLogger(__name__)
+
+try:
+    import xformers
+    import xformers.ops
+
+    XFORMERS_IS_AVAILABLE = True
+except:
+    XFORMERS_IS_AVAILABLE = False
+    logpy.warning("no module 'xformers'. Processing without...")
+
+from lvdm.modules.attention_svd import LinearAttention, MemoryEfficientCrossAttention
+
+
+def nonlinearity(x):
+    # swish
+    return x * torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(
+        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
+                )
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def attention(self, h_: torch.Tensor) -> torch.Tensor:
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        b, c, h, w = q.shape
+        q, k, v = map(
+            lambda x: rearrange(x, "b c h w -> b 1 (h w) c").contiguous(), (q, k, v)
+        )
+        h_ = torch.nn.functional.scaled_dot_product_attention(
+            q, k, v
+        )  # scale is dim ** -0.5 per default
+        # compute attention
+
+        return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b)
+
+    def forward(self, x, **kwargs):
+        h_ = x
+        h_ = self.attention(h_)
+        h_ = self.proj_out(h_)
+        return x + h_
+
+
+class MemoryEfficientAttnBlock(nn.Module):
+    """
+    Uses xformers efficient implementation,
+    see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    Note: this is a single-head self-attention operation
+    """
+
+    #
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.attention_op: Optional[Any] = None
+
+    def attention(self, h_: torch.Tensor) -> torch.Tensor:
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        B, C, H, W = q.shape
+        q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v))
+
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(B, t.shape[1], 1, C)
+            .permute(0, 2, 1, 3)
+            .reshape(B * 1, t.shape[1], C)
+            .contiguous(),
+            (q, k, v),
+        )
+        out = xformers.ops.memory_efficient_attention(
+            q, k, v, attn_bias=None, op=self.attention_op
+        )
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(B, 1, out.shape[1], C)
+            .permute(0, 2, 1, 3)
+            .reshape(B, out.shape[1], C)
+        )
+        return rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C)
+
+    def forward(self, x, **kwargs):
+        h_ = x
+        h_ = self.attention(h_)
+        h_ = self.proj_out(h_)
+        return x + h_
+
+
+class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
+    def forward(self, x, context=None, mask=None, **unused_kwargs):
+        b, c, h, w = x.shape
+        x = rearrange(x, "b c h w -> b (h w) c")
+        out = super().forward(x, context=context, mask=mask)
+        out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c)
+        return x + out
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
+    assert attn_type in [
+        "vanilla",
+        "vanilla-xformers",
+        "memory-efficient-cross-attn",
+        "linear",
+        "none",
+        "memory-efficient-cross-attn-fusion",
+    ], f"attn_type {attn_type} unknown"
+    if (
+        version.parse(torch.__version__) < version.parse("2.0.0")
+        and attn_type != "none"
+    ):
+        assert XFORMERS_IS_AVAILABLE, (
+            f"We do not support vanilla attention in {torch.__version__} anymore, "
+            f"as it is too expensive. Please install xformers via e.g. 'pip install xformers==0.0.16'"
+        )
+        # attn_type = "vanilla-xformers"
+    logpy.info(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        assert attn_kwargs is None
+        return AttnBlock(in_channels)
+    elif attn_type == "vanilla-xformers":
+        logpy.info(
+            f"building MemoryEfficientAttnBlock with {in_channels} in_channels..."
+        )
+        return MemoryEfficientAttnBlock(in_channels)
+    elif attn_type == "memory-efficient-cross-attn":
+        attn_kwargs["query_dim"] = in_channels
+        return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
+    elif attn_type == "memory-efficient-cross-attn-fusion":
+        attn_kwargs["query_dim"] = in_channels
+        return MemoryEfficientCrossAttentionWrapperFusion(**attn_kwargs)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+class MemoryEfficientCrossAttentionWrapperFusion(MemoryEfficientCrossAttention):
+    # print('x.shape: ',x.shape, 'context.shape: ',context.shape) ##torch.Size([8, 128, 256, 256]) torch.Size([1, 128, 2, 256, 256])
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0, **kwargs):
+        super().__init__(query_dim, context_dim, heads, dim_head, dropout, **kwargs)
+        self.norm = Normalize(query_dim)
+        nn.init.zeros_(self.to_out[0].weight)
+        nn.init.zeros_(self.to_out[0].bias)
+
+    def forward(self, x, context=None, mask=None):
+        if self.training:
+            return checkpoint(self._forward, x, context, mask, use_reentrant=False)
+        else:
+            return self._forward(x, context, mask)
+
+    def _forward(
+        self,
+        x,
+        context=None,
+        mask=None,
+    ):
+        bt, c, h, w = x.shape
+        h_ = self.norm(x)
+        h_ = rearrange(h_, "b c h w -> b (h w) c")
+        q = self.to_q(h_)
+
+
+        b, c, l, h, w = context.shape
+        context = rearrange(context, "b c l h w -> (b l) (h w) c")
+        k = self.to_k(context)
+        v = self.to_v(context)
+        k = rearrange(k, "(b l) d c -> b l d c", l=l)
+        k = torch.cat([k[:, [0] * (bt//b)], k[:, [1]*(bt//b)]], dim=2)
+        k = rearrange(k, "b l d c -> (b l) d c")
+
+        v = rearrange(v, "(b l) d c -> b l d c", l=l)
+        v = torch.cat([v[:, [0] * (bt//b)], v[:, [1]*(bt//b)]], dim=2)
+        v = rearrange(v, "b l d c -> (b l) d c")
+
+
+        b, _, _ = q.shape  ##actually bt
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        if version.parse(xformers.__version__) >= version.parse("0.0.21"):
+            # NOTE: workaround for
+            # https://github.com/facebookresearch/xformers/issues/845
+            max_bs = 32768
+            N = q.shape[0]
+            n_batches = math.ceil(N / max_bs)
+            out = list()
+            for i_batch in range(n_batches):
+                batch = slice(i_batch * max_bs, (i_batch + 1) * max_bs)
+                out.append(
+                    xformers.ops.memory_efficient_attention(
+                        q[batch],
+                        k[batch],
+                        v[batch],
+                        attn_bias=None,
+                        op=self.attention_op,
+                    )
+                )
+            out = torch.cat(out, 0)
+        else:
+            out = xformers.ops.memory_efficient_attention(
+                q, k, v, attn_bias=None, op=self.attention_op
+            )
+
+        # TODO: Use this directly in the attention operation, as a bias
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        out = self.to_out(out)
+        out = rearrange(out, "bt (h w) c -> bt c h w", h=h, w=w, c=c)
+        return x + out 
+
+class Combiner(nn.Module):
+    def __init__(self, ch) -> None:
+        super().__init__()
+        self.conv = nn.Conv2d(ch,ch,1,padding=0)
+
+        nn.init.zeros_(self.conv.weight)
+        nn.init.zeros_(self.conv.bias)
+
+    def forward(self, x, context):
+        if self.training:
+            return checkpoint(self._forward, x, context, use_reentrant=False)
+        else:
+            return self._forward(x, context)
+    
+    def _forward(self, x, context):
+        ## x: b c h w, context: b c 2 h w
+        b, c, l, h, w = context.shape
+        bt, c, h, w = x.shape
+        context = rearrange(context, "b c l h w -> (b l) c h w")
+        context = self.conv(context)
+        context = rearrange(context, "(b l) c h w -> b c l h w", l=l)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=bt//b)
+        x[:,:,0] = x[:,:,0] + context[:,:,0]
+        x[:,:,-1] = x[:,:,-1] + context[:,:,1]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        use_linear_attn=False,
+        attn_type="vanilla-xformers",
+        attn_level=[2,3], 
+        **ignorekwargs,
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+        self.attn_level = attn_level
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,) + tuple(ch_mult)
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        logpy.info(
+            "Working with z of shape {} = {} dimensions.".format(
+                self.z_shape, np.prod(self.z_shape)
+            )
+        )
+
+        make_attn_cls = self._make_attn()
+        make_resblock_cls = self._make_resblock()
+        make_conv_cls = self._make_conv()
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = make_resblock_cls(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn_cls(block_in, attn_type=attn_type)
+        self.mid.block_2 = make_resblock_cls(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        self.attn_refinement = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    make_resblock_cls(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn_cls(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+            if i_level in self.attn_level:
+                self.attn_refinement.insert(0, make_attn_cls(block_in, attn_type='memory-efficient-cross-attn-fusion', attn_kwargs={}))
+            else:
+                self.attn_refinement.insert(0, Combiner(block_in))
+        # end
+        self.norm_out = Normalize(block_in)
+        self.attn_refinement.append(Combiner(block_in))
+        self.conv_out = make_conv_cls(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def _make_attn(self) -> Callable:
+        return make_attn
+
+    def _make_resblock(self) -> Callable:
+        return ResnetBlock
+
+    def _make_conv(self) -> Callable:
+        return torch.nn.Conv2d
+
+    def get_last_layer(self, **kwargs):
+        return self.conv_out.weight
+
+    def forward(self, z, ref_context=None, **kwargs):
+        ## ref_context: b c 2 h w, 2 means starting and ending frame
+        # assert z.shape[1:] == self.z_shape[1:]
+        # print("ref context:", ref_context)
+        self.last_z_shape = z.shape
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb, **kwargs)
+        h = self.mid.attn_1(h, **kwargs)
+        h = self.mid.block_2(h, temb, **kwargs)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb, **kwargs)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, **kwargs)
+            if ref_context:
+                h = self.attn_refinement[i_level](x=h, context=ref_context[i_level])
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        if ref_context:
+            # print(h.shape, ref_context[i_level].shape) #torch.Size([8, 128, 256, 256]) torch.Size([1, 128, 2, 256, 256])
+            h = self.attn_refinement[-1](x=h, context=ref_context[-1])
+        h = self.conv_out(h, **kwargs)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+#####
+
+
+from abc import abstractmethod
+from lvdm.models.utils_diffusion import timestep_embedding
+
+from torch.utils.checkpoint import checkpoint
+from lvdm.basics import (
+    zero_module,
+    conv_nd,
+    linear,
+    normalization,
+)
+from lvdm.modules.networks.openaimodel3d import Upsample, Downsample
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x: torch.Tensor, emb: torch.Tensor):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        emb_channels: int,
+        dropout: float,
+        out_channels: Optional[int] = None,
+        use_conv: bool = False,
+        use_scale_shift_norm: bool = False,
+        dims: int = 2,
+        use_checkpoint: bool = False,
+        up: bool = False,
+        down: bool = False,
+        kernel_size: int = 3,
+        exchange_temb_dims: bool = False,
+        skip_t_emb: bool = False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.exchange_temb_dims = exchange_temb_dims
+
+        if isinstance(kernel_size, Iterable):
+            padding = [k // 2 for k in kernel_size]
+        else:
+            padding = kernel_size // 2
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.skip_t_emb = skip_t_emb
+        self.emb_out_channels = (
+            2 * self.out_channels if use_scale_shift_norm else self.out_channels
+        )
+        if self.skip_t_emb:
+            # print(f"Skipping timestep embedding in {self.__class__.__name__}")
+            assert not self.use_scale_shift_norm
+            self.emb_layers = None
+            self.exchange_temb_dims = False
+        else:
+            self.emb_layers = nn.Sequential(
+                nn.SiLU(),
+                linear(
+                    emb_channels,
+                    self.emb_out_channels,
+                ),
+            )
+
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(
+                    dims,
+                    self.out_channels,
+                    self.out_channels,
+                    kernel_size,
+                    padding=padding,
+                )
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, kernel_size, padding=padding
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        if self.use_checkpoint:
+            return checkpoint(self._forward, x, emb, use_reentrant=False)
+        else:
+            return self._forward(x, emb)
+
+    def _forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+
+        if self.skip_t_emb:
+            emb_out = torch.zeros_like(h)
+        else:
+            emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            if self.exchange_temb_dims:
+                emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+#####
+
+#####
+from lvdm.modules.attention_svd import *
+class VideoTransformerBlock(nn.Module):
+    ATTENTION_MODES = {
+        "softmax": CrossAttention,
+        "softmax-xformers": MemoryEfficientCrossAttention,
+    }
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        context_dim=None,
+        gated_ff=True,
+        checkpoint=True,
+        timesteps=None,
+        ff_in=False,
+        inner_dim=None,
+        attn_mode="softmax",
+        disable_self_attn=False,
+        disable_temporal_crossattention=False,
+        switch_temporal_ca_to_sa=False,
+    ):
+        super().__init__()
+
+        attn_cls = self.ATTENTION_MODES[attn_mode]
+
+        self.ff_in = ff_in or inner_dim is not None
+        if inner_dim is None:
+            inner_dim = dim
+
+        assert int(n_heads * d_head) == inner_dim
+
+        self.is_res = inner_dim == dim
+
+        if self.ff_in:
+            self.norm_in = nn.LayerNorm(dim)
+            self.ff_in = FeedForward(
+                dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff
+            )
+
+        self.timesteps = timesteps
+        self.disable_self_attn = disable_self_attn
+        if self.disable_self_attn:
+            self.attn1 = attn_cls(
+                query_dim=inner_dim,
+                heads=n_heads,
+                dim_head=d_head,
+                context_dim=context_dim,
+                dropout=dropout,
+            )  # is a cross-attention
+        else:
+            self.attn1 = attn_cls(
+                query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout
+            )  # is a self-attention
+
+        self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff)
+
+        if disable_temporal_crossattention:
+            if switch_temporal_ca_to_sa:
+                raise ValueError
+            else:
+                self.attn2 = None
+        else:
+            self.norm2 = nn.LayerNorm(inner_dim)
+            if switch_temporal_ca_to_sa:
+                self.attn2 = attn_cls(
+                    query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout
+                )  # is a self-attention
+            else:
+                self.attn2 = attn_cls(
+                    query_dim=inner_dim,
+                    context_dim=context_dim,
+                    heads=n_heads,
+                    dim_head=d_head,
+                    dropout=dropout,
+                )  # is self-attn if context is none
+
+        self.norm1 = nn.LayerNorm(inner_dim)
+        self.norm3 = nn.LayerNorm(inner_dim)
+        self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
+
+        self.checkpoint = checkpoint
+        if self.checkpoint:
+            print(f"====>{self.__class__.__name__} is using checkpointing")
+        else:
+            print(f"====>{self.__class__.__name__} is NOT using checkpointing")
+
+    def forward(
+        self, x: torch.Tensor, context: torch.Tensor = None, timesteps: int = None
+    ) -> torch.Tensor:
+        if self.checkpoint:
+            return checkpoint(self._forward, x, context, timesteps, use_reentrant=False)
+        else:
+            return self._forward(x, context, timesteps=timesteps)
+
+    def _forward(self, x, context=None, timesteps=None):
+        assert self.timesteps or timesteps
+        assert not (self.timesteps and timesteps) or self.timesteps == timesteps
+        timesteps = self.timesteps or timesteps
+        B, S, C = x.shape
+        x = rearrange(x, "(b t) s c -> (b s) t c", t=timesteps)
+
+        if self.ff_in:
+            x_skip = x
+            x = self.ff_in(self.norm_in(x))
+            if self.is_res:
+                x += x_skip
+
+        if self.disable_self_attn:
+            x = self.attn1(self.norm1(x), context=context) + x
+        else:
+            x = self.attn1(self.norm1(x)) + x
+
+        if self.attn2 is not None:
+            if self.switch_temporal_ca_to_sa:
+                x = self.attn2(self.norm2(x)) + x
+            else:
+                x = self.attn2(self.norm2(x), context=context) + x
+        x_skip = x
+        x = self.ff(self.norm3(x))
+        if self.is_res:
+            x += x_skip
+
+        x = rearrange(
+            x, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
+        )
+        return x
+
+    def get_last_layer(self):
+        return self.ff.net[-1].weight
+
+#####
+
+#####
+import functools
+def partialclass(cls, *args, **kwargs):
+    class NewCls(cls):
+        __init__ = functools.partialmethod(cls.__init__, *args, **kwargs)
+
+    return NewCls
+######
+
+class VideoResBlock(ResnetBlock):
+    def __init__(
+        self,
+        out_channels,
+        *args,
+        dropout=0.0,
+        video_kernel_size=3,
+        alpha=0.0,
+        merge_strategy="learned",
+        **kwargs,
+    ):
+        super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs)
+        if video_kernel_size is None:
+            video_kernel_size = [3, 1, 1]
+        self.time_stack = ResBlock(
+            channels=out_channels,
+            emb_channels=0,
+            dropout=dropout,
+            dims=3,
+            use_scale_shift_norm=False,
+            use_conv=False,
+            up=False,
+            down=False,
+            kernel_size=video_kernel_size,
+            use_checkpoint=True,
+            skip_t_emb=True,
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, bs):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError()
+
+    def forward(self, x, temb, skip_video=False, timesteps=None):
+        if timesteps is None:
+            timesteps = self.timesteps
+
+        b, c, h, w = x.shape
+
+        x = super().forward(x, temb)
+
+        if not skip_video:
+            x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = self.time_stack(x, temb)
+
+            alpha = self.get_alpha(bs=b // timesteps)
+            x = alpha * x + (1.0 - alpha) * x_mix
+
+            x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class AE3DConv(torch.nn.Conv2d):
+    def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs):
+        super().__init__(in_channels, out_channels, *args, **kwargs)
+        if isinstance(video_kernel_size, Iterable):
+            padding = [int(k // 2) for k in video_kernel_size]
+        else:
+            padding = int(video_kernel_size // 2)
+
+        self.time_mix_conv = torch.nn.Conv3d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=video_kernel_size,
+            padding=padding,
+        )
+
+    def forward(self, input, timesteps, skip_video=False):
+        x = super().forward(input)
+        if skip_video:
+            return x
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        x = self.time_mix_conv(x)
+        return rearrange(x, "b c t h w -> (b t) c h w")
+
+
+class VideoBlock(AttnBlock):
+    def __init__(
+        self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned"
+    ):
+        super().__init__(in_channels)
+        # no context, single headed, as in base class
+        self.time_mix_block = VideoTransformerBlock(
+            dim=in_channels,
+            n_heads=1,
+            d_head=in_channels,
+            checkpoint=True,
+            ff_in=True,
+            attn_mode="softmax",
+        )
+
+        time_embed_dim = self.in_channels * 4
+        self.video_time_embed = torch.nn.Sequential(
+            torch.nn.Linear(self.in_channels, time_embed_dim),
+            torch.nn.SiLU(),
+            torch.nn.Linear(time_embed_dim, self.in_channels),
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def forward(self, x, timesteps, skip_video=False):
+        if skip_video:
+            return super().forward(x)
+
+        x_in = x
+        x = self.attention(x)
+        h, w = x.shape[2:]
+        x = rearrange(x, "b c h w -> b (h w) c")
+
+        x_mix = x
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False)
+        emb = self.video_time_embed(t_emb)  # b, n_channels
+        emb = emb[:, None, :]
+        x_mix = x_mix + emb
+
+        alpha = self.get_alpha()
+        x_mix = self.time_mix_block(x_mix, timesteps=timesteps)
+        x = alpha * x + (1.0 - alpha) * x_mix  # alpha merge
+
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+
+        return x_in + x
+
+    def get_alpha(
+        self,
+    ):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}")
+
+
+class MemoryEfficientVideoBlock(MemoryEfficientAttnBlock):
+    def __init__(
+        self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned"
+    ):
+        super().__init__(in_channels)
+        # no context, single headed, as in base class
+        self.time_mix_block = VideoTransformerBlock(
+            dim=in_channels,
+            n_heads=1,
+            d_head=in_channels,
+            checkpoint=True,
+            ff_in=True,
+            attn_mode="softmax-xformers",
+        )
+
+        time_embed_dim = self.in_channels * 4
+        self.video_time_embed = torch.nn.Sequential(
+            torch.nn.Linear(self.in_channels, time_embed_dim),
+            torch.nn.SiLU(),
+            torch.nn.Linear(time_embed_dim, self.in_channels),
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def forward(self, x, timesteps, skip_time_block=False):
+        if skip_time_block:
+            return super().forward(x)
+
+        x_in = x
+        x = self.attention(x)
+        h, w = x.shape[2:]
+        x = rearrange(x, "b c h w -> b (h w) c")
+
+        x_mix = x
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False)
+        emb = self.video_time_embed(t_emb)  # b, n_channels
+        emb = emb[:, None, :]
+        x_mix = x_mix + emb
+
+        alpha = self.get_alpha()
+        x_mix = self.time_mix_block(x_mix, timesteps=timesteps)
+        x = alpha * x + (1.0 - alpha) * x_mix  # alpha merge
+
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+
+        return x_in + x
+
+    def get_alpha(
+        self,
+    ):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}")
+
+
+def make_time_attn(
+    in_channels,
+    attn_type="vanilla",
+    attn_kwargs=None,
+    alpha: float = 0,
+    merge_strategy: str = "learned",
+):
+    assert attn_type in [
+        "vanilla",
+        "vanilla-xformers",
+    ], f"attn_type {attn_type} not supported for spatio-temporal attention"
+    print(
+        f"making spatial and temporal attention of type '{attn_type}' with {in_channels} in_channels"
+    )
+    if not XFORMERS_IS_AVAILABLE and attn_type == "vanilla-xformers":
+        print(
+            f"Attention mode '{attn_type}' is not available. Falling back to vanilla attention. "
+            f"This is not a problem in Pytorch >= 2.0. FYI, you are running with PyTorch version {torch.__version__}"
+        )
+        attn_type = "vanilla"
+
+    if attn_type == "vanilla":
+        assert attn_kwargs is None
+        return partialclass(
+            VideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy
+        )
+    elif attn_type == "vanilla-xformers":
+        print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
+        return partialclass(
+            MemoryEfficientVideoBlock,
+            in_channels,
+            alpha=alpha,
+            merge_strategy=merge_strategy,
+        )
+    else:
+        return NotImplementedError()
+
+
+class Conv2DWrapper(torch.nn.Conv2d):
+    def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor:
+        return super().forward(input)
+
+
+class VideoDecoder(Decoder):
+    available_time_modes = ["all", "conv-only", "attn-only"]
+
+    def __init__(
+        self,
+        *args,
+        video_kernel_size: Union[int, list] = [3,1,1],
+        alpha: float = 0.0,
+        merge_strategy: str = "learned",
+        time_mode: str = "conv-only",
+        **kwargs,
+    ):
+        self.video_kernel_size = video_kernel_size
+        self.alpha = alpha
+        self.merge_strategy = merge_strategy
+        self.time_mode = time_mode
+        assert (
+            self.time_mode in self.available_time_modes
+        ), f"time_mode parameter has to be in {self.available_time_modes}"
+        super().__init__(*args, **kwargs)
+
+    def get_last_layer(self, skip_time_mix=False, **kwargs):
+        if self.time_mode == "attn-only":
+            raise NotImplementedError("TODO")
+        else:
+            return (
+                self.conv_out.time_mix_conv.weight
+                if not skip_time_mix
+                else self.conv_out.weight
+            )
+
+    def _make_attn(self) -> Callable:
+        if self.time_mode not in ["conv-only", "only-last-conv"]:
+            return partialclass(
+                make_time_attn,
+                alpha=self.alpha,
+                merge_strategy=self.merge_strategy,
+            )
+        else:
+            return super()._make_attn()
+
+    def _make_conv(self) -> Callable:
+        if self.time_mode != "attn-only":
+            return partialclass(AE3DConv, video_kernel_size=self.video_kernel_size)
+        else:
+            return Conv2DWrapper
+
+    def _make_resblock(self) -> Callable:
+        if self.time_mode not in ["attn-only", "only-last-conv"]:
+            return partialclass(
+                VideoResBlock,
+                video_kernel_size=self.video_kernel_size,
+                alpha=self.alpha,
+                merge_strategy=self.merge_strategy,
+            )
+        else:
+            return super()._make_resblock()

