copied from dreamcreature main repo

.gitignore

@@ -0,0 +1,8 @@
+.idea/
+src/data/cub200_2011/train
+src/data/dogs/Images
+__pycache__
+*/.ipynb_checkpoints
+/.ipynb_checkpoints/requirements-checkpoint.txt
+
+*.bin

app.py

@@ -0,0 +1,215 @@
+import gc
+import os
+import re
+import shutil
+
+import gradio as gr
+import requests
+import torch
+
+from dreamcreature.pipeline import create_args, load_pipeline
+
+CUB_DESCRIPTION = """
+# DreamCreature (CUB-200-2011)
+To create your own creature, you can type:
+
+`"a photo of a <head:id> <wing:id> bird"` where `id` ranges from 1~200 (200 classes corresponding to CUB-200-2011)
+
+For instance `"a photo of a <head:17> <wing:18> bird"` using head of `cardinal (17)` and wing of `spotted catbird (18)`
+
+Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/cub200_2011/class_names.txt
+
+You can also try any prompt you like such as:
+
+Sub-concept transfer: `"a photo of a <wing:17> cat"`
+
+Inspiring design: `"a photo of a <head:101> <wing:191> teddy bear"`
+
+(Experimental) You can also use two parts together such as:
+
+`"a photo of a <head:17> <head:18> bird"` mixing head of `cardinal (17)` and `spotted catbird (18)`
+
+The current available parts are: `head`, `body`, `wing`, `tail`, and `leg`
+
+"""
+
+DOG_DESCRIPTION = """
+# DreamCreature (Stanford Dogs)
+To create your own creature, you can type:
+
+`"a photo of a <nose:id> <ear:id> dog"` where `id` ranges from 0~119 (120 classes corresponding to Stanford Dogs)
+
+For instance `"a photo of a <nose:2> <ear:112> dog"` using head of `maltese dog (2)` and wing of `cardigan (112)`
+
+Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/dogs/class_names.txt
+
+Sub-concept transfer: `"a photo of a <ear:112> cat"`
+
+Inspiring design: `"a photo of a <eye:38> <body:38> teddy bear"`
+
+(Experimental) You can also use two parts together such as:
+
+`"a photo of a <nose:1> <nose:112> dog"` mixing head of `maltese dog (2)` and `spotted cardigan (112)`
+
+The current available parts are: `eye`, `neck`, `ear`, `body`, `leg`, `nose` and `forehead`
+
+"""
+
+
+def prepare_pipeline(model_name):
+    is_cub = 'cub' in model_name
+
+    checkpoint_name = {
+        'dreamcreature-sd1.5-cub200': 'checkpoint-74900',
+        'dreamcreature-sd1.5-dog': 'checkpoint-150000'
+    }
+
+    repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
+    file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
+    local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
+    os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
+    download_file(file_url, local_path)
+
+    file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
+    local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
+    download_file(file_url, local_path)
+
+    OUTPUT_DIR = model_name
+
+    args = create_args(OUTPUT_DIR)
+    if 'dpo' in OUTPUT_DIR:
+        args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
+
+    pipe = load_pipeline(args, torch.float16, 'cuda')
+    pipe = pipe.to(torch.float16)
+
+    pipe.verbose = True
+    pipe.v = 're'
+
+    if is_cub:
+        pipe.num_k_per_part = 200
+
+        MAPPING = {
+            'body': 0,
+            'tail': 1,
+            'head': 2,
+            'wing': 4,
+            'leg': 6
+        }
+
+        ID2NAME = open('data/cub200_2011/class_names.txt').readlines()
+        ID2NAME = [line.strip() for line in ID2NAME]
+
+    else:
+        pipe.num_k_per_part = 120
+
+        MAPPING = {
+            'eye': 0,
+            'neck': 2,
+            'ear': 3,
+            'body': 4,
+            'leg': 5,
+            'nose': 6,
+            'forehead': 7
+        }
+
+        ID2NAME = open('data/dogs/class_names.txt').readlines()
