gradio_app.py

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
from cldm.model import load_state_dict
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_sketch'
    filename_list = ['tc_sketch.pt']
    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/tc_sketch.pt'
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!"
# ckpt_path = "/group/40005/gzhiwang/tc_sketch.pt"
ckpt_path = "/group/40034/gzhiwang/tc_sketch.pt"
model = load_model_checkpoint(model, ckpt_path)
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():
        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]
]




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_320x512'):
            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 End SKetch",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...")