lvdm/models/ddpm3d.py

@@ -0,0 +1,1312 @@
+"""
+wild mixture of
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+
+from functools import partial
+from contextlib import contextmanager
+import numpy as np
+from tqdm import tqdm
+from einops import rearrange, repeat
+import logging
+mainlogger = logging.getLogger('mainlogger')
+import random
+import torch
+import torch.nn as nn
+from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR
+from torchvision.utils import make_grid
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_only
+from utils.utils import instantiate_from_config
+from lvdm.ema import LitEma
+from lvdm.models.samplers.ddim import DDIMSampler
+from lvdm.distributions import DiagonalGaussianDistribution
+from lvdm.models.utils_diffusion import make_beta_schedule, rescale_zero_terminal_snr
+from lvdm.basics import disabled_train
+from lvdm.common import (
+    extract_into_tensor,
+    noise_like,
+    exists,
+    default
+)
+import math
+from lvdm.models.autoencoder_dualref import VideoDecoder
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+class DDPM(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor=None,
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all assuming fixed variance schedules
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.,
+                 rescale_betas_zero_snr=False,
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"'
+        self.parameterization = parameterization
+        mainlogger.info(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.channels = channels
+        self.temporal_length = unet_config.params.temporal_length
+        self.image_size = image_size  # try conv?
+        if isinstance(self.image_size, int):
+            self.image_size = [self.image_size, self.image_size]
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        #count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        self.rescale_betas_zero_snr = rescale_betas_zero_snr
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            mainlogger.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        ## for reschedule
+        self.given_betas = given_betas
+        self.beta_schedule = beta_schedule
+        self.timesteps = timesteps
+        self.cosine_s = cosine_s
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                       cosine_s=cosine_s)
+        if self.rescale_betas_zero_snr:
+            betas = rescale_zero_terminal_snr(betas)
+        
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+
+        if self.parameterization != 'v':
+            self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+            self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+        else:
+            self.register_buffer('sqrt_recip_alphas_cumprod', torch.zeros_like(to_torch(alphas_cumprod)))
+            self.register_buffer('sqrt_recipm1_alphas_cumprod', torch.zeros_like(to_torch(alphas_cumprod)))
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
+                    1. - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
+        elif self.parameterization == "v":
+            lvlb_weights = torch.ones_like(self.betas ** 2 / (
+                    2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)))
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                mainlogger.info(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    mainlogger.info(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    mainlogger.info("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        mainlogger.info(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            mainlogger.info(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            mainlogger.info(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def predict_start_from_z_and_v(self, x_t, t, v):
+        # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        return (
+                extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
+        )
+
+    def predict_eps_from_z_and_v(self, x_t, t, v):
+        return (
+                extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def get_v(self, x, noise, t):
+        return (
+                extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise -
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x
+        )
+
+    def get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.model(x_noisy, t)
+
+        loss_dict = {}
+        if self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "v":
+            target = self.get_v(x_start, noise, t)
+        else:
+            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+
+        loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
+
+        log_prefix = 'train' if self.training else 'val'
+
+        loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
+        loss_simple = loss.mean() * self.l_simple_weight
+
+        loss_vlb = (self.lvlb_weights[t] * loss).mean()
+        loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
+
+        loss = loss_simple + self.original_elbo_weight * loss_vlb
+
+        loss_dict.update({f'{log_prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def forward(self, x, *args, **kwargs):
+        # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        '''
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        '''
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def shared_step(self, batch):
+        x = self.get_input(batch, self.first_stage_key)
+        loss, loss_dict = self(x)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(loss_dict, prog_bar=True,
+                      logger=True, on_step=True, on_epoch=True)
+
+        self.log("global_step", self.global_step,
+                 prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        return loss
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        _, loss_dict_no_ema = self.shared_step(batch)
+        with self.ema_scope():
+            _, loss_dict_ema = self.shared_step(batch)
+            loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
+        self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+        self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.learn_logvar:
+            params = params + [self.logvar]
+        opt = torch.optim.AdamW(params, lr=lr)
+        return opt
+
+class LatentDiffusion(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="caption",
+                 cond_stage_trainable=False,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 uncond_prob=0.2,
+                 uncond_type="empty_seq",
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 encoder_type="2d",
+                 only_model=False,
+                 noise_strength=0,
+                 use_dynamic_rescale=False,
+                 base_scale=0.7,
+                 turning_step=400,
+                 loop_video=False,
+                 fps_condition_type='fs',
+                 perframe_ae=False,
+                 # added
+                 logdir=None,
+                 rand_cond_frame=False,
+                 en_and_decode_n_samples_a_time=None,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        conditioning_key = default(conditioning_key, 'crossattn')
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        self.noise_strength = noise_strength
+        self.use_dynamic_rescale = use_dynamic_rescale
+        self.loop_video = loop_video
+        self.fps_condition_type = fps_condition_type
+        self.perframe_ae = perframe_ae
+
+        self.logdir = logdir
+        self.rand_cond_frame = rand_cond_frame
+        self.en_and_decode_n_samples_a_time = en_and_decode_n_samples_a_time
+
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+
+        if use_dynamic_rescale:
+            scale_arr1 = np.linspace(1.0, base_scale, turning_step)
+            scale_arr2 = np.full(self.num_timesteps, base_scale)
+            scale_arr = np.concatenate((scale_arr1, scale_arr2))
+            to_torch = partial(torch.tensor, dtype=torch.float32)
+            self.register_buffer('scale_arr', to_torch(scale_arr))
+
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.first_stage_config = first_stage_config
+        self.cond_stage_config = cond_stage_config        
+        self.clip_denoised = False
+
+        self.cond_stage_forward = cond_stage_forward
+        self.encoder_type = encoder_type
+        assert(encoder_type in ["2d", "3d"])
+        self.uncond_prob = uncond_prob
+        self.classifier_free_guidance = True if uncond_prob > 0 else False
+        assert(uncond_type in ["zero_embed", "empty_seq"])
+        self.uncond_type = uncond_type
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
+            self.restarted_from_ckpt = True
+                
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx=None):
+        # only for very first batch, reset the self.scale_factor
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and \
+                not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            mainlogger.info("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            mainlogger.info(f"setting self.scale_factor to {self.scale_factor}")
+            mainlogger.info("### USING STD-RESCALING ###")
+            mainlogger.info(f"std={z.flatten().std()}")
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model.eval()
+            self.cond_stage_model.train = disabled_train
+            for param in self.cond_stage_model.parameters():
+                param.requires_grad = False
+        else:
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+    
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def get_first_stage_encoding(self, encoder_posterior, noise=None):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample(noise=noise)
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+   
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if self.encoder_type == "2d" and x.dim() == 5:
+            b, _, t, _, _ = x.shape
+            x = rearrange(x, 'b c t h w -> (b t) c h w')
+            reshape_back = True
+        else:
+            reshape_back = False
+        
+        ## consume more GPU memory but faster
+        if not self.perframe_ae:
+            encoder_posterior = self.first_stage_model.encode(x)
+            results = self.get_first_stage_encoding(encoder_posterior).detach()
+        else:  ## consume less GPU memory but slower
+            results = []
+            for index in range(x.shape[0]):
+                frame_batch = self.first_stage_model.encode(x[index:index+1,:,:,:])
+                frame_result = self.get_first_stage_encoding(frame_batch).detach()
+                results.append(frame_result)
+            results = torch.cat(results, dim=0)
+
+        if reshape_back:
+            results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
+        
+        return results
+    
+    def decode_core(self, z, **kwargs):
+        if self.encoder_type == "2d" and z.dim() == 5:
+            b, _, t, _, _ = z.shape
+            z = rearrange(z, 'b c t h w -> (b t) c h w')
+            reshape_back = True
+        else:
+            reshape_back = False
+
+        z = 1. / self.scale_factor * z 
+        if not self.perframe_ae: 
+            results = self.first_stage_model.decode(z, **kwargs)
+        else:
+
+            results = []
+            
+            n_samples = default(self.en_and_decode_n_samples_a_time, self.temporal_length) 
+            n_rounds = math.ceil(z.shape[0] / n_samples)
+            with torch.autocast("cuda", enabled=True):
+                for n in range(n_rounds):
+                    if isinstance(self.first_stage_model.decoder, VideoDecoder):
+                        kwargs.update({"timesteps": len(z[n * n_samples : (n + 1) * n_samples])})
+                    else:
+                        kwargs = {}
+                    
+                    out = self.first_stage_model.decode(
+                        z[n * n_samples : (n + 1) * n_samples], **kwargs
+                    )
+                    results.append(out)
+            results = torch.cat(results, dim=0)
+
+        if reshape_back:
+            results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
+        return results
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, **kwargs):
+        return self.decode_core(z, **kwargs)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, **kwargs):
+        return self.decode_core(z, **kwargs)
+    
+    @torch.no_grad()
+    def get_batch_input(self, batch, random_uncond, return_first_stage_outputs=False, return_original_cond=False):
+        ## video shape: b, c, t, h, w
+        x = super().get_input(batch, self.first_stage_key)
+
+        ## encode video frames x to z via a 2D encoder
+        z = self.encode_first_stage(x)
+                
+        ## get caption condition
+        cond = batch[self.cond_stage_key]
+        if random_uncond and self.uncond_type == 'empty_seq':
+            for i, ci in enumerate(cond):
+                if random.random() < self.uncond_prob:
+                    cond[i] = ""
+        if isinstance(cond, dict) or isinstance(cond, list):
+            cond_emb = self.get_learned_conditioning(cond)
+        else:
+            cond_emb = self.get_learned_conditioning(cond.to(self.device))
+        if random_uncond and self.uncond_type == 'zero_embed':
+            for i, ci in enumerate(cond):
+                if random.random() < self.uncond_prob:
+                    cond_emb[i] = torch.zeros_like(cond_emb[i])
+        
+        out = [z, cond_emb]
+        ## optional output: self-reconst or caption
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([xrec])
+
+        if return_original_cond:
+            out.append(cond)
+
+        return out
+
+    def forward(self, x, c, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.use_dynamic_rescale:
+            x = x * extract_into_tensor(self.scale_arr, t, x.shape)
+        return self.p_losses(x, c, t, **kwargs)
+
+    def shared_step(self, batch, random_uncond, **kwargs):
+        x, c = self.get_batch_input(batch, random_uncond=random_uncond)
+        loss, loss_dict = self(x, c, **kwargs)
+
+        return loss, loss_dict
+
+    def apply_model(self, x_noisy, t, cond, **kwargs):
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        x_recon = self.model(x_noisy, t, **cond, **kwargs)
+
+        if isinstance(x_recon, tuple):
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def p_losses(self, x_start, cond, t, noise=None, **kwargs):
+        if self.noise_strength > 0:
+            b, c, f, _, _ = x_start.shape
+            offset_noise = torch.randn(b, c, f, 1, 1, device=x_start.device)
+            noise = default(noise, lambda: torch.randn_like(x_start) + self.noise_strength * offset_noise)
+        else:
+            noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+
+        model_output = self.apply_model(x_noisy, t, cond, **kwargs)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "v":
+            target = self.get_v(x_start, noise, t)
+        else:
+            raise NotImplementedError()
+        
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3, 4])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        if self.logvar.device is not self.device:
+            self.logvar = self.logvar.to(self.device)
+        logvar_t = self.logvar[t]
+        # logvar_t = self.logvar[t.item()].to(self.device) # device conflict when ddp shared
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3, 4))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict  
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch, random_uncond=self.classifier_free_guidance)
+        ## sync_dist | rank_zero_only 
+        self.log_dict(loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=True, sync_dist=False)
+        #self.log("epoch/global_step", self.global_step.float(), prog_bar=True, logger=True, on_step=True, on_epoch=False)
+        '''
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False, rank_zero_only=True)
+        '''
+        if (batch_idx+1) % self.log_every_t == 0:
+            mainlogger.info(f"batch:{batch_idx}|epoch:{self.current_epoch} [globalstep:{self.global_step}]: loss={loss}")
+        return loss
+    
+    def _get_denoise_row_from_list(self, samples, desc=''):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device)))
+        n_log_timesteps = len(denoise_row)
+
+        denoise_row = torch.stack(denoise_row)  # n_log_timesteps, b, C, H, W
+        
+        if denoise_row.dim() == 5:
+            denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+            denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+            denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps)
+        elif denoise_row.dim() == 6:
+            # video, grid_size=[n_log_timesteps*bs, t]
+            video_length = denoise_row.shape[3]
+            denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w')
+            denoise_grid = rearrange(denoise_grid, 'b n c t h w -> (b n) c t h w')
+            denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w')
+            denoise_grid = make_grid(denoise_grid, nrow=video_length)
+        else:
+            raise ValueError
+
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, sample=True, ddim_steps=200, ddim_eta=1., plot_denoise_rows=False, \
+                    unconditional_guidance_scale=1.0, **kwargs):
+        """ log images for LatentDiffusion """
+        ##### control sampled imgae for logging, larger value may cause OOM
+        sampled_img_num = 2
+        for key in batch.keys():
+            batch[key] = batch[key][:sampled_img_num]
+
+        ## TBD: currently, classifier_free_guidance sampling is only supported by DDIM
+        use_ddim = ddim_steps is not None
+        log = dict()
+        z, c, xrec, xc = self.get_batch_input(batch, random_uncond=False,
+                                                return_first_stage_outputs=True,
+                                                return_original_cond=True)
+
+        N = xrec.shape[0]
+        log["reconst"] = xrec
+        log["condition"] = xc
+        
+
+        if sample:
+            # get uncond embedding for classifier-free guidance sampling
+            if unconditional_guidance_scale != 1.0:
+                if isinstance(c, dict):
+                    c_cat, c_emb = c["c_concat"][0], c["c_crossattn"][0]
+                    log["condition_cat"] = c_cat
+                else:
+                    c_emb = c
+
+                if self.uncond_type == "empty_seq":
+                    prompts = N * [""]
+                    uc = self.get_learned_conditioning(prompts)
+                elif self.uncond_type == "zero_embed":
+                    uc = torch.zeros_like(c_emb)
+                ## hybrid case
+                if isinstance(c, dict):
+                    uc_hybrid = {"c_concat": [c_cat], "c_crossattn": [uc]}
+                    uc = uc_hybrid
+            else:
+                uc = None
+
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta,
+                                                         unconditional_guidance_scale=unconditional_guidance_scale,
+                                                         unconditional_conditioning=uc, x0=z, **kwargs)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+        return log
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_x0=False, score_corrector=None, corrector_kwargs=None, **kwargs):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, **kwargs)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+