+        ID2NAME = [line.strip() for line in ID2NAME]
+
+    return pipe, MAPPING, ID2NAME
+
+
+def download_file(url, local_path):
+    if os.path.exists(local_path):
+        return
+
+    with requests.get(url, stream=True) as r:
+        with open(local_path, 'wb') as f:
+            shutil.copyfileobj(r.raw, f)
+
+
+def process_text(text, MAPPING, ID2NAME):
+    pattern = r"<([^:>]+):(\d+)>"
+    result = text
+    offset = 0
+
+    part2id = []
+
+    for match in re.finditer(pattern, text):
+        key = match.group(1)
+        clsid = int(match.group(2))
+        clsid = min(max(clsid, 1), 200)  # must be 1~200
+
+        replacement = f"<{MAPPING[key]}:{clsid - 1}>"
+        start, end = match.span()
+
+        # Adjust the start and end positions based on the offset from previous replacements
+        start += offset
+        end += offset
+
+        # Replace the matched text with the replacement
+        result = result[:start] + replacement + result[end:]
+
+        # Update the offset for the next replacement
+        offset += len(replacement) - (end - start)
+
+        part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
+
+    return result, part2id
+
+
+def generate_images(model_name, prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+
+    try:
+        pipe, MAPPING, ID2NAME = prepare_pipeline(model_name)
+
+        prompt, part2id = process_text(prompt, MAPPING, ID2NAME)
+        negative_prompt, _ = process_text(negative_prompt, MAPPING, ID2NAME)
+
+        images = pipe(prompt,
+                      negative_prompt=negative_prompt, generator=generator,
+                      num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                      num_images_per_prompt=num_images).images
+
+        del pipe
+    except Exception as e:
+        raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
+                       f"The error message: {e}")
+    finally:
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    return images, '; '.join(part2id)
+
+
+with gr.Blocks(title="DreamCreature") as demo:
+    with gr.Row():
+        main_desc = gr.Markdown(CUB_DESCRIPTION)
+    with gr.Column():
+        with gr.Row():
+            with gr.Group():
+                dropdown = gr.Dropdown(choices=["dreamcreature-sd1.5-cub200",
+                                                "dreamcreature-sd1.5-dog"],
+                                       value="dreamcreature-sd1.5-cub200")
+                prompt = gr.Textbox(label="Prompt", value="a photo of a <head:101> <wing:191> teddy bear")
+                negative_prompt = gr.Textbox(label="Negative Prompt",
+                                             value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+                num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
+                guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+                num_images = gr.Slider(minimum=1, maximum=4, step=1, value=4, label="Number of Images")
+                seed = gr.Number(label="Seed", value=777881414)
+                button = gr.Button()
+
+            with gr.Column():
+                output_images = gr.Gallery(columns=4, label='Output')
+                markdown_labels = gr.Markdown("")
+
+    dropdown.change(fn=lambda x: {'dreamcreature-sd1.5-cub200': CUB_DESCRIPTION,
+                                  'dreamcreature-sd1.5-dog': DOG_DESCRIPTION}[x], inputs=dropdown, outputs=main_desc)
+    button.click(fn=generate_images,
+                 inputs=[dropdown, prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
+                         seed], outputs=[output_images, markdown_labels], show_progress=True)
+
+demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')