+        if return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, return_x0=False, \
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, **kwargs):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs)
+        if return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False, x_T=None, verbose=True, callback=None, \
+                      timesteps=None, mask=None, x0=None, img_callback=None, start_T=None, log_every_t=None, **kwargs):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]        
+        # sample an initial noise
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts, clip_denoised=self.clip_denoised, **kwargs)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, \
+               verbose=True, timesteps=None, mask=None, x0=None, shape=None, **kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.temporal_length, *self.image_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps,
+                                  mask=mask, x0=x0, **kwargs)
+
+    @torch.no_grad()
+    def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs):
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.temporal_length, *self.image_size)
+            samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size, return_intermediates=True, **kwargs)
+
+        return samples, intermediates
+
+    def configure_schedulers(self, optimizer):
+        assert 'target' in self.scheduler_config
+        scheduler_name = self.scheduler_config.target.split('.')[-1]
+        interval = self.scheduler_config.interval
+        frequency = self.scheduler_config.frequency
+        if scheduler_name == "LambdaLRScheduler":
+            scheduler = instantiate_from_config(self.scheduler_config)
+            scheduler.start_step = self.global_step
+            lr_scheduler = {
+                            'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
+                            'interval': interval,
+                            'frequency': frequency
+            }
+        elif scheduler_name == "CosineAnnealingLRScheduler":
+            scheduler = instantiate_from_config(self.scheduler_config)
+            decay_steps = scheduler.decay_steps
+            last_step = -1 if self.global_step == 0 else scheduler.start_step
+            lr_scheduler = {
+                            'scheduler': CosineAnnealingLR(optimizer, T_max=decay_steps, last_epoch=last_step),
+                            'interval': interval,
+                            'frequency': frequency
+            }
+        else:
+            raise NotImplementedError
+        return lr_scheduler
+
+class LatentVisualDiffusion(LatentDiffusion):
+    def __init__(self, img_cond_stage_config, image_proj_stage_config, freeze_embedder=True, image_proj_model_trainable=True, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.image_proj_model_trainable = image_proj_model_trainable
+        self._init_embedder(img_cond_stage_config, freeze_embedder)
+        self._init_img_ctx_projector(image_proj_stage_config, image_proj_model_trainable)
+
+    def _init_img_ctx_projector(self, config, trainable):
+        self.image_proj_model = instantiate_from_config(config)
+        if not trainable:
+            self.image_proj_model.eval()
+            self.image_proj_model.train = disabled_train
+            for param in self.image_proj_model.parameters():
+                param.requires_grad = False
+
+    def _init_embedder(self, config, freeze=True):
+        self.embedder = instantiate_from_config(config)
+        if freeze:
+            self.embedder.eval()
+            self.embedder.train = disabled_train
+            for param in self.embedder.parameters():
+                param.requires_grad = False
+
+    def shared_step(self, batch, random_uncond, **kwargs):
+        x, c, fs = self.get_batch_input(batch, random_uncond=random_uncond, return_fs=True)
+        kwargs.update({"fs": fs.long()})
+        loss, loss_dict = self(x, c, **kwargs)
+        return loss, loss_dict
+    
+    def get_batch_input(self, batch, random_uncond, return_first_stage_outputs=False, return_original_cond=False, return_fs=False, return_cond_frame=False, return_original_input=False, **kwargs):
+        ## x: b c t h w
+        x = super().get_input(batch, self.first_stage_key)
+        ## encode video frames x to z via a 2D encoder        
+        z = self.encode_first_stage(x)
+        
+        ## get caption condition
+        cond_input = batch[self.cond_stage_key]
+
+        if isinstance(cond_input, dict) or isinstance(cond_input, list):
+            cond_emb = self.get_learned_conditioning(cond_input)
+        else:
+            cond_emb = self.get_learned_conditioning(cond_input.to(self.device))
+                
+        cond = {}
+        ## to support classifier-free guidance, randomly drop out only text conditioning 5%, only image conditioning 5%, and both 5%.
+        if random_uncond:
+            random_num = torch.rand(x.size(0), device=x.device)
+        else:
+            random_num = torch.ones(x.size(0), device=x.device)  ## by doning so, we can get text embedding and complete img emb for inference
+        prompt_mask = rearrange(random_num < 2 * self.uncond_prob, "n -> n 1 1")
+        input_mask = 1 - rearrange((random_num >= self.uncond_prob).float() * (random_num < 3 * self.uncond_prob).float(), "n -> n 1 1 1")
+
+        null_prompt = self.get_learned_conditioning([""])
+        prompt_imb = torch.where(prompt_mask, null_prompt, cond_emb.detach())
+
+        ## get conditioning frame
+        cond_frame_index = 0
+        if self.rand_cond_frame:
+            cond_frame_index = random.randint(0, self.model.diffusion_model.temporal_length-1)
+
+        img = x[:,:,cond_frame_index,...]
+        img = input_mask * img
+        ## img: b c h w
+        img_emb = self.embedder(img) ## b l c
+        img_emb = self.image_proj_model(img_emb)
+
+        if self.model.conditioning_key == 'hybrid':
+            ## simply repeat the cond_frame to match the seq_len of z
+            img_cat_cond = z[:,:,cond_frame_index,:,:]
+            img_cat_cond = img_cat_cond.unsqueeze(2)
+            img_cat_cond = repeat(img_cat_cond, 'b c t h w -> b c (repeat t) h w', repeat=z.shape[2])
+
+            cond["c_concat"] = [img_cat_cond] # b c t h w
+        cond["c_crossattn"] = [torch.cat([prompt_imb, img_emb], dim=1)] ## concat in the seq_len dim
+
+        out = [z, cond]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([xrec])
+
+        if return_original_cond:
+            out.append(cond_input)
+        if return_fs:
+            if self.fps_condition_type == 'fs':
+                fs = super().get_input(batch, 'frame_stride')
+            elif self.fps_condition_type == 'fps':
+                fs = super().get_input(batch, 'fps')
+            out.append(fs)
+        if return_cond_frame:
+            out.append(x[:,:,cond_frame_index,...].unsqueeze(2))
+        if return_original_input:
+            out.append(x)
+
+        return out
+
+    @torch.no_grad()
+    def log_images(self, batch, sample=True, ddim_steps=50, ddim_eta=1., plot_denoise_rows=False, \
+                    unconditional_guidance_scale=1.0, mask=None, **kwargs):
+        """ log images for LatentVisualDiffusion """
+        ##### sampled_img_num: control sampled imgae for logging, larger value may cause OOM
+        sampled_img_num = 1
+        for key in batch.keys():
+            batch[key] = batch[key][:sampled_img_num]
+
+        ## TBD: currently, classifier_free_guidance sampling is only supported by DDIM
+        use_ddim = ddim_steps is not None
+        log = dict()
+
+        z, c, xrec, xc, fs, cond_x = self.get_batch_input(batch, random_uncond=False,
+                                                return_first_stage_outputs=True,
+                                                return_original_cond=True,
+                                                return_fs=True,
+                                                return_cond_frame=True)
+
+        N = xrec.shape[0]
+        log["image_condition"] = cond_x
+        log["reconst"] = xrec
+        xc_with_fs = []
+        for idx, content in enumerate(xc):
+            xc_with_fs.append(content + '_fs=' + str(fs[idx].item()))
+        log["condition"] = xc_with_fs
+        kwargs.update({"fs": fs.long()})
+
+        c_cat = None
+        if sample:
+            # get uncond embedding for classifier-free guidance sampling
+            if unconditional_guidance_scale != 1.0:
+                if isinstance(c, dict):
+                    c_emb = c["c_crossattn"][0]
+                    if 'c_concat' in c.keys():
+                        c_cat = c["c_concat"][0]
+                else:
+                    c_emb = c
+
+                if self.uncond_type == "empty_seq":
+                    prompts = N * [""]
+                    uc_prompt = self.get_learned_conditioning(prompts)
+                elif self.uncond_type == "zero_embed":
+                    uc_prompt = torch.zeros_like(c_emb)
+                
+                img = torch.zeros_like(xrec[:,:,0]) ## b c h w
+                ## img: b c h w
+                img_emb = self.embedder(img) ## b l c
+                uc_img = self.image_proj_model(img_emb)
+
+                uc = torch.cat([uc_prompt, uc_img], dim=1)
+                ## hybrid case
+                if isinstance(c, dict):
+                    uc_hybrid = {"c_concat": [c_cat], "c_crossattn": [uc]}
+                    uc = uc_hybrid
+            else:
+                uc = None
+
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta,
+                                                         unconditional_guidance_scale=unconditional_guidance_scale,
+                                                         unconditional_conditioning=uc, x0=z, **kwargs)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+        return log
+
+    def configure_optimizers(self):
+        """ configure_optimizers for LatentDiffusion """
+        lr = self.learning_rate
+
+        params = list(self.model.parameters())
+        mainlogger.info(f"@Training [{len(params)}] Full Paramters.")
+
+        if self.cond_stage_trainable:
+            params_cond_stage = [p for p in self.cond_stage_model.parameters() if p.requires_grad == True]
+            mainlogger.info(f"@Training [{len(params_cond_stage)}] Paramters for Cond_stage_model.")
+            params.extend(params_cond_stage)
+        
+        if self.image_proj_model_trainable:
+            mainlogger.info(f"@Training [{len(list(self.image_proj_model.parameters()))}] Paramters for Image_proj_model.")
+            params.extend(list(self.image_proj_model.parameters()))   
+
+        if self.learn_logvar:
+            mainlogger.info('Diffusion model optimizing logvar')
+            if isinstance(params[0], dict):
+                params.append({"params": [self.logvar]})
+            else:
+                params.append(self.logvar)
+
+        ## optimizer
+        optimizer = torch.optim.AdamW(params, lr=lr)
+
+        ## lr scheduler
+        if self.use_scheduler:
+            mainlogger.info("Setting up scheduler...")
+            lr_scheduler = self.configure_schedulers(optimizer)
+            return [optimizer], [lr_scheduler]
+        
+        return optimizer
+
+
+class DiffusionWrapper(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None,
+                c_adm=None, s=None, mask=None, **kwargs):
+        # temporal_context = fps is foNone
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t, **kwargs)
+        elif self.conditioning_key == 'crossattn':
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc, **kwargs)
+        elif self.conditioning_key == 'hybrid':
+            ## it is just right [b,c,t,h,w]: concatenate in channel dim
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, **kwargs)
+        elif self.conditioning_key == 'resblockcond':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        elif self.conditioning_key == 'hybrid-adm':
+            assert c_adm is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, y=c_adm, **kwargs)
+        elif self.conditioning_key == 'hybrid-time':
+            assert s is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, s=s)
+        elif self.conditioning_key == 'concat-time-mask':
+            # assert s is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t, context=None, s=s, mask=mask)
+        elif self.conditioning_key == 'concat-adm-mask':
+            # assert s is not None
+            if c_concat is not None:
+                xc = torch.cat([x] + c_concat, dim=1)
+            else:
+                xc = x
+            out = self.diffusion_model(xc, t, context=None, y=s, mask=mask)
+        elif self.conditioning_key == 'hybrid-adm-mask':
+            cc = torch.cat(c_crossattn, 1)
+            if c_concat is not None:
+                xc = torch.cat([x] + c_concat, dim=1)
+            else:
+                xc = x
+            out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask)
+        elif self.conditioning_key == 'hybrid-time-adm': # adm means y, e.g., class index
+            # assert s is not None
+            assert c_adm is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm)
+        elif self.conditioning_key == 'crossattn-adm':
+            assert c_adm is not None
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc, y=c_adm)
+        else:
+            raise NotImplementedError()
+
+        return out

lvdm/models/samplers/ddim.py

@@ -0,0 +1,317 @@
+import numpy as np
+from tqdm import tqdm
+import torch
+from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps, rescale_noise_cfg
+from lvdm.common import noise_like
+from lvdm.common import extract_into_tensor
+import copy
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+        self.counter = 0
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        if self.model.use_dynamic_rescale:
+            self.ddim_scale_arr = self.model.scale_arr[self.ddim_timesteps]
+            self.ddim_scale_arr_prev = torch.cat([self.ddim_scale_arr[0:1], self.ddim_scale_arr[:-1]])
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               schedule_verbose=False,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               precision=None,
+               fs=None,
+               timestep_spacing='uniform', #uniform_trailing for starting from last timestep
+               guidance_rescale=0.0,
+               **kwargs
+               ):
+        
+        # check condition bs
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                try:
+                    cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                except:
+                    cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
+
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_discretize=timestep_spacing, ddim_eta=eta, verbose=schedule_verbose)
+        
+        # make shape
+        if len(shape) == 3:
+            C, H, W = shape
+            size = (batch_size, C, H, W)
+        elif len(shape) == 4:
+            C, T, H, W = shape
+            size = (batch_size, C, T, H, W)
+
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    verbose=verbose,
+                                                    precision=precision,
+                                                    fs=fs,
+                                                    guidance_rescale=guidance_rescale,
+                                                    **kwargs)
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,precision=None,fs=None,guidance_rescale=0.0,
+                      **kwargs):
+        device = self.model.betas.device        
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+        if precision is not None:
+            if precision == 16:
+                img = img.to(dtype=torch.float16)
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+            
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        if verbose:
+            iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+        else:
+            iterator = time_range
+
+        clean_cond = kwargs.pop("clean_cond", False)
+
+        # cond_copy, unconditional_conditioning_copy = copy.deepcopy(cond), copy.deepcopy(unconditional_conditioning)
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            ## use mask to blend noised original latent (img_orig) & new sampled latent (img)
+            if mask is not None:
+                assert x0 is not None
+                if clean_cond:
+                    img_orig = x0
+                else:
+                    img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass? <ddim inversion>
+                img = img_orig * mask + (1. - mask) * img # keep original & modify use img
+
+
+
+
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      mask=mask,x0=x0,fs=fs,guidance_rescale=guidance_rescale,
+                                      **kwargs)
+            
+
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      uc_type=None, conditional_guidance_scale_temporal=None,mask=None,x0=None,guidance_rescale=0.0,**kwargs):
+        b, *_, device = *x.shape, x.device
+        if x.dim() == 5:
+            is_video = True
+        else:
+            is_video = False
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            model_output = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
+        else:
+            ### do_classifier_free_guidance
+            if isinstance(c, torch.Tensor) or isinstance(c, dict):
+                e_t_cond = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            else:
+                raise NotImplementedError
+
+            model_output = e_t_uncond + unconditional_guidance_scale * (e_t_cond - e_t_uncond)
+
+            if guidance_rescale > 0.0:
+                model_output = rescale_noise_cfg(model_output, e_t_cond, guidance_rescale=guidance_rescale)
+
+        if self.model.parameterization == "v":
+            e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
+        else:
+            e_t = model_output
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps", 'not implemented'
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        # sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        
+        if is_video:
+            size = (b, 1, 1, 1, 1)
+        else:
+            size = (b, 1, 1, 1)
+        a_t = torch.full(size, alphas[index], device=device)
+        a_prev = torch.full(size, alphas_prev[index], device=device)
+        sigma_t = torch.full(size, sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        if self.model.parameterization != "v":
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        else:
+            pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
+        
+        if self.model.use_dynamic_rescale:
+            scale_t = torch.full(size, self.ddim_scale_arr[index], device=device)
+            prev_scale_t = torch.full(size, self.ddim_scale_arr_prev[index], device=device)
+            rescale = (prev_scale_t / scale_t)
+            pred_x0 *= rescale
+
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+    
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False, callback=None):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+            if callback: callback(i)
+        return x_dec
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)