data/cub200_2011/class_names.txt

@@ -0,0 +1,200 @@
+black footed albatross
+laysan albatross
+sooty albatross
+groove billed ani
+crested auklet
+least auklet
+parakeet auklet
+rhinoceros auklet
+brewer blackbird
+red winged blackbird
+rusty blackbird
+yellow headed blackbird
+bobolink
+indigo bunting
+lazuli bunting
+painted bunting
+cardinal
+spotted catbird
+gray catbird
+yellow breasted chat
+eastern towhee
+chuck will widow
+brandt cormorant
+red faced cormorant
+pelagic cormorant
+bronzed cowbird
+shiny cowbird
+brown creeper
+american crow
+fish crow
+black billed cuckoo
+mangrove cuckoo
+yellow billed cuckoo
+gray crowned rosy finch
+purple finch
+northern flicker
+acadian flycatcher
+great crested flycatcher
+least flycatcher
+olive sided flycatcher
+scissor tailed flycatcher
+vermilion flycatcher
+yellow bellied flycatcher
+frigatebird
+northern fulmar
+gadwall
+american goldfinch
+european goldfinch
+boat tailed grackle
+eared grebe
+horned grebe
+pied billed grebe
+western grebe
+blue grosbeak
+evening grosbeak
+pine grosbeak
+rose breasted grosbeak
+pigeon guillemot
+california gull
+glaucous winged gull
+heermann gull
+herring gull
+ivory gull
+ring billed gull
+slaty backed gull
+western gull
+anna hummingbird
+ruby throated hummingbird
+rufous hummingbird
+green violetear
+long tailed jaeger
+pomarine jaeger
+blue jay
+florida jay
+green jay
+dark eyed junco
+tropical kingbird
+gray kingbird
+belted kingfisher
+green kingfisher
+pied kingfisher
+ringed kingfisher
+white breasted kingfisher
+red legged kittiwake
+horned lark
+pacific loon
+mallard
+western meadowlark
+hooded merganser
+red breasted merganser
+mockingbird
+nighthawk
+clark nutcracker
+white breasted nuthatch
+baltimore oriole
+hooded oriole
+orchard oriole
+scott oriole
+ovenbird
+brown pelican
+white pelican
+western wood pewee
+sayornis
+american pipit
+whip poor will
+horned puffin
+common raven
+white necked raven
+american redstart
+geococcyx
+loggerhead shrike
+great grey shrike
+baird sparrow
+black throated sparrow
+brewer sparrow
+chipping sparrow
+clay colored sparrow
+house sparrow
+field sparrow
+fox sparrow
+grasshopper sparrow
+harris sparrow
+henslow sparrow
+le conte sparrow
+lincoln sparrow
+nelson sharp tailed sparrow
+savannah sparrow
+seaside sparrow
+song sparrow
+tree sparrow
+vesper sparrow
+white crowned sparrow
+white throated sparrow
+cape glossy starling
+bank swallow
+barn swallow
+cliff swallow
+tree swallow
+scarlet tanager
+summer tanager
+artic tern
+black tern
+caspian tern
+common tern
+elegant tern
+forsters tern
+least tern
+green tailed towhee
+brown thrasher
+sage thrasher
+black capped vireo
+blue headed vireo
+philadelphia vireo
+red eyed vireo
+warbling vireo
+white eyed vireo
+yellow throated vireo
+bay breasted warbler
+black and white warbler
+black throated blue warbler
+blue winged warbler
+canada warbler
+cape may warbler
+cerulean warbler
+chestnut sided warbler
+golden winged warbler
+hooded warbler
+kentucky warbler
+magnolia warbler
+mourning warbler
+myrtle warbler
+nashville warbler
+orange crowned warbler
+palm warbler
+pine warbler
+prairie warbler
+prothonotary warbler
+swainson warbler
+tennessee warbler
+wilson warbler
+worm eating warbler
+yellow warbler
+northern waterthrush
+louisiana waterthrush
+bohemian waxwing
+cedar waxwing
+american three toed woodpecker
+pileated woodpecker
+red bellied woodpecker
+red cockaded woodpecker
+red headed woodpecker
+downy woodpecker
+bewick wren
+cactus wren
+carolina wren
+house wren
+marsh wren
+rock wren
+winter wren
+common yellowthroat

data/cub200_2011/pretrained_kmeans.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00b2ff84f80daa3cdbd4b18e4088fd900b70a8a192a70957c34e4369d6065e65
+size 6874495

data/dogs/class_names.txt

@@ -0,0 +1,120 @@
+Chihuahua
+Japanese spaniel
+Maltese dog
+Pekinese
+Shih Tzu
+Blenheim spaniel
+papillon
+toy terrier
+Rhodesian ridgeback
+Afghan hound
+basset
+beagle
+bloodhound
+bluetick
+black and tan coonhound
+Walker hound
+English foxhound
+redbone
+borzoi
+Irish wolfhound
+Italian greyhound
+whippet
+Ibizan hound
+Norwegian elkhound
+otterhound
+Saluki
+Scottish deerhound
+Weimaraner
+Staffordshire bullterrier
+American Staffordshire terrier
+Bedlington terrier
+Border terrier
+Kerry blue terrier
+Irish terrier
+Norfolk terrier
+Norwich terrier
+Yorkshire terrier
+wire haired fox terrier
+Lakeland terrier
+Sealyham terrier
+Airedale
+cairn
+Australian terrier
+Dandie Dinmont
+Boston bull
+miniature schnauzer
+giant schnauzer
+standard schnauzer
+Scotch terrier
+Tibetan terrier
+silky terrier
+soft coated wheaten terrier
+West Highland white terrier
+Lhasa
+flat coated retriever
+curly coated retriever
+golden retriever
+Labrador retriever
+Chesapeake Bay retriever
+German short haired pointer
+vizsla
+English setter
+Irish setter
+Gordon setter
+Brittany spaniel
+clumber
+English springer
+Welsh springer spaniel
+cocker spaniel
+Sussex spaniel
+Irish water spaniel
+kuvasz
+schipperke
+groenendael
+malinois
+briard
+kelpie
+komondor
+Old English sheepdog
+Shetland sheepdog
+collie
+Border collie
+Bouvier des Flandres
+Rottweiler
+German shepherd
+Doberman
+miniature pinscher
+Greater Swiss Mountain dog
+Bernese mountain dog
+Appenzeller
+EntleBucher
+boxer
+bull mastiff
+Tibetan mastiff
+French bulldog
+Great Dane
+Saint Bernard
+Eskimo dog
+malamute
+Siberian husky
+affenpinscher
+basenji
+pug
+Leonberg
+Newfoundland
+Great Pyrenees
+Samoyed
+Pomeranian
+chow
+keeshond
+Brabancon griffon
+Pembroke
+Cardigan
+toy poodle
+miniature poodle
+standard poodle
+Mexican hairless
+dingo
+dhole
+African hunting dog