lvdm/models/samplers/ddim_multiplecond.py

@@ -0,0 +1,323 @@
+import numpy as np
+from tqdm import tqdm
+import torch
+from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps, rescale_noise_cfg
+from lvdm.common import noise_like
+from lvdm.common import extract_into_tensor
+import copy
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+        self.counter = 0
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        if self.model.use_dynamic_rescale:
+            self.ddim_scale_arr = self.model.scale_arr[self.ddim_timesteps]
+            self.ddim_scale_arr_prev = torch.cat([self.ddim_scale_arr[0:1], self.ddim_scale_arr[:-1]])
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               schedule_verbose=False,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               precision=None,
+               fs=None,
+               timestep_spacing='uniform', #uniform_trailing for starting from last timestep
+               guidance_rescale=0.0,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        
+        # check condition bs
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                try:
+                    cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                except:
+                    cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
+
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        # print('==> timestep_spacing: ', timestep_spacing, guidance_rescale)
+        self.make_schedule(ddim_num_steps=S, ddim_discretize=timestep_spacing, ddim_eta=eta, verbose=schedule_verbose)
+        
+        # make shape
+        if len(shape) == 3:
+            C, H, W = shape
+            size = (batch_size, C, H, W)
+        elif len(shape) == 4:
+            C, T, H, W = shape
+            size = (batch_size, C, T, H, W)
+        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+        
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    verbose=verbose,
+                                                    precision=precision,
+                                                    fs=fs,
+                                                    guidance_rescale=guidance_rescale,
+                                                    **kwargs)
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,precision=None,fs=None,guidance_rescale=0.0,
+                      **kwargs):
+        device = self.model.betas.device        
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+        if precision is not None:
+            if precision == 16:
+                img = img.to(dtype=torch.float16)
+
+        
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+            
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        if verbose:
+            iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+        else:
+            iterator = time_range
+
+        clean_cond = kwargs.pop("clean_cond", False)
+
+        # cond_copy, unconditional_conditioning_copy = copy.deepcopy(cond), copy.deepcopy(unconditional_conditioning)
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            ## use mask to blend noised original latent (img_orig) & new sampled latent (img)
+            if mask is not None:
+                assert x0 is not None
+                if clean_cond:
+                    img_orig = x0
+                else:
+                    img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass? <ddim inversion>
+                img = img_orig * mask + (1. - mask) * img # keep original & modify use img
+
+
+
+
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      mask=mask,x0=x0,fs=fs,guidance_rescale=guidance_rescale,
+                                      **kwargs)
+            
+
+
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      uc_type=None, cfg_img=None,mask=None,x0=None,guidance_rescale=0.0, **kwargs):
+        b, *_, device = *x.shape, x.device
+        if x.dim() == 5:
+            is_video = True
+        else:
+            is_video = False
+        if cfg_img is None:
+            cfg_img = unconditional_guidance_scale
+
+        unconditional_conditioning_img_nonetext = kwargs['unconditional_conditioning_img_nonetext']
+
+        
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            model_output = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
+        else:
+            ### with unconditional condition
+            e_t_cond = self.model.apply_model(x, t, c, **kwargs)
+            e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            e_t_uncond_img = self.model.apply_model(x, t, unconditional_conditioning_img_nonetext, **kwargs)
+            # text cfg
+            model_output = e_t_uncond + cfg_img * (e_t_uncond_img - e_t_uncond) + unconditional_guidance_scale * (e_t_cond - e_t_uncond_img)
+            if guidance_rescale > 0.0:
+                model_output = rescale_noise_cfg(model_output, e_t_cond, guidance_rescale=guidance_rescale)
+        
+        if self.model.parameterization == "v":
+            e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
+        else:
+            e_t = model_output
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps", 'not implemented'
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        
+        if is_video:
+            size = (b, 1, 1, 1, 1)
+        else:
+            size = (b, 1, 1, 1)
+        a_t = torch.full(size, alphas[index], device=device)
+        a_prev = torch.full(size, alphas_prev[index], device=device)
+        sigma_t = torch.full(size, sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        if self.model.parameterization != "v":
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        else:
+            pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
+        
+        if self.model.use_dynamic_rescale:
+            scale_t = torch.full(size, self.ddim_scale_arr[index], device=device)
+            prev_scale_t = torch.full(size, self.ddim_scale_arr_prev[index], device=device)
+            rescale = (prev_scale_t / scale_t)
+            pred_x0 *= rescale
+
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+    
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False, callback=None):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+            if callback: callback(i)
+        return x_dec
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)

lvdm/models/utils_diffusion.py

@@ -0,0 +1,158 @@
+import math
+import numpy as np
+import torch
+import torch.nn.functional as F
+from einops import repeat
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+        steps_out = ddim_timesteps + 1
+    elif ddim_discr_method == 'uniform_trailing':
+        c = num_ddpm_timesteps / num_ddim_timesteps
+        ddim_timesteps = np.flip(np.round(np.arange(num_ddpm_timesteps, 0, -c))).astype(np.int64)
+        steps_out = ddim_timesteps - 1
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+        steps_out = ddim_timesteps + 1
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    # steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    # print(f'ddim_timesteps={ddim_timesteps}, len_alphacums={len(alphacums)}')
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+def rescale_zero_terminal_snr(betas):
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+    Args:
+        betas (`numpy.ndarray`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `numpy.ndarray`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = np.cumprod(alphas, axis=0)
+    alphas_bar_sqrt = np.sqrt(alphas_cumprod)
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].copy()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].copy()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = np.concatenate([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg

lvdm/modules/attention.py

@@ -0,0 +1,514 @@
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from functools import partial
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+from lvdm.common import (
+    checkpoint,
+    exists,
+    default,
+)
+from lvdm.basics import zero_module
+
+
+class RelativePosition(nn.Module):
+    """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
+
+    def __init__(self, num_units, max_relative_position):
+        super().__init__()
+        self.num_units = num_units
+        self.max_relative_position = max_relative_position
+        self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
+        nn.init.xavier_uniform_(self.embeddings_table)
+
+    def forward(self, length_q, length_k):
+        device = self.embeddings_table.device
+        range_vec_q = torch.arange(length_q, device=device)
+        range_vec_k = torch.arange(length_k, device=device)
+        distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
+        distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
+        final_mat = distance_mat_clipped + self.max_relative_position
+        final_mat = final_mat.long()
+        embeddings = self.embeddings_table[final_mat]
+        return embeddings
+
+
+class CrossAttention(nn.Module):
+
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., 
+                 relative_position=False, temporal_length=None, video_length=None, image_cross_attention=False, image_cross_attention_scale=1.0, image_cross_attention_scale_learnable=False, text_context_len=77):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head**-0.5
+        self.heads = heads
+        self.dim_head = dim_head
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+        
+        self.relative_position = relative_position
+        if self.relative_position:
+            assert(temporal_length is not None)
+            self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+            self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+        else:
+            ## only used for spatial attention, while NOT for temporal attention
+            if XFORMERS_IS_AVAILBLE and temporal_length is None:
+                self.forward = self.efficient_forward
+
+        self.video_length = video_length
+        self.image_cross_attention = image_cross_attention
+        self.image_cross_attention_scale = image_cross_attention_scale
+        self.text_context_len = text_context_len
+        self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable
+        if self.image_cross_attention:
+            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            if image_cross_attention_scale_learnable:
+                self.register_parameter('alpha', nn.Parameter(torch.tensor(0.)) )
+
+
+    def forward(self, x, context=None, mask=None):
+        spatial_self_attn = (context is None)
+        k_ip, v_ip, out_ip = None, None, None
+
+        h = self.heads
+        q = self.to_q(x)
+        context = default(context, x)
+
+        if self.image_cross_attention and not spatial_self_attn:
+            context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_image)
+            v_ip = self.to_v_ip(context_image)
+        else:
+            if not spatial_self_attn:
+                context = context[:,:self.text_context_len,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
+        if self.relative_position:
+            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
+            k2 = self.relative_position_k(len_q, len_k)
+            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check 
+            sim += sim2
+        del k
+
+        if exists(mask):
+            ## feasible for causal attention mask only
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b i j -> (b h) i j', h=h)
+            sim.masked_fill_(~(mask>0.5), max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+
+        out = torch.einsum('b i j, b j d -> b i d', sim, v)
+        if self.relative_position:
+            v2 = self.relative_position_v(len_q, len_v)
+            out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
+            out += out2
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+
+
+        ## for image cross-attention
+        if k_ip is not None:
+            k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
+            sim_ip =  torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
+            del k_ip
+            sim_ip = sim_ip.softmax(dim=-1)
+            out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
+            out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
+
+
+        if out_ip is not None:
+            if self.image_cross_attention_scale_learnable:
+                out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1)
+            else:
+                out = out + self.image_cross_attention_scale * out_ip
+        
+        return self.to_out(out)
+    
+    def efficient_forward(self, x, context=None, mask=None):
+        spatial_self_attn = (context is None)
+        k_ip, v_ip, out_ip = None, None, None
+
+        q = self.to_q(x)
+        context = default(context, x)
+
+        if self.image_cross_attention and not spatial_self_attn:
+            context, context_image = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_image)
+            v_ip = self.to_v_ip(context_image)
+        else:
+            if not spatial_self_attn:
+                context = context[:,:self.text_context_len,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+        # actually compute the attention, what we cannot get enough of
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)
+        
+        ## for image cross-attention
+        if k_ip is not None:
+            k_ip, v_ip = map(
+                lambda t: t.unsqueeze(3)
+                .reshape(b, t.shape[1], self.heads, self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b * self.heads, t.shape[1], self.dim_head)
+                .contiguous(),
+                (k_ip, v_ip),
+            )
+            out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None)
+            out_ip = (
+                out_ip.unsqueeze(0)
+                .reshape(b, self.heads, out.shape[1], self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b, out.shape[1], self.heads * self.dim_head)
+            )
+
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        if out_ip is not None:
+            if self.image_cross_attention_scale_learnable:
+                out = out + self.image_cross_attention_scale * out_ip * (torch.tanh(self.alpha)+1)
+            else:
+                out = out + self.image_cross_attention_scale * out_ip
+           
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                disable_self_attn=False, attention_cls=None, video_length=None, image_cross_attention=False, image_cross_attention_scale=1.0, image_cross_attention_scale_learnable=False, text_context_len=77):
+        super().__init__()
+        attn_cls = CrossAttention if attention_cls is None else attention_cls
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+            context_dim=context_dim if self.disable_self_attn else None)
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout, video_length=video_length, image_cross_attention=image_cross_attention, image_cross_attention_scale=image_cross_attention_scale, image_cross_attention_scale_learnable=image_cross_attention_scale_learnable,text_context_len=text_context_len)
+        self.image_cross_attention = image_cross_attention
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+
+    def forward(self, x, context=None, mask=None, **kwargs):
+        ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
+        input_tuple = (x,)      ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
+        if context is not None:
+            input_tuple = (x, context)
+        if mask is not None:
+            forward_mask = partial(self._forward, mask=mask)
+            return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)
+
+
+    def _forward(self, x, context=None, mask=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x
+        x = self.attn2(self.norm2(x), context=context, mask=mask) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data in spatial axis.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, disable_self_attn=False, use_linear=False, video_length=None,
+                 image_cross_attention=False, image_cross_attention_scale_learnable=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        attention_cls = None
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                disable_self_attn=disable_self_attn,
+                checkpoint=use_checkpoint,
+                attention_cls=attention_cls,
+                video_length=video_length,
+                image_cross_attention=image_cross_attention,
+                image_cross_attention_scale_learnable=image_cross_attention_scale_learnable,
+                ) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+
+    def forward(self, x, context=None, **kwargs):
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context, **kwargs)
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+    
+    
+class TemporalTransformer(nn.Module):
+    """
+    Transformer block for image-like data in temporal axis.
+    First, reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False, causal_block_size=1,
+                 relative_position=False, temporal_length=None):
+        super().__init__()
+        self.only_self_att = only_self_att
+        self.relative_position = relative_position
+        self.causal_attention = causal_attention
+        self.causal_block_size = causal_block_size
+
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        if not use_linear:
+            self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        if relative_position:
+            assert(temporal_length is not None)
+            attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
+        else:
+            attention_cls = partial(CrossAttention, temporal_length=temporal_length)
+        if self.causal_attention:
+            assert(temporal_length is not None)
+            self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))
+
+        if self.only_self_att:
+            context_dim = None
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                attention_cls=attention_cls,
+                checkpoint=use_checkpoint) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None):
+        b, c, t, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+
+        temp_mask = None
+        if self.causal_attention:
+            # slice the from mask map
+            temp_mask = self.mask[:,:t,:t].to(x.device)
+
+        if temp_mask is not None:
+            mask = temp_mask.to(x.device)
+            mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
+        else:
+            mask = None
+
+        if self.only_self_att:
+            ## note: if no context is given, cross-attention defaults to self-attention
+            for i, block in enumerate(self.transformer_blocks):
+                x = block(x, mask=mask)
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+        else:
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+            context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
+            for i, block in enumerate(self.transformer_blocks):
+                # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
+                for j in range(b):
+                    context_j = repeat(
+                        context[j],
+                        't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
+                    ## note: causal mask will not applied in cross-attention case
+                    x[j] = block(x[j], context=context_j)
+        
+        if self.use_linear:
+            x = self.proj_out(x)
+            x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
+            x = self.proj_out(x)
+            x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()
+
+        return x + x_in
+    
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_

lvdm/modules/attention_svd.py

@@ -0,0 +1,759 @@
+import logging
+import math
+from inspect import isfunction
+from typing import Any, Optional
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from packaging import version
+from torch import nn
+from torch.utils.checkpoint import checkpoint
+
+logpy = logging.getLogger(__name__)
+
+if version.parse(torch.__version__) >= version.parse("2.0.0"):
+    SDP_IS_AVAILABLE = True
+    from torch.backends.cuda import SDPBackend, sdp_kernel
+
+    BACKEND_MAP = {
+        SDPBackend.MATH: {
+            "enable_math": True,
+            "enable_flash": False,
+            "enable_mem_efficient": False,
+        },
+        SDPBackend.FLASH_ATTENTION: {
+            "enable_math": False,
+            "enable_flash": True,
+            "enable_mem_efficient": False,
+        },
+        SDPBackend.EFFICIENT_ATTENTION: {
+            "enable_math": False,
+            "enable_flash": False,
+            "enable_mem_efficient": True,
+        },
+        None: {"enable_math": True, "enable_flash": True, "enable_mem_efficient": True},
+    }
+else:
+    from contextlib import nullcontext
+
+    SDP_IS_AVAILABLE = False
+    sdp_kernel = nullcontext
+    BACKEND_MAP = {}
+    logpy.warn(
+        f"No SDP backend available, likely because you are running in pytorch "
+        f"versions < 2.0. In fact, you are using PyTorch {torch.__version__}. "
+        f"You might want to consider upgrading."
+    )
+
+try:
+    import xformers
+    import xformers.ops
+
+    XFORMERS_IS_AVAILABLE = True
+except:
+    XFORMERS_IS_AVAILABLE = False
+    logpy.warn("no module 'xformers'. Processing without...")
+
+# from .diffusionmodules.util import mixed_checkpoint as checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return {el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = (
+            nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
+            if not glu
+            else GEGLU(dim, inner_dim)
+        )
+
+        self.net = nn.Sequential(
+            project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(
+        num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(
+            qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
+        )
+        k = k.softmax(dim=-1)
+        context = torch.einsum("bhdn,bhen->bhde", k, v)
+        out = torch.einsum("bhde,bhdn->bhen", context, q)
+        out = rearrange(
+            out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
+        )
+        return self.to_out(out)
+
+
+class SelfAttention(nn.Module):
+    ATTENTION_MODES = ("xformers", "torch", "math")
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = False,
+        qk_scale: Optional[float] = None,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        attn_mode: str = "xformers",
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        assert attn_mode in self.ATTENTION_MODES
+        self.attn_mode = attn_mode
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, L, C = x.shape
+
+        qkv = self.qkv(x)
+        if self.attn_mode == "torch":
+            qkv = rearrange(
+                qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads
+            ).float()
+            q, k, v = qkv[0], qkv[1], qkv[2]  # B H L D
+            x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+            x = rearrange(x, "B H L D -> B L (H D)")
+        elif self.attn_mode == "xformers":
+            qkv = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.num_heads)
+            q, k, v = qkv[0], qkv[1], qkv[2]  # B L H D
+            x = xformers.ops.memory_efficient_attention(q, k, v)
+            x = rearrange(x, "B L H D -> B L (H D)", H=self.num_heads)
+        elif self.attn_mode == "math":
+            qkv = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+            q, k, v = qkv[0], qkv[1], qkv[2]  # B H L D
+            attn = (q @ k.transpose(-2, -1)) * self.scale
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = (attn @ v).transpose(1, 2).reshape(B, L, C)
+        else:
+            raise NotImplemented
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b (h w) c")
+        k = rearrange(k, "b c h w -> b c (h w)")
+        w_ = torch.einsum("bij,bjk->bik", q, k)
+
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, "b c h w -> b c (h w)")
+        w_ = rearrange(w_, "b i j -> b j i")
+        h_ = torch.einsum("bij,bjk->bik", v, w_)
+        h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(
+        self,
+        query_dim,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        backend=None,
+    ):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head**-0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
+        )
+        self.backend = backend
+
+    def forward(
+        self,
+        x,
+        context=None,
+        mask=None,
+        additional_tokens=None,
+        n_times_crossframe_attn_in_self=0,
+    ):
+        h = self.heads
+
+        if additional_tokens is not None:
+            # get the number of masked tokens at the beginning of the output sequence
+            n_tokens_to_mask = additional_tokens.shape[1]
+            # add additional token
+            x = torch.cat([additional_tokens, x], dim=1)
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        if n_times_crossframe_attn_in_self:
+            # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439
+            assert x.shape[0] % n_times_crossframe_attn_in_self == 0
+            n_cp = x.shape[0] // n_times_crossframe_attn_in_self
+            k = repeat(
+                k[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp
+            )
+            v = repeat(
+                v[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp
+            )
+
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
+
+        ## old
+        """
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        del q, k
+
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', sim, v)
+        """
+        ## new
+        with sdp_kernel(**BACKEND_MAP[self.backend]):
+            # print("dispatching into backend", self.backend, "q/k/v shape: ", q.shape, k.shape, v.shape)
+            out = F.scaled_dot_product_attention(
+                q, k, v, attn_mask=mask
+            )  # scale is dim_head ** -0.5 per default
+
+        del q, k, v
+        out = rearrange(out, "b h n d -> b n (h d)", h=h)
+
+        if additional_tokens is not None:
+            # remove additional token
+            out = out[:, n_tokens_to_mask:]
+        return self.to_out(out)
+
+
+class MemoryEfficientCrossAttention(nn.Module):
+    # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    def __init__(
+        self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs
+    ):
+        super().__init__()
+        logpy.debug(
+            f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, "
+            f"context_dim is {context_dim} and using {heads} heads with a "
+            f"dimension of {dim_head}."
+        )
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
+        )
+        self.attention_op: Optional[Any] = None
+
+    def forward(
+        self,
+        x,
+        context=None,
+        mask=None,
+        additional_tokens=None,
+        n_times_crossframe_attn_in_self=0,
+    ):
+        if additional_tokens is not None:
+            # get the number of masked tokens at the beginning of the output sequence
+            n_tokens_to_mask = additional_tokens.shape[1]
+            # add additional token
+            x = torch.cat([additional_tokens, x], dim=1)
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        if n_times_crossframe_attn_in_self:
+            # reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439
+            assert x.shape[0] % n_times_crossframe_attn_in_self == 0
+            # n_cp = x.shape[0]//n_times_crossframe_attn_in_self
+            k = repeat(
+                k[::n_times_crossframe_attn_in_self],
+                "b ... -> (b n) ...",
+                n=n_times_crossframe_attn_in_self,
+            )
+            v = repeat(
+                v[::n_times_crossframe_attn_in_self],
+                "b ... -> (b n) ...",
+                n=n_times_crossframe_attn_in_self,
+            )
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        if version.parse(xformers.__version__) >= version.parse("0.0.21"):
+            # NOTE: workaround for
+            # https://github.com/facebookresearch/xformers/issues/845
+            max_bs = 32768
+            N = q.shape[0]
+            n_batches = math.ceil(N / max_bs)
+            out = list()
+            for i_batch in range(n_batches):
+                batch = slice(i_batch * max_bs, (i_batch + 1) * max_bs)
+                out.append(
+                    xformers.ops.memory_efficient_attention(
+                        q[batch],
+                        k[batch],
+                        v[batch],
+                        attn_bias=None,
+                        op=self.attention_op,
+                    )
+                )
+            out = torch.cat(out, 0)
+        else:
+            out = xformers.ops.memory_efficient_attention(
+                q, k, v, attn_bias=None, op=self.attention_op
+            )
+
+        # TODO: Use this directly in the attention operation, as a bias
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        if additional_tokens is not None:
+            # remove additional token
+            out = out[:, n_tokens_to_mask:]
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    ATTENTION_MODES = {
+        "softmax": CrossAttention,  # vanilla attention
+        "softmax-xformers": MemoryEfficientCrossAttention,  # ampere
+    }
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        context_dim=None,
+        gated_ff=True,
+        checkpoint=True,
+        disable_self_attn=False,
+        attn_mode="softmax",
+        sdp_backend=None,
+    ):
+        super().__init__()
+        assert attn_mode in self.ATTENTION_MODES
+        if attn_mode != "softmax" and not XFORMERS_IS_AVAILABLE:
+            logpy.warn(
+                f"Attention mode '{attn_mode}' is not available. Falling "
+                f"back to native attention. This is not a problem in "
+                f"Pytorch >= 2.0. FYI, you are running with PyTorch "
+                f"version {torch.__version__}."
+            )
+            attn_mode = "softmax"
+        elif attn_mode == "softmax" and not SDP_IS_AVAILABLE:
+            logpy.warn(
+                "We do not support vanilla attention anymore, as it is too "
+                "expensive. Sorry."
+            )
+            if not XFORMERS_IS_AVAILABLE:
+                assert (
+                    False
+                ), "Please install xformers via e.g. 'pip install xformers==0.0.16'"
+            else:
+                logpy.info("Falling back to xformers efficient attention.")
+                attn_mode = "softmax-xformers"
+        attn_cls = self.ATTENTION_MODES[attn_mode]
+        if version.parse(torch.__version__) >= version.parse("2.0.0"):
+            assert sdp_backend is None or isinstance(sdp_backend, SDPBackend)
+        else:
+            assert sdp_backend is None
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            context_dim=context_dim if self.disable_self_attn else None,
+            backend=sdp_backend,
+        )  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(
+            query_dim=dim,
+            context_dim=context_dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            backend=sdp_backend,
+        )  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+        if self.checkpoint:
+            logpy.debug(f"{self.__class__.__name__} is using checkpointing")
+
+    def forward(
+        self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0
+    ):
+        kwargs = {"x": x}
+
+        if context is not None:
+            kwargs.update({"context": context})
+
+        if additional_tokens is not None:
+            kwargs.update({"additional_tokens": additional_tokens})
+
+        if n_times_crossframe_attn_in_self:
+            kwargs.update(
+                {"n_times_crossframe_attn_in_self": n_times_crossframe_attn_in_self}
+            )
+
+        # return mixed_checkpoint(self._forward, kwargs, self.parameters(), self.checkpoint)
+        if self.checkpoint:
+            # inputs = {"x": x, "context": context}
+            return checkpoint(self._forward, x, context)
+            # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint)
+        else:
+            return self._forward(**kwargs)
+
+    def _forward(
+        self, x, context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0
+    ):
+        x = (
+            self.attn1(
+                self.norm1(x),
+                context=context if self.disable_self_attn else None,
+                additional_tokens=additional_tokens,
+                n_times_crossframe_attn_in_self=n_times_crossframe_attn_in_self
+                if not self.disable_self_attn
+                else 0,
+            )
+            + x
+        )
+        x = (
+            self.attn2(
+                self.norm2(x), context=context, additional_tokens=additional_tokens
+            )
+            + x
+        )
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class BasicTransformerSingleLayerBlock(nn.Module):
+    ATTENTION_MODES = {
+        "softmax": CrossAttention,  # vanilla attention
+        "softmax-xformers": MemoryEfficientCrossAttention  # on the A100s not quite as fast as the above version
+        # (todo might depend on head_dim, check, falls back to semi-optimized kernels for dim!=[16,32,64,128])
+    }
+
+    def __init__(
+        self,
+        dim,
+        n_heads,
+        d_head,
+        dropout=0.0,
+        context_dim=None,
+        gated_ff=True,
+        checkpoint=True,
+        attn_mode="softmax",
+    ):
+        super().__init__()
+        assert attn_mode in self.ATTENTION_MODES
+        attn_cls = self.ATTENTION_MODES[attn_mode]
+        self.attn1 = attn_cls(
+            query_dim=dim,
+            heads=n_heads,
+            dim_head=d_head,
+            dropout=dropout,
+            context_dim=context_dim,
+        )
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        # inputs = {"x": x, "context": context}
+        # return checkpoint(self._forward, inputs, self.parameters(), self.checkpoint)
+        return checkpoint(self._forward, x, context)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x), context=context) + x
+        x = self.ff(self.norm2(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        depth=1,
+        dropout=0.0,
+        context_dim=None,
+        disable_self_attn=False,
+        use_linear=False,
+        attn_type="softmax",
+        use_checkpoint=True,
+        # sdp_backend=SDPBackend.FLASH_ATTENTION
+        sdp_backend=None,
+    ):
+        super().__init__()
+        logpy.debug(
+            f"constructing {self.__class__.__name__} of depth {depth} w/ "
+            f"{in_channels} channels and {n_heads} heads."
+        )
+
+        if exists(context_dim) and not isinstance(context_dim, list):
+            context_dim = [context_dim]
+        if exists(context_dim) and isinstance(context_dim, list):
+            if depth != len(context_dim):
+                logpy.warn(
+                    f"{self.__class__.__name__}: Found context dims "
+                    f"{context_dim} of depth {len(context_dim)}, which does not "
+                    f"match the specified 'depth' of {depth}. Setting context_dim "
+                    f"to {depth * [context_dim[0]]} now."
+                )
+                # depth does not match context dims.
+                assert all(
+                    map(lambda x: x == context_dim[0], context_dim)
+                ), "need homogenous context_dim to match depth automatically"
+                context_dim = depth * [context_dim[0]]
+        elif context_dim is None:
+            context_dim = [None] * depth
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    n_heads,
+                    d_head,
+                    dropout=dropout,
+                    context_dim=context_dim[d],
+                    disable_self_attn=disable_self_attn,
+                    attn_mode=attn_type,
+                    checkpoint=use_checkpoint,
+                    sdp_backend=sdp_backend,
+                )
+                for d in range(depth)
+            ]
+        )
+        if not use_linear:
+            self.proj_out = zero_module(
+                nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+            )
+        else:
+            # self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context]
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, "b c h w -> b (h w) c").contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            if i > 0 and len(context) == 1:
+                i = 0  # use same context for each block
+            x = block(x, context=context[i])
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+
+
+class SimpleTransformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        context_dim: Optional[int] = None,
+        dropout: float = 0.0,
+        checkpoint: bool = True,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                BasicTransformerBlock(
+                    dim,
+                    heads,
+                    dim_head,
+                    dropout=dropout,
+                    context_dim=context_dim,
+                    attn_mode="softmax-xformers",
+                    checkpoint=checkpoint,
+                )
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for layer in self.layers:
+            x = layer(x, context)
+        return x

lvdm/modules/encoders/condition.py

@@ -0,0 +1,389 @@
+import torch
+import torch.nn as nn
+import kornia
+import open_clip
+from torch.utils.checkpoint import checkpoint
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
+from lvdm.common import autocast
+from utils.utils import count_params
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class IdentityEncoder(AbstractEncoder):
+    def encode(self, x):
+        return x
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+        self.n_classes = n_classes
+        self.ucg_rate = ucg_rate
+
+    def forward(self, batch, key=None, disable_dropout=False):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        if self.ucg_rate > 0. and not disable_dropout:
+            mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
+            c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1)
+            c = c.long()
+        c = self.embedding(c)
+        return c
+
+    def get_unconditional_conditioning(self, bs, device="cuda"):
+        uc_class = self.n_classes - 1  # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
+        uc = torch.ones((bs,), device=device) * uc_class
+        uc = {self.key: uc}
+        return uc
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77,
+                 freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length  # TODO: typical value?
+        if freeze:
+            self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        # self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    LAYERS = [
+        "last",
+        "pooled",
+        "hidden"
+    ]
+
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
+                 freeze=True, layer="last", layer_idx=None):  # clip-vit-base-patch32
+        super().__init__()
+        assert layer in self.LAYERS
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        self.transformer = CLIPTextModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        self.layer_idx = layer_idx
+        if layer == "hidden":
+            assert layer_idx is not None
+            assert 0 <= abs(layer_idx) <= 12
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        # self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden")
+        if self.layer == "last":
+            z = outputs.last_hidden_state
+        elif self.layer == "pooled":
+            z = outputs.pooler_output[:, None, :]
+        else:
+            z = outputs.hidden_states[self.layer_idx]
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class ClipImageEmbedder(nn.Module):
+    def __init__(
+            self,
+            model,
+            jit=False,
+            device='cuda' if torch.cuda.is_available() else 'cpu',
+            antialias=True,
+            ucg_rate=0.
+    ):
+        super().__init__()
+        from clip import load as load_clip
+        self.model, _ = load_clip(name=model, device=device, jit=jit)
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.ucg_rate = ucg_rate
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # re-normalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x, no_dropout=False):
+        # x is assumed to be in range [-1,1]
+        out = self.model.encode_image(self.preprocess(x))
+        out = out.to(x.dtype)
+        if self.ucg_rate > 0. and not no_dropout:
+            out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out
+        return out
+
+
+class FrozenOpenCLIPEmbedder(AbstractEncoder):
+    """
+    Uses the OpenCLIP transformer encoder for text
+    """
+    LAYERS = [
+        # "pooled",
+        "last",
+        "penultimate"
+    ]
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
+                 freeze=True, layer="last"):
+        super().__init__()
+        assert layer in self.LAYERS
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version)
+        del model.visual
+        self.model = model
+
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "last":
+            self.layer_idx = 0
+        elif self.layer == "penultimate":
+            self.layer_idx = 1
+        else:
+            raise NotImplementedError()
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        tokens = open_clip.tokenize(text) ## all clip models use 77 as context length
+        z = self.encode_with_transformer(tokens.to(self.device))
+        return z
+
+    def encode_with_transformer(self, text):
+        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
+        x = x + self.model.positional_embedding
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.model.ln_final(x)
+        return x
+
+    def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
+        for i, r in enumerate(self.model.transformer.resblocks):
+            if i == len(self.model.transformer.resblocks) - self.layer_idx:
+                break
+            if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint(r, x, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenOpenCLIPImageEmbedder(AbstractEncoder):
+    """
+    Uses the OpenCLIP vision transformer encoder for images
+    """
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
+                 freeze=True, layer="pooled", antialias=True, ucg_rate=0.):
+        super().__init__()
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
+                                                            pretrained=version, )
+        del model.transformer
+        self.model = model
+        # self.mapper = torch.nn.Linear(1280, 1024)
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "penultimate":
+            raise NotImplementedError()
+            self.layer_idx = 1
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.ucg_rate = ucg_rate
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+    
+    @autocast
+    def forward(self, image, no_dropout=False):
+        z = self.encode_with_vision_transformer(image)
+        if self.ucg_rate > 0. and not no_dropout:
+            z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z
+        return z
+
+    def encode_with_vision_transformer(self, img):
+        img = self.preprocess(img)
+        x = self.model.visual(img)
+        return x
+
+    def encode(self, text):
+        return self(text)
+
+class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder):
+    """
+    Uses the OpenCLIP vision transformer encoder for images
+    """
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda",
+                 freeze=True, layer="pooled", antialias=True):
+        super().__init__()
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
+                                                            pretrained=version, )
+        del model.transformer
+        self.model = model
+        self.device = device
+
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "penultimate":
+            raise NotImplementedError()
+            self.layer_idx = 1
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+
+    def forward(self, image, no_dropout=False): 
+        ## image: b c h w
+        z = self.encode_with_vision_transformer(image)
+        return z
+
+    def encode_with_vision_transformer(self, x):
+        x = self.preprocess(x)
+
+        # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
+        if self.model.visual.input_patchnorm:
+            # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
+            x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1])
+            x = x.permute(0, 2, 4, 1, 3, 5)
+            x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1)
+            x = self.model.visual.patchnorm_pre_ln(x)
+            x = self.model.visual.conv1(x)
+        else:
+            x = self.model.visual.conv1(x)  # shape = [*, width, grid, grid]
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+        # class embeddings and positional embeddings
+        x = torch.cat(
+            [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
+             x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.model.visual.positional_embedding.to(x.dtype)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        x = self.model.visual.patch_dropout(x)
+        x = self.model.visual.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.model.visual.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        return x
+
+class FrozenCLIPT5Encoder(AbstractEncoder):
+    def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
+                 clip_max_length=77, t5_max_length=77):
+        super().__init__()
+        self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
+        self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
+        print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
+              f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.")
+
+    def encode(self, text):
+        return self(text)
+
+    def forward(self, text):
+        clip_z = self.clip_encoder.encode(text)
+        t5_z = self.t5_encoder.encode(text)
+        return [clip_z, t5_z]