data/dogs/pretrained_kmeans.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bf723669b2d6dad50d58a6d7b3dad7fafa6b49d8a3fca3fd5b713662ccd4b88
+size 6874495

dreamcreature/attn_processor.py

@@ -0,0 +1,143 @@
+from diffusers.models.attention_processor import *
+
+
+class LoRAAttnProcessorCustom(nn.Module, AttnProcessor):
+    r"""
+    Processor for implementing the LoRA attention mechanism.
+
+    Args:
+        hidden_size (`int`, *optional*):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the `encoder_hidden_states`.
+        rank (`int`, defaults to 4):
+            The dimension of the LoRA update matrices.
+        network_alpha (`int`, *optional*):
+            Equivalent to `alpha` but it's usage is specific to Kohya (A1111) style LoRAs.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, rank=4, network_alpha=None, **kwargs):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.rank = rank
+
+        q_rank = kwargs.pop("q_rank", None)
+        q_hidden_size = kwargs.pop("q_hidden_size", None)
+        q_rank = q_rank if q_rank is not None else rank
+        q_hidden_size = q_hidden_size if q_hidden_size is not None else hidden_size
+
+        v_rank = kwargs.pop("v_rank", None)
+        v_hidden_size = kwargs.pop("v_hidden_size", None)
+        v_rank = v_rank if v_rank is not None else rank
+        v_hidden_size = v_hidden_size if v_hidden_size is not None else hidden_size
+
+        out_rank = kwargs.pop("out_rank", None)
+        out_hidden_size = kwargs.pop("out_hidden_size", None)
+        out_rank = out_rank if out_rank is not None else rank
+        out_hidden_size = out_hidden_size if out_hidden_size is not None else hidden_size
+
+        self.to_q_lora = LoRALinearLayer(q_hidden_size, q_hidden_size, q_rank, network_alpha)
+        self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+        self.to_v_lora = LoRALinearLayer(cross_attention_dim or v_hidden_size, v_hidden_size, v_rank, network_alpha)
+        self.to_out_lora = LoRALinearLayer(out_hidden_size, out_hidden_size, out_rank, network_alpha)
+
+    def __call__(self, attn: Attention, hidden_states, *args, **kwargs):
+        self_cls_name = self.__class__.__name__
+        deprecate(
+            self_cls_name,
+            "0.26.0",
+            (
+                f"Make sure use {self_cls_name[4:]} instead by setting"
+                "LoRA layers to `self.{to_q,to_k,to_v,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
+                " `LoraLoaderMixin.load_lora_weights`"
+            ),
+        )
+        attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
+        attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
+        attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
+        attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
+
+        attn._modules.pop("processor")
+        attn.processor = AttnProcessorCustom(16)
+        return attn.processor(attn, hidden_states, *args, **kwargs)
+
+
+class AttnProcessorCustom(AttnProcessor):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, attn_size):
+        self.attn_size = attn_size
+
+    def __call__(
+            self,
+            attn: Attention,
+            hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=None,
+            temb=None,
+            scale=1.0,
+    ):
+        residual = hidden_states
+
+        args = () if USE_PEFT_BACKEND else (scale,)
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states, *args)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states, *args)
+        value = attn.to_v(encoder_hidden_states, *args)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attn_size = self.attn_size
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        if attention_probs.size(2) == 77 and attention_probs.size(1) == (attn_size * attn_size):  # (B*Head,HW,L)
+            attn_probs_cache = attention_probs.reshape(batch_size, -1, attn_size, attn_size, 77)
+            attn.attn_probs = attn_probs_cache
+        else:
+            attn.attn_probs = None
+
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states, *args)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