lvdm/modules/encoders/resampler.py

@@ -0,0 +1,145 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+# and https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/resampler.py
+import math
+import torch
+import torch.nn as nn
+
+
+class ImageProjModel(nn.Module):
+    """Projection Model"""
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()        
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = nn.LayerNorm(cross_attention_dim)
+        
+    def forward(self, image_embeds):
+        #embeds = image_embeds
+        embeds = image_embeds.type(list(self.proj.parameters())[0].dtype)
+        clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim)
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+    
+    
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    #(bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+        
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+        
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+        
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        video_length=None, # using frame-wise version or not
+    ):
+        super().__init__()
+        ## queries for a single frame / image
+        self.num_queries = num_queries 
+        self.video_length = video_length
+
+        ## <num_queries> queries for each frame
+        if video_length is not None: 
+            num_queries = num_queries * video_length
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+        self.proj_in = nn.Linear(embedding_dim, dim)
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+        
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1) ## B (T L) C
+        x = self.proj_in(x)
+        
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+            
+        latents = self.proj_out(latents)
+        latents = self.norm_out(latents) # B L C or B (T L) C
+
+        return latents

lvdm/modules/networks/ae_modules.py

@@ -0,0 +1,856 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+
+import torch
+import numpy as np
+import torch.nn as nn
+from einops import rearrange
+
+from utils.utils import instantiate_from_config
+from lvdm.modules.attention import LinearAttention
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w) # bcl
+        q = q.permute(0,2,1)   # bcl -> blc l=hw
+        k = k.reshape(b,c,h*w) # bcl
+        
+        w_ = torch.bmm(q,k)    # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    #print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+ 
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        self.in_channels = in_channels
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        self.in_channels = in_channels
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, return_hidden_states=False):
+        # timestep embedding
+        temb = None
+
+        # print(f'encoder-input={x.shape}')
+        # downsampling
+        hs = [self.conv_in(x)]
+
+        ## if we return hidden states for decoder usage, we will store them in a list
+        if return_hidden_states:
+            hidden_states = []
+        # print(f'encoder-conv in feat={hs[0].shape}')
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                # print(f'encoder-down feat={h.shape}')
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if return_hidden_states:
+                hidden_states.append(h)
+            if i_level != self.num_resolutions-1:
+                # print(f'encoder-downsample (input)={hs[-1].shape}')
+                hs.append(self.down[i_level].downsample(hs[-1]))
+                # print(f'encoder-downsample (output)={hs[-1].shape}')
+        if return_hidden_states:
+            hidden_states.append(hs[0])
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        # print(f'encoder-mid1 feat={h.shape}')
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+        # print(f'encoder-mid2 feat={h.shape}')
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        # print(f'end feat={h.shape}')
+        if return_hidden_states:
+            return h, hidden_states
+        else:
+            return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("AE working on z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # print(f'decoder-input={z.shape}')
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+        # print(f'decoder-conv in feat={h.shape}')
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+        # print(f'decoder-mid feat={h.shape}')
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+                # print(f'decoder-up feat={h.shape}')
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+                # print(f'decoder-upsample feat={h.shape}')
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        # print(f'decoder-conv_out feat={h.shape}')
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z

lvdm/modules/networks/openaimodel3d.py

@@ -0,0 +1,603 @@
+from functools import partial
+from abc import abstractmethod
+import torch
+import torch.nn as nn
+from einops import rearrange
+import torch.nn.functional as F
+from lvdm.models.utils_diffusion import timestep_embedding
+from lvdm.common import checkpoint
+from lvdm.basics import (
+    zero_module,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    normalization
+)
+from lvdm.modules.attention import SpatialTransformer, TemporalTransformer
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, batch_size=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb, batch_size=batch_size)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            elif isinstance(layer, TemporalTransformer):
+                x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size)
+                x = layer(x, context)
+                x = rearrange(x, 'b c f h w -> (b f) c h w')
+            else:
+                x = layer(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest')
+        else:
+            x = F.interpolate(x, scale_factor=2, mode='nearest')
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    :param use_temporal_conv: if True, use the temporal convolution.
+    :param use_image_dataset: if True, the temporal parameters will not be optimized.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        use_conv=False,
+        up=False,
+        down=False,
+        use_temporal_conv=False,
+        tempspatial_aware=False
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.use_temporal_conv = use_temporal_conv
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1)
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+        if self.use_temporal_conv:
+            self.temopral_conv = TemporalConvBlock(
+                self.out_channels,
+                self.out_channels,
+                dropout=0.1,
+                spatial_aware=tempspatial_aware
+            )
+
+    def forward(self, x, emb, batch_size=None):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        input_tuple = (x, emb)
+        if batch_size:
+            forward_batchsize = partial(self._forward, batch_size=batch_size)
+            return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint)
+
+    def _forward(self, x, emb, batch_size=None):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        h = self.skip_connection(x) + h
+
+        if self.use_temporal_conv and batch_size:
+            h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size)
+            h = self.temopral_conv(h)
+            h = rearrange(h, 'b c t h w -> (b t) c h w')
+        return h
+
+
+class TemporalConvBlock(nn.Module):
+    """
+    Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py
+    """
+    def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False):
+        super(TemporalConvBlock, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        th_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 1)
+        th_padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 0)
+        tw_kernel_shape = (3, 1, 1) if not spatial_aware else (3, 1, 3)
+        tw_padding_shape = (1, 0, 0) if not spatial_aware else (1, 0, 1)
+
+        # conv layers
+        self.conv1 = nn.Sequential(
+            nn.GroupNorm(32, in_channels), nn.SiLU(),
+            nn.Conv3d(in_channels, out_channels, th_kernel_shape, padding=th_padding_shape))
+        self.conv2 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape))
+        self.conv3 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, th_kernel_shape, padding=th_padding_shape))
+        self.conv4 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, tw_kernel_shape, padding=tw_padding_shape))
+
+        # zero out the last layer params,so the conv block is identity
+        nn.init.zeros_(self.conv4[-1].weight)
+        nn.init.zeros_(self.conv4[-1].bias)
+
+    def forward(self, x):
+        identity = x
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        return identity + x
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: in_channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 model_channels,
+                 out_channels,
+                 num_res_blocks,
+                 attention_resolutions,
+                 dropout=0.0,
+                 channel_mult=(1, 2, 4, 8),
+                 conv_resample=True,
+                 dims=2,
+                 context_dim=None,
+                 use_scale_shift_norm=False,
+                 resblock_updown=False,
+                 num_heads=-1,
+                 num_head_channels=-1,
+                 transformer_depth=1,
+                 use_linear=False,
+                 use_checkpoint=False,
+                 temporal_conv=False,
+                 tempspatial_aware=False,
+                 temporal_attention=True,
+                 use_relative_position=True,
+                 use_causal_attention=False,
+                 temporal_length=None,
+                 use_fp16=False,
+                 addition_attention=False,
+                 temporal_selfatt_only=True,
+                 image_cross_attention=False,
+                 image_cross_attention_scale_learnable=False,
+                 default_fs=4,
+                 fs_condition=False,
+                ):
+        super(UNetModel, self).__init__()
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.temporal_attention = temporal_attention
+        time_embed_dim = model_channels * 4
+        self.use_checkpoint = use_checkpoint
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+        temporal_self_att_only = True
+        self.addition_attention = addition_attention
+        self.temporal_length = temporal_length
+        self.image_cross_attention = image_cross_attention
+        self.image_cross_attention_scale_learnable = image_cross_attention_scale_learnable
+        self.default_fs = default_fs
+        self.fs_condition = fs_condition
+
+        ## Time embedding blocks
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+        if fs_condition:
+            self.fps_embedding = nn.Sequential(
+                linear(model_channels, time_embed_dim),
+                nn.SiLU(),
+                linear(time_embed_dim, time_embed_dim),
+            )
+            nn.init.zeros_(self.fps_embedding[-1].weight)
+            nn.init.zeros_(self.fps_embedding[-1].bias)
+        ## Input Block
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))
+            ]
+        )
+        if self.addition_attention:
+            self.init_attn=TimestepEmbedSequential(
+                TemporalTransformer(
+                    model_channels,
+                    n_heads=8,
+                    d_head=num_head_channels,
+                    depth=transformer_depth,
+                    context_dim=context_dim,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                    causal_attention=False, relative_position=use_relative_position, 
+                    temporal_length=temporal_length))
+
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(ch, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False, 
+                            video_length=temporal_length, image_cross_attention=self.image_cross_attention,
+                            image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable,                      
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(ch, time_embed_dim, dropout, 
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True
+                        )
+                        if resblock_updown
+                        else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        layers = [
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                use_temporal_conv=temporal_conv
+            ),
+            SpatialTransformer(ch, num_heads, dim_head, 
+                depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length, 
+                image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable                
+            )
+        ]
+        if self.temporal_attention:
+            layers.append(
+                TemporalTransformer(ch, num_heads, dim_head,
+                    depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, 
+                    causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                    temporal_length=temporal_length
+                )
+            )
+        layers.append(
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware, 
+                use_temporal_conv=temporal_conv
+                )
+        )
+
+        ## Middle Block
+        self.middle_block = TimestepEmbedSequential(*layers)
+
+        ## Output Block
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(ch + ich, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False, video_length=temporal_length,
+                            image_cross_attention=self.image_cross_attention,image_cross_attention_scale_learnable=self.image_cross_attention_scale_learnable    
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_self_att_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(ch, time_embed_dim, dropout,
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+
+    def forward(self, x, timesteps, context=None, features_adapter=None, fs=None, **kwargs):
+        b,_,t,_,_ = x.shape
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).type(x.dtype)
+        emb = self.time_embed(t_emb)
+        
+        ## repeat t times for context [(b t) 77 768] & time embedding
+        ## check if we use per-frame image conditioning
+        _, l_context, _ = context.shape
+        if l_context == 77 + t*16: ## !!! HARD CODE here
+            context_text, context_img = context[:,:77,:], context[:,77:,:]
+            context_text = context_text.repeat_interleave(repeats=t, dim=0)
+            context_img = rearrange(context_img, 'b (t l) c -> (b t) l c', t=t)
+            context = torch.cat([context_text, context_img], dim=1)
+        else:
+            context = context.repeat_interleave(repeats=t, dim=0)
+        emb = emb.repeat_interleave(repeats=t, dim=0)
+        
+        ## always in shape (b t) c h w, except for temporal layer
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+
+        ## combine emb
+        if self.fs_condition:
+            if fs is None:
+                fs = torch.tensor(
+                    [self.default_fs] * b, dtype=torch.long, device=x.device)
+            fs_emb = timestep_embedding(fs, self.model_channels, repeat_only=False).type(x.dtype)
+
+            fs_embed = self.fps_embedding(fs_emb)
+            fs_embed = fs_embed.repeat_interleave(repeats=t, dim=0)
+            emb = emb + fs_embed
+
+        h = x.type(self.dtype)
+        adapter_idx = 0
+        hs = []
+        for id, module in enumerate(self.input_blocks):
+            h = module(h, emb, context=context, batch_size=b)
+            if id ==0 and self.addition_attention:
+                h = self.init_attn(h, emb, context=context, batch_size=b)
+            ## plug-in adapter features
+            if ((id+1)%3 == 0) and features_adapter is not None:
+                h = h + features_adapter[adapter_idx]
+                adapter_idx += 1
+            hs.append(h)
+        if features_adapter is not None:
+            assert len(features_adapter)==adapter_idx, 'Wrong features_adapter'
+
+        h = self.middle_block(h, emb, context=context, batch_size=b)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context=context, batch_size=b)
+        h = h.type(x.dtype)
+        y = self.out(h)
+        
+        # reshape back to (b c t h w)
+        y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
+        return y

lvdm/modules/x_transformer.py

@@ -0,0 +1,639 @@
+"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
+from functools import partial
+from inspect import isfunction
+from collections import namedtuple
+from einops import rearrange, repeat
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+# constants
+DEFAULT_DIM_HEAD = 64
+
+Intermediates = namedtuple('Intermediates', [
+    'pre_softmax_attn',
+    'post_softmax_attn'
+])
+
+LayerIntermediates = namedtuple('Intermediates', [
+    'hiddens',
+    'attn_intermediates'
+])
+
+
+class AbsolutePositionalEmbedding(nn.Module):
+    def __init__(self, dim, max_seq_len):
+        super().__init__()
+        self.emb = nn.Embedding(max_seq_len, dim)
+        self.init_()
+
+    def init_(self):
+        nn.init.normal_(self.emb.weight, std=0.02)
+
+    def forward(self, x):
+        n = torch.arange(x.shape[1], device=x.device)
+        return self.emb(n)[None, :, :]
+
+
+class FixedPositionalEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+
+    def forward(self, x, seq_dim=1, offset=0):
+        t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
+        sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
+        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
+        return emb[None, :, :]
+
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def always(val):
+    def inner(*args, **kwargs):
+        return val
+    return inner
+
+
+def not_equals(val):
+    def inner(x):
+        return x != val
+    return inner
+
+
+def equals(val):
+    def inner(x):
+        return x == val
+    return inner
+
+
+def max_neg_value(tensor):
+    return -torch.finfo(tensor.dtype).max
+
+
+# keyword argument helpers
+
+def pick_and_pop(keys, d):
+    values = list(map(lambda key: d.pop(key), keys))
+    return dict(zip(keys, values))
+
+
+def group_dict_by_key(cond, d):
+    return_val = [dict(), dict()]
+    for key in d.keys():
+        match = bool(cond(key))
+        ind = int(not match)
+        return_val[ind][key] = d[key]
+    return (*return_val,)
+
+
+def string_begins_with(prefix, str):
+    return str.startswith(prefix)
+
+
+def group_by_key_prefix(prefix, d):
+    return group_dict_by_key(partial(string_begins_with, prefix), d)
+
+
+def groupby_prefix_and_trim(prefix, d):
+    kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
+    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
+    return kwargs_without_prefix, kwargs
+
+
+# classes
+class Scale(nn.Module):
+    def __init__(self, value, fn):
+        super().__init__()
+        self.value = value
+        self.fn = fn
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.value, *rest)
+
+
+class Rezero(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+        self.g = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.g, *rest)
+
+
+class ScaleNorm(nn.Module):
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-8):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class Residual(nn.Module):
+    def forward(self, x, residual):
+        return x + residual
+
+
+class GRUGating(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.gru = nn.GRUCell(dim, dim)
+
+    def forward(self, x, residual):
+        gated_output = self.gru(
+            rearrange(x, 'b n d -> (b n) d'),
+            rearrange(residual, 'b n d -> (b n) d')
+        )
+
+        return gated_output.reshape_as(x)
+
+
+# feedforward
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+# attention.
+class Attention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            dim_head=DEFAULT_DIM_HEAD,
+            heads=8,
+            causal=False,
+            mask=None,
+            talking_heads=False,
+            sparse_topk=None,
+            use_entmax15=False,
+            num_mem_kv=0,
+            dropout=0.,
+            on_attn=False
+    ):
+        super().__init__()
+        if use_entmax15:
+            raise NotImplementedError("Check out entmax activation instead of softmax activation!")
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        self.causal = causal
+        self.mask = mask
+
+        inner_dim = dim_head * heads
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(dim, inner_dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+        # talking heads
+        self.talking_heads = talking_heads
+        if talking_heads:
+            self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+            self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+
+        # explicit topk sparse attention
+        self.sparse_topk = sparse_topk
+
+        # entmax
+        #self.attn_fn = entmax15 if use_entmax15 else F.softmax
+        self.attn_fn = F.softmax
+
+        # add memory key / values
+        self.num_mem_kv = num_mem_kv
+        if num_mem_kv > 0:
+            self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+            self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+
+        # attention on attention
+        self.attn_on_attn = on_attn
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            rel_pos=None,
+            sinusoidal_emb=None,
+            prev_attn=None,
+            mem=None
+    ):
+        b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
+        kv_input = default(context, x)
+
+        q_input = x
+        k_input = kv_input
+        v_input = kv_input
+
+        if exists(mem):
+            k_input = torch.cat((mem, k_input), dim=-2)
+            v_input = torch.cat((mem, v_input), dim=-2)
+
+        if exists(sinusoidal_emb):
+            # in shortformer, the query would start at a position offset depending on the past cached memory
+            offset = k_input.shape[-2] - q_input.shape[-2]
+            q_input = q_input + sinusoidal_emb(q_input, offset=offset)
+            k_input = k_input + sinusoidal_emb(k_input)
+
+        q = self.to_q(q_input)
+        k = self.to_k(k_input)
+        v = self.to_v(v_input)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
+
+        input_mask = None
+        if any(map(exists, (mask, context_mask))):
+            q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
+            k_mask = q_mask if not exists(context) else context_mask
+            k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
+            q_mask = rearrange(q_mask, 'b i -> b () i ()')
+            k_mask = rearrange(k_mask, 'b j -> b () () j')
+            input_mask = q_mask * k_mask
+
+        if self.num_mem_kv > 0:
+            mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
+            k = torch.cat((mem_k, k), dim=-2)
+            v = torch.cat((mem_v, v), dim=-2)
+            if exists(input_mask):
+                input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+        mask_value = max_neg_value(dots)
+
+        if exists(prev_attn):
+            dots = dots + prev_attn
+
+        pre_softmax_attn = dots
+
+        if talking_heads:
+            dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
+
+        if exists(rel_pos):
+            dots = rel_pos(dots)
+
+        if exists(input_mask):
+            dots.masked_fill_(~input_mask, mask_value)
+            del input_mask
+
+        if self.causal:
+            i, j = dots.shape[-2:]
+            r = torch.arange(i, device=device)
+            mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
+            mask = F.pad(mask, (j - i, 0), value=False)
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
+            top, _ = dots.topk(self.sparse_topk, dim=-1)
+            vk = top[..., -1].unsqueeze(-1).expand_as(dots)
+            mask = dots < vk
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        attn = self.attn_fn(dots, dim=-1)
+        post_softmax_attn = attn
+
+        attn = self.dropout(attn)
+
+        if talking_heads:
+            attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
+
+        out = einsum('b h i j, b h j d -> b h i d', attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+
+        intermediates = Intermediates(
+            pre_softmax_attn=pre_softmax_attn,
+            post_softmax_attn=post_softmax_attn
+        )
+
+        return self.to_out(out), intermediates
+
+
+class AttentionLayers(nn.Module):
+    def __init__(
+            self,
+            dim,
+            depth,
+            heads=8,
+            causal=False,
+            cross_attend=False,
+            only_cross=False,
+            use_scalenorm=False,
+            use_rmsnorm=False,
+            use_rezero=False,
+            rel_pos_num_buckets=32,
+            rel_pos_max_distance=128,
+            position_infused_attn=False,
+            custom_layers=None,
+            sandwich_coef=None,
+            par_ratio=None,
+            residual_attn=False,
+            cross_residual_attn=False,
+            macaron=False,
+            pre_norm=True,
+            gate_residual=False,
+            **kwargs
+    ):
+        super().__init__()
+        ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
+        attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
+
+        dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
+
+        self.dim = dim
+        self.depth = depth
+        self.layers = nn.ModuleList([])
+
+        self.has_pos_emb = position_infused_attn
+        self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
+        self.rotary_pos_emb = always(None)
+
+        assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
+        self.rel_pos = None
+
+        self.pre_norm = pre_norm
+
+        self.residual_attn = residual_attn
+        self.cross_residual_attn = cross_residual_attn
+
+        norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
+        norm_class = RMSNorm if use_rmsnorm else norm_class
+        norm_fn = partial(norm_class, dim)
+
+        norm_fn = nn.Identity if use_rezero else norm_fn
+        branch_fn = Rezero if use_rezero else None
+
+        if cross_attend and not only_cross:
+            default_block = ('a', 'c', 'f')
+        elif cross_attend and only_cross:
+            default_block = ('c', 'f')
+        else:
+            default_block = ('a', 'f')
+
+        if macaron:
+            default_block = ('f',) + default_block
+
+        if exists(custom_layers):
+            layer_types = custom_layers
+        elif exists(par_ratio):
+            par_depth = depth * len(default_block)
+            assert 1 < par_ratio <= par_depth, 'par ratio out of range'
+            default_block = tuple(filter(not_equals('f'), default_block))
+            par_attn = par_depth // par_ratio
+            depth_cut = par_depth * 2 // 3  # 2 / 3 attention layer cutoff suggested by PAR paper
+            par_width = (depth_cut + depth_cut // par_attn) // par_attn
+            assert len(default_block) <= par_width, 'default block is too large for par_ratio'
+            par_block = default_block + ('f',) * (par_width - len(default_block))
+            par_head = par_block * par_attn
+            layer_types = par_head + ('f',) * (par_depth - len(par_head))
+        elif exists(sandwich_coef):
+            assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
+            layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
+        else:
+            layer_types = default_block * depth
+
+        self.layer_types = layer_types
+        self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
+
+        for layer_type in self.layer_types:
+            if layer_type == 'a':
+                layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
+            elif layer_type == 'c':
+                layer = Attention(dim, heads=heads, **attn_kwargs)
+            elif layer_type == 'f':
+                layer = FeedForward(dim, **ff_kwargs)
+                layer = layer if not macaron else Scale(0.5, layer)
+            else:
+                raise Exception(f'invalid layer type {layer_type}')
+
+            if isinstance(layer, Attention) and exists(branch_fn):
+                layer = branch_fn(layer)
+
+            if gate_residual:
+                residual_fn = GRUGating(dim)
+            else:
+                residual_fn = Residual()
+
+            self.layers.append(nn.ModuleList([
+                norm_fn(),
+                layer,
+                residual_fn
+            ]))
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            mems=None,
+            return_hiddens=False
+    ):
+        hiddens = []
+        intermediates = []
+        prev_attn = None
+        prev_cross_attn = None
+
+        mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
+
+        for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
+            is_last = ind == (len(self.layers) - 1)
+
+            if layer_type == 'a':
+                hiddens.append(x)
+                layer_mem = mems.pop(0)
+
+            residual = x
+
+            if self.pre_norm:
+                x = norm(x)
+
+            if layer_type == 'a':
+                out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
+                                   prev_attn=prev_attn, mem=layer_mem)
+            elif layer_type == 'c':
+                out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
+            elif layer_type == 'f':
+                out = block(x)
+
+            x = residual_fn(out, residual)
+
+            if layer_type in ('a', 'c'):
+                intermediates.append(inter)
+
+            if layer_type == 'a' and self.residual_attn:
+                prev_attn = inter.pre_softmax_attn
+            elif layer_type == 'c' and self.cross_residual_attn:
+                prev_cross_attn = inter.pre_softmax_attn
+
+            if not self.pre_norm and not is_last:
+                x = norm(x)
+
+        if return_hiddens:
+            intermediates = LayerIntermediates(
+                hiddens=hiddens,
+                attn_intermediates=intermediates
+            )
+
+            return x, intermediates
+
+        return x
+
+
+class Encoder(AttentionLayers):
+    def __init__(self, **kwargs):
+        assert 'causal' not in kwargs, 'cannot set causality on encoder'
+        super().__init__(causal=False, **kwargs)
+
+
+
+class TransformerWrapper(nn.Module):
+    def __init__(
+            self,
+            *,
+            num_tokens,
+            max_seq_len,
+            attn_layers,
+            emb_dim=None,
+            max_mem_len=0.,
+            emb_dropout=0.,
+            num_memory_tokens=None,
+            tie_embedding=False,
+            use_pos_emb=True
+    ):
+        super().__init__()
+        assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
+
+        dim = attn_layers.dim
+        emb_dim = default(emb_dim, dim)
+
+        self.max_seq_len = max_seq_len
+        self.max_mem_len = max_mem_len
+        self.num_tokens = num_tokens
+
+        self.token_emb = nn.Embedding(num_tokens, emb_dim)
+        self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
+                    use_pos_emb and not attn_layers.has_pos_emb) else always(0)
+        self.emb_dropout = nn.Dropout(emb_dropout)
+
+        self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
+        self.attn_layers = attn_layers
+        self.norm = nn.LayerNorm(dim)
+
+        self.init_()
+
+        self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
+
+        # memory tokens (like [cls]) from Memory Transformers paper
+        num_memory_tokens = default(num_memory_tokens, 0)
+        self.num_memory_tokens = num_memory_tokens
+        if num_memory_tokens > 0:
+            self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
+
+            # let funnel encoder know number of memory tokens, if specified
+            if hasattr(attn_layers, 'num_memory_tokens'):
+                attn_layers.num_memory_tokens = num_memory_tokens
+
+    def init_(self):
+        nn.init.normal_(self.token_emb.weight, std=0.02)
+
+    def forward(
+            self,
+            x,
+            return_embeddings=False,
+            mask=None,
+            return_mems=False,
+            return_attn=False,
+            mems=None,
+            **kwargs
+    ):
+        b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
+        x = self.token_emb(x)
+        x += self.pos_emb(x)
+        x = self.emb_dropout(x)
+
+        x = self.project_emb(x)
+
+        if num_mem > 0:
+            mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
+            x = torch.cat((mem, x), dim=1)
+
+            # auto-handle masking after appending memory tokens
+            if exists(mask):
+                mask = F.pad(mask, (num_mem, 0), value=True)
+
+        x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
+        x = self.norm(x)
+
+        mem, x = x[:, :num_mem], x[:, num_mem:]
+
+        out = self.to_logits(x) if not return_embeddings else x
+
+        if return_mems:
+            hiddens = intermediates.hiddens
+            new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
+            new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
+            return out, new_mems
+
+        if return_attn:
+            attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
+            return out, attn_maps
+
+        return out

main/callbacks.py

@@ -0,0 +1,133 @@
+import os
+import time
+import logging
+mainlogger = logging.getLogger('mainlogger')
+
+import torch
+import torchvision
+import pytorch_lightning as pl
+from pytorch_lightning.callbacks import Callback
+from pytorch_lightning.utilities import rank_zero_only
+from pytorch_lightning.utilities import rank_zero_info
+from utils.save_video import log_local, prepare_to_log
+
+
+class ImageLogger(Callback):
+    def __init__(self, batch_frequency, max_images=8, clamp=True, rescale=True, save_dir=None, \
+                to_local=False, log_images_kwargs=None):
+        super().__init__()
+        self.rescale = rescale
+        self.batch_freq = batch_frequency
+        self.max_images = max_images
+        self.to_local = to_local
+        self.clamp = clamp
+        self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {}
+        if self.to_local:
+            ## default save dir
+            self.save_dir = os.path.join(save_dir, "images")
+            os.makedirs(os.path.join(self.save_dir, "train"), exist_ok=True)
+            os.makedirs(os.path.join(self.save_dir, "val"), exist_ok=True)
+
+    def log_to_tensorboard(self, pl_module, batch_logs, filename, split, save_fps=8):
+        """ log images and videos to tensorboard """        
+        global_step = pl_module.global_step
+        for key in batch_logs:
+            value = batch_logs[key]
+            tag = "gs%d-%s/%s-%s"%(global_step, split, filename, key)
+            if isinstance(value, list) and isinstance(value[0], str):
+                captions = ' |------| '.join(value)
+                pl_module.logger.experiment.add_text(tag, captions, global_step=global_step)
+            elif isinstance(value, torch.Tensor) and value.dim() == 5:
+                video = value
+                n = video.shape[0]
+                video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+                frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0) for framesheet in video] #[3, n*h, 1*w]
+                grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
+                grid = (grid + 1.0) / 2.0
+                grid = grid.unsqueeze(dim=0)
+                pl_module.logger.experiment.add_video(tag, grid, fps=save_fps, global_step=global_step)
+            elif isinstance(value, torch.Tensor) and value.dim() == 4:
+                img = value
+                grid = torchvision.utils.make_grid(img, nrow=int(n), padding=0)
+                grid = (grid + 1.0) / 2.0  # -1,1 -> 0,1; c,h,w
+                pl_module.logger.experiment.add_image(tag, grid, global_step=global_step)
+            else:
+                pass
+
+    @rank_zero_only
+    def log_batch_imgs(self, pl_module, batch, batch_idx, split="train"):
+        """ generate images, then save and log to tensorboard """
+        skip_freq = self.batch_freq if split == "train" else 5
+        if (batch_idx+1) % skip_freq == 0:
+            is_train = pl_module.training
+            if is_train:
+                pl_module.eval()
+            torch.cuda.empty_cache()
+            with torch.no_grad():
+                log_func = pl_module.log_images
+                batch_logs = log_func(batch, split=split, **self.log_images_kwargs)
+            
+            ## process: move to CPU and clamp
+            batch_logs = prepare_to_log(batch_logs, self.max_images, self.clamp)
+            torch.cuda.empty_cache()
+            
+            filename = "ep{}_idx{}_rank{}".format(
+                pl_module.current_epoch,
+                batch_idx,
+                pl_module.global_rank)
+            if self.to_local:
+                mainlogger.info("Log [%s] batch <%s> to local ..."%(split, filename))
+                filename = "gs{}_".format(pl_module.global_step) + filename
+                log_local(batch_logs, os.path.join(self.save_dir, split), filename, save_fps=10)
+            else:
+                mainlogger.info("Log [%s] batch <%s> to tensorboard ..."%(split, filename))
+                self.log_to_tensorboard(pl_module, batch_logs, filename, split, save_fps=10)
+            mainlogger.info('Finish!')
+
+            if is_train:
+                pl_module.train()
+
+    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=None):
+        if self.batch_freq != -1 and pl_module.logdir:
+            self.log_batch_imgs(pl_module, batch, batch_idx, split="train")
+
+    def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=None):
+        ## different with validation_step() that saving the whole validation set and only keep the latest,
+        ## it records the performance of every validation (without overwritten) by only keep a subset
+        if self.batch_freq != -1 and pl_module.logdir:
+            self.log_batch_imgs(pl_module, batch, batch_idx, split="val")
+        if hasattr(pl_module, 'calibrate_grad_norm'):
+            if (pl_module.calibrate_grad_norm and batch_idx % 25 == 0) and batch_idx > 0:
+                self.log_gradients(trainer, pl_module, batch_idx=batch_idx)
+
+
+class CUDACallback(Callback):
+    # see https://github.com/SeanNaren/minGPT/blob/master/mingpt/callback.py
+    def on_train_epoch_start(self, trainer, pl_module):
+        # Reset the memory use counter
+        # lightning update
+        if int((pl.__version__).split('.')[1])>=7:
+            gpu_index = trainer.strategy.root_device.index
+        else:
+            gpu_index = trainer.root_gpu
+        torch.cuda.reset_peak_memory_stats(gpu_index)
+        torch.cuda.synchronize(gpu_index)
+        self.start_time = time.time()
+
+    def on_train_epoch_end(self, trainer, pl_module):
+        if int((pl.__version__).split('.')[1])>=7:
+            gpu_index = trainer.strategy.root_device.index
+        else:
+            gpu_index = trainer.root_gpu
+        torch.cuda.synchronize(gpu_index)
+        max_memory = torch.cuda.max_memory_allocated(gpu_index) / 2 ** 20
+        epoch_time = time.time() - self.start_time
+
+        try:
+            max_memory = trainer.training_type_plugin.reduce(max_memory)
+            epoch_time = trainer.training_type_plugin.reduce(epoch_time)
+
+            rank_zero_info(f"Average Epoch time: {epoch_time:.2f} seconds")
+            rank_zero_info(f"Average Peak memory {max_memory:.2f}MiB")
+        except AttributeError:
+            pass

main/trainer.py

@@ -0,0 +1,168 @@
+import argparse, os, sys, datetime
+from omegaconf import OmegaConf
+from transformers import logging as transf_logging
+import pytorch_lightning as pl
+from pytorch_lightning import seed_everything
+from pytorch_lightning.trainer import Trainer
+import torch
+sys.path.insert(1, os.path.join(sys.path[0], '..'))
+from utils.utils import instantiate_from_config
+from utils_train import get_trainer_callbacks, get_trainer_logger, get_trainer_strategy
+from utils_train import set_logger, init_workspace, load_checkpoints
+
+
+def get_parser(**parser_kwargs):
+    parser = argparse.ArgumentParser(**parser_kwargs)
+    parser.add_argument("--seed", "-s", type=int, default=20230211, help="seed for seed_everything")
+    parser.add_argument("--name", "-n", type=str, default="", help="experiment name, as saving folder")
+
+    parser.add_argument("--base", "-b", nargs="*", metavar="base_config.yaml", help="paths to base configs. Loaded from left-to-right. "
+                            "Parameters can be overwritten or added with command-line options of the form `--key value`.", default=list())
+    
+    parser.add_argument("--train", "-t", action='store_true', default=False, help='train')
+    parser.add_argument("--val", "-v", action='store_true', default=False, help='val')
+    parser.add_argument("--test", action='store_true', default=False, help='test')
+
+    parser.add_argument("--logdir", "-l", type=str, default="logs", help="directory for logging dat shit")
+    parser.add_argument("--auto_resume", action='store_true', default=False, help="resume from full-info checkpoint")
+    parser.add_argument("--auto_resume_weight_only", action='store_true', default=False, help="resume from weight-only checkpoint")
+    parser.add_argument("--debug", "-d", action='store_true', default=False, help="enable post-mortem debugging")
+
+    return parser
+    
+def get_nondefault_trainer_args(args):
+    parser = argparse.ArgumentParser()
+    parser = Trainer.add_argparse_args(parser)
+    default_trainer_args = parser.parse_args([])
+    return sorted(k for k in vars(default_trainer_args) if getattr(args, k) != getattr(default_trainer_args, k))
+
+
+if __name__ == "__main__":
+    now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
+    local_rank = int(os.environ.get('LOCAL_RANK'))
+    global_rank = int(os.environ.get('RANK'))
+    num_rank = int(os.environ.get('WORLD_SIZE'))
+
+    parser = get_parser()
+    ## Extends existing argparse by default Trainer attributes
+    parser = Trainer.add_argparse_args(parser)
+    args, unknown = parser.parse_known_args()
+    ## disable transformer warning
+    transf_logging.set_verbosity_error()
+    seed_everything(args.seed)
+
+    ## yaml configs: "model" | "data" | "lightning"
+    configs = [OmegaConf.load(cfg) for cfg in args.base]
+    cli = OmegaConf.from_dotlist(unknown)
+    config = OmegaConf.merge(*configs, cli)
+    lightning_config = config.pop("lightning", OmegaConf.create())
+    trainer_config = lightning_config.get("trainer", OmegaConf.create()) 
+
+    ## setup workspace directories
+    workdir, ckptdir, cfgdir, loginfo = init_workspace(args.name, args.logdir, config, lightning_config, global_rank)
+    logger = set_logger(logfile=os.path.join(loginfo, 'log_%d:%s.txt'%(global_rank, now)))
+    logger.info("@lightning version: %s [>=1.8 required]"%(pl.__version__))  
+
+    ## MODEL CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
+    logger.info("***** Configing Model *****")
+    config.model.params.logdir = workdir
+    model = instantiate_from_config(config.model)
+
+    ## load checkpoints
+    model = load_checkpoints(model, config.model)
+
+    ## register_schedule again to make ZTSNR work
+    if model.rescale_betas_zero_snr:
+        model.register_schedule(given_betas=model.given_betas, beta_schedule=model.beta_schedule, timesteps=model.timesteps,
+                                linear_start=model.linear_start, linear_end=model.linear_end, cosine_s=model.cosine_s)
+
+    ## update trainer config
+    for k in get_nondefault_trainer_args(args):
+        trainer_config[k] = getattr(args, k)
+        
+    num_nodes = trainer_config.num_nodes
+    ngpu_per_node = trainer_config.devices
+    logger.info(f"Running on {num_rank}={num_nodes}x{ngpu_per_node} GPUs")
+
+    ## setup learning rate
+    base_lr = config.model.base_learning_rate
+    bs = config.data.params.batch_size
+    if getattr(config.model, 'scale_lr', True):
+        model.learning_rate = num_rank * bs * base_lr
+    else:
+        model.learning_rate = base_lr
+
+
+    ## DATA CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
+    logger.info("***** Configing Data *****")
+    data = instantiate_from_config(config.data)
+    data.setup()
+    for k in data.datasets:
+        logger.info(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}")
+
+
+    ## TRAINER CONFIG >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
+    logger.info("***** Configing Trainer *****")
+    if "accelerator" not in trainer_config:
+        trainer_config["accelerator"] = "gpu"
+
+    ## setup trainer args: pl-logger and callbacks
+    trainer_kwargs = dict()
+    trainer_kwargs["num_sanity_val_steps"] = 0
+    logger_cfg = get_trainer_logger(lightning_config, workdir, args.debug)
+    trainer_kwargs["logger"] = instantiate_from_config(logger_cfg)
+    
+    ## setup callbacks
+    callbacks_cfg = get_trainer_callbacks(lightning_config, config, workdir, ckptdir, logger)
+    trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg]
+    strategy_cfg = get_trainer_strategy(lightning_config)
+    trainer_kwargs["strategy"] = strategy_cfg if type(strategy_cfg) == str else instantiate_from_config(strategy_cfg)
+    trainer_kwargs['precision'] = lightning_config.get('precision', 32)
+    trainer_kwargs["sync_batchnorm"] = False
+
+    ## trainer config: others
+
+    trainer_args = argparse.Namespace(**trainer_config)
+    trainer = Trainer.from_argparse_args(trainer_args, **trainer_kwargs)
+
+    ## allow checkpointing via USR1
+    def melk(*args, **kwargs):
+        ## run all checkpoint hooks
+        if trainer.global_rank == 0:
+            print("Summoning checkpoint.")
+            ckpt_path = os.path.join(ckptdir, "last_summoning.ckpt")
+            trainer.save_checkpoint(ckpt_path)
+
+    def divein(*args, **kwargs):
+        if trainer.global_rank == 0:
+            import pudb;
+            pudb.set_trace()
+
+    import signal
+    signal.signal(signal.SIGUSR1, melk)
+    signal.signal(signal.SIGUSR2, divein)
+
+    ## Running LOOP >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
+    logger.info("***** Running the Loop *****")
+    if args.train:
+        try:
+            if "strategy" in lightning_config and lightning_config['strategy'].startswith('deepspeed'):
+                logger.info("<Training in DeepSpeed Mode>")
+                ## deepspeed
+                if trainer_kwargs['precision'] == 16:
+                    with torch.cuda.amp.autocast():
+                        trainer.fit(model, data)
+                else:
+                    trainer.fit(model, data)
+            else:
+                logger.info("<Training in DDPSharded Mode>") ## this is default
+                ## ddpsharded
+                trainer.fit(model, data)
+        except Exception:
+            #melk()
+            raise
+
+    # if args.val:
+    #     trainer.validate(model, data)
+    # if args.test or not trainer.interrupted:
+    #     trainer.test(model, data)

main/utils_data.py

@@ -0,0 +1,136 @@
+from functools import partial
+import numpy as np
+
+import torch
+import pytorch_lightning as pl
+from torch.utils.data import DataLoader, Dataset
+
+import os, sys
+os.chdir(sys.path[0])
+sys.path.append("..")
+from lvdm.data.base import Txt2ImgIterableBaseDataset
+from utils.utils import instantiate_from_config
+
+
+def worker_init_fn(_):
+    worker_info = torch.utils.data.get_worker_info()
+
+    dataset = worker_info.dataset
+    worker_id = worker_info.id
+
+    if isinstance(dataset, Txt2ImgIterableBaseDataset):
+        split_size = dataset.num_records // worker_info.num_workers
+        # reset num_records to the true number to retain reliable length information
+        dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
+        current_id = np.random.choice(len(np.random.get_state()[1]), 1)
+        return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
+    else:
+        return np.random.seed(np.random.get_state()[1][0] + worker_id)
+
+
+class WrappedDataset(Dataset):
+    """Wraps an arbitrary object with __len__ and __getitem__ into a pytorch dataset"""
+
+    def __init__(self, dataset):
+        self.data = dataset
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        return self.data[idx]
+
+
+class DataModuleFromConfig(pl.LightningDataModule):
+    def __init__(self, batch_size, train=None, validation=None, test=None, predict=None,
+                 wrap=False, num_workers=None, shuffle_test_loader=False, use_worker_init_fn=False,
+                 shuffle_val_dataloader=False, train_img=None,
+                 test_max_n_samples=None):
+        super().__init__()
+        self.batch_size = batch_size
+        self.dataset_configs = dict()
+        self.num_workers = num_workers if num_workers is not None else batch_size * 2
+        self.use_worker_init_fn = use_worker_init_fn
+        if train is not None:
+            self.dataset_configs["train"] = train
+            self.train_dataloader = self._train_dataloader
+        if validation is not None:
+            self.dataset_configs["validation"] = validation
+            self.val_dataloader = partial(self._val_dataloader, shuffle=shuffle_val_dataloader)
+        if test is not None:
+            self.dataset_configs["test"] = test
+            self.test_dataloader = partial(self._test_dataloader, shuffle=shuffle_test_loader)
+        if predict is not None:
+            self.dataset_configs["predict"] = predict
+            self.predict_dataloader = self._predict_dataloader
+
+        self.img_loader = None
+        self.wrap = wrap
+        self.test_max_n_samples = test_max_n_samples
+        self.collate_fn = None
+
+    def prepare_data(self):
+        pass
+
+    def setup(self, stage=None):
+        self.datasets = dict((k, instantiate_from_config(self.dataset_configs[k])) for k in self.dataset_configs)
+        if self.wrap:
+            for k in self.datasets:
+                self.datasets[k] = WrappedDataset(self.datasets[k])
+
+    def _train_dataloader(self):
+        is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
+        if is_iterable_dataset or self.use_worker_init_fn:
+            init_fn = worker_init_fn
+        else:
+            init_fn = None
+        loader = DataLoader(self.datasets["train"], batch_size=self.batch_size,
+                          num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True,
+                          worker_init_fn=init_fn, collate_fn=self.collate_fn,
+                          )
+        return loader
+
+    def _val_dataloader(self, shuffle=False):
+        if isinstance(self.datasets['validation'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
+            init_fn = worker_init_fn
+        else:
+            init_fn = None
+        return DataLoader(self.datasets["validation"],
+                          batch_size=self.batch_size,
+                          num_workers=self.num_workers,
+                          worker_init_fn=init_fn,
+                          shuffle=shuffle, 
+                          collate_fn=self.collate_fn,
+                          )
+
+    def _test_dataloader(self, shuffle=False):
+        try:
+            is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
+        except:
+            is_iterable_dataset = isinstance(self.datasets['test'], Txt2ImgIterableBaseDataset)
+
+        if is_iterable_dataset or self.use_worker_init_fn:
+            init_fn = worker_init_fn
+        else:
+            init_fn = None
+
+        # do not shuffle dataloader for iterable dataset
+        shuffle = shuffle and (not is_iterable_dataset)
+        if self.test_max_n_samples is not None:
+            dataset = torch.utils.data.Subset(self.datasets["test"], list(range(self.test_max_n_samples)))
+        else:
+            dataset = self.datasets["test"]
+        return DataLoader(dataset, batch_size=self.batch_size,
+                          num_workers=self.num_workers, worker_init_fn=init_fn, shuffle=shuffle,
+                          collate_fn=self.collate_fn,
+                          )
+
+    def _predict_dataloader(self, shuffle=False):
+        if isinstance(self.datasets['predict'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
+            init_fn = worker_init_fn
+        else:
+            init_fn = None
+        return DataLoader(self.datasets["predict"], batch_size=self.batch_size,
+                          num_workers=self.num_workers, worker_init_fn=init_fn,
+                          collate_fn=self.collate_fn,
+                          )

main/utils_train.py

@@ -0,0 +1,173 @@
+import os, re
+from omegaconf import OmegaConf
+import logging
+mainlogger = logging.getLogger('mainlogger')
+
+import torch
+from collections import OrderedDict
+
+def init_workspace(name, logdir, model_config, lightning_config, rank=0):
+    workdir = os.path.join(logdir, name)
+    ckptdir = os.path.join(workdir, "checkpoints")
+    cfgdir = os.path.join(workdir, "configs")
+    loginfo = os.path.join(workdir, "loginfo")
+
+    # Create logdirs and save configs (all ranks will do to avoid missing directory error if rank:0 is slower)
+    os.makedirs(workdir, exist_ok=True)
+    os.makedirs(ckptdir, exist_ok=True)
+    os.makedirs(cfgdir, exist_ok=True)
+    os.makedirs(loginfo, exist_ok=True)
+
+    if rank == 0:
+        if "callbacks" in lightning_config and 'metrics_over_trainsteps_checkpoint' in lightning_config.callbacks:
+            os.makedirs(os.path.join(ckptdir, 'trainstep_checkpoints'), exist_ok=True)
+        OmegaConf.save(model_config, os.path.join(cfgdir, "model.yaml"))
+        OmegaConf.save(OmegaConf.create({"lightning": lightning_config}), os.path.join(cfgdir, "lightning.yaml"))
+    return workdir, ckptdir, cfgdir, loginfo
+
+def check_config_attribute(config, name):
+    if name in config:
+        value = getattr(config, name)
+        return value
+    else:
+        return None
+
+def get_trainer_callbacks(lightning_config, config, logdir, ckptdir, logger):
+    default_callbacks_cfg = {
+        "model_checkpoint": {
+            "target": "pytorch_lightning.callbacks.ModelCheckpoint",
+            "params": {
+                "dirpath": ckptdir,
+                "filename": "{epoch}",
+                "verbose": True,
+                "save_last": False,
+            }
+        },
+        "batch_logger": {
+            "target": "callbacks.ImageLogger",
+            "params": {
+                "save_dir": logdir,
+                "batch_frequency": 1000,
+                "max_images": 4,
+                "clamp": True,
+            }
+        },    
+        "learning_rate_logger": {
+            "target": "pytorch_lightning.callbacks.LearningRateMonitor",
+            "params": {
+                "logging_interval": "step",
+                "log_momentum": False
+            }
+        },
+        "cuda_callback": {
+            "target": "callbacks.CUDACallback"
+        },
+    }
+
+    ## optional setting for saving checkpoints
+    monitor_metric = check_config_attribute(config.model.params, "monitor")
+    if monitor_metric is not None:
+        mainlogger.info(f"Monitoring {monitor_metric} as checkpoint metric.")
+        default_callbacks_cfg["model_checkpoint"]["params"]["monitor"] = monitor_metric
+        default_callbacks_cfg["model_checkpoint"]["params"]["save_top_k"] = 3
+        default_callbacks_cfg["model_checkpoint"]["params"]["mode"] = "min"
+
+    if 'metrics_over_trainsteps_checkpoint' in lightning_config.callbacks:
+        mainlogger.info('Caution: Saving checkpoints every n train steps without deleting. This might require some free space.')
+        default_metrics_over_trainsteps_ckpt_dict = {
+            'metrics_over_trainsteps_checkpoint': {"target": 'pytorch_lightning.callbacks.ModelCheckpoint',
+                                                   'params': {
+                                                        "dirpath": os.path.join(ckptdir, 'trainstep_checkpoints'),
+                                                        "filename": "{epoch}-{step}",
+                                                        "verbose": True,
+                                                        'save_top_k': -1,
+                                                        'every_n_train_steps': 10000,
+                                                        'save_weights_only': True
+                                                    }
+                                                }
+        }
+        default_callbacks_cfg.update(default_metrics_over_trainsteps_ckpt_dict)
+
+    if "callbacks" in lightning_config:
+        callbacks_cfg = lightning_config.callbacks
+    else:
+        callbacks_cfg = OmegaConf.create()
+    callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg)
+
+    return callbacks_cfg
+
+def get_trainer_logger(lightning_config, logdir, on_debug):
+    default_logger_cfgs = {
+        "tensorboard": {
+            "target": "pytorch_lightning.loggers.TensorBoardLogger",
+            "params": {
+                "save_dir": logdir,
+                "name": "tensorboard",
+            }
+        },
+        "testtube": {
+            "target": "pytorch_lightning.loggers.CSVLogger",
+            "params": {
+                    "name": "testtube",
+                    "save_dir": logdir,
+                }
+            },
+    }
+    os.makedirs(os.path.join(logdir, "tensorboard"), exist_ok=True)
+    default_logger_cfg = default_logger_cfgs["tensorboard"]
+    if "logger" in lightning_config:
+        logger_cfg = lightning_config.logger
+    else:
+        logger_cfg = OmegaConf.create()
+    logger_cfg = OmegaConf.merge(default_logger_cfg, logger_cfg)
+    return logger_cfg
+
+def get_trainer_strategy(lightning_config):
+    default_strategy_dict = {
+        "target": "pytorch_lightning.strategies.DDPShardedStrategy"
+    }
+    if "strategy" in lightning_config:
+        strategy_cfg = lightning_config.strategy
+        return strategy_cfg
+    else:
+        strategy_cfg = OmegaConf.create()
+
+    strategy_cfg = OmegaConf.merge(default_strategy_dict, strategy_cfg)
+    return strategy_cfg
+
+def load_checkpoints(model, model_cfg):
+    if check_config_attribute(model_cfg, "pretrained_checkpoint"):
+        pretrained_ckpt = model_cfg.pretrained_checkpoint
+        assert os.path.exists(pretrained_ckpt), "Error: Pre-trained checkpoint NOT found at:%s"%pretrained_ckpt
+        mainlogger.info(">>> Load weights from pretrained checkpoint")
+
+        pl_sd = torch.load(pretrained_ckpt, map_location="cpu")
+        try:
+            if 'state_dict' in pl_sd.keys():
+                model.load_state_dict(pl_sd["state_dict"], strict=True)
+                mainlogger.info(">>> Loaded weights from pretrained checkpoint: %s"%pretrained_ckpt)
+            else:
+                # deepspeed
+                new_pl_sd = OrderedDict()
+                for key in pl_sd['module'].keys():
+                    new_pl_sd[key[16:]]=pl_sd['module'][key]
+                model.load_state_dict(new_pl_sd, strict=True)
+        except:
+            model.load_state_dict(pl_sd)
+    else:
+        mainlogger.info(">>> Start training from scratch")
+
+    return model
+
+def set_logger(logfile, name='mainlogger'):
+    logger = logging.getLogger(name)
+    logger.setLevel(logging.INFO)
+    fh = logging.FileHandler(logfile, mode='w')
+    fh.setLevel(logging.INFO)
+    ch = logging.StreamHandler()
+    ch.setLevel(logging.DEBUG)
+    fh.setFormatter(logging.Formatter("%(asctime)s-%(levelname)s: %(message)s"))
+    ch.setFormatter(logging.Formatter("%(message)s"))
+    logger.addHandler(fh)
+    logger.addHandler(ch)
+    return logger

prompts/512_interp/prompts.txt

@@ -0,0 +1,3 @@
+walking man
+an anime scene
+an anime scene