dreamcreature/dataset.py

@@ -0,0 +1,155 @@
+import os
+import random
+
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import Dataset
+
+
+class ImageDataset(Dataset):
+
+    def __init__(self,
+                 rootdir,
+                 filename='train.txt',
+                 path_prefix='',
+                 transform=None,
+                 target_transform=None):
+        super().__init__()
+
+        self.rootdir = rootdir
+        self.filename = filename
+        self.path_prefix = path_prefix
+
+        self.image_paths = []
+        self.image_labels = []
+
+        filename = os.path.join(self.rootdir, self.filename)
+
+        with open(filename, 'r') as f:
+            while True:
+                lines = f.readline()
+                if not lines:
+                    break
+
+                lines = lines.strip()
+                split_lines = lines.split(' ')
+                path_tmp = split_lines[0]
+                label_tmp = split_lines[1:]
+                self.is_onehot = len(label_tmp) != 1
+                if not self.is_onehot:
+                    label_tmp = label_tmp[0]
+                self.image_paths.append(path_tmp)
+                self.image_labels.append(label_tmp)
+
+        self.image_paths = np.array(self.image_paths)
+        self.image_labels = np.array(self.image_labels, dtype=np.float32)
+
+        self.transform = transform
+        self.target_transform = target_transform
+
+    def __getitem__(self, index):
+        """
+        Args:
+            index (int): Index
+        Returns:
+            tuple: (image, target) where target is index of the target class.
+        """
+        path, target = self.image_paths[index], self.image_labels[index]
+        target = torch.tensor(target)
+
+        img = Image.open(f'{self.path_prefix}{path}').convert('RGB')
+
+        if self.transform is not None:
+            img = self.transform(img)
+
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        return img, target, index
+
+    def __len__(self):
+        return len(self.image_paths)
+
+
+class DreamCreatureDataset(ImageDataset):
+
+    def __init__(self,
+                 rootdir,
+                 filename='train.txt',
+                 path_prefix='',
+                 code_filename='train_caps.txt',
+                 num_parts=8, num_k_per_part=256, repeat=1,
+                 use_gt_label=False,
+                 bg_code=7,
+                 transform=None,
+                 target_transform=None):
+        super().__init__(rootdir, filename, path_prefix, transform, target_transform)
+
+        self.image_codes = np.array(open(rootdir + '/' + code_filename).readlines())
+        self.num_parts = num_parts
+        self.num_k_per_part = num_k_per_part
+        self.repeat = repeat
+        self.use_gt_label = use_gt_label
+        self.bg_code = bg_code
+
+    def filter_by_class(self, target):
+        target_mask = self.image_labels == target
+        self.image_paths = self.image_paths[target_mask]
+        self.image_codes = self.image_codes[target_mask]
+        self.image_labels = self.image_labels[target_mask]
+
+    def set_max_samples(self, n, seed):
+        np.random.seed(seed)
+        rand_idx = np.arange(len(self.image_paths))
+        np.random.shuffle(rand_idx)
+
+        self.image_paths = self.image_paths[rand_idx[:n]]
+        self.image_codes = self.image_codes[rand_idx[:n]]
+        self.image_labels = self.image_labels[rand_idx[:n]]
+
+    def __len__(self):
+        return len(self.image_paths) * self.repeat
+
+    def __getitem__(self, index):
+        """
+        Args:
+            index (int): Index
+        Returns:
+            tuple: (image, target) where target is index of the target class.
+        """
+        index = index % len(self.image_paths)
+        path, target = self.image_paths[index], self.image_labels[index]
+        target = torch.tensor(target)
+
+        img = Image.open(f'{self.path_prefix}{path}').convert('RGB')
+
+        cap = self.image_codes[index].strip()
+
+        if self.transform is not None:
+            img = self.transform(img)
+
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        appeared = []
+
+        code = torch.ones(self.num_parts) * self.num_k_per_part  # represents not exists
+        splits = cap.strip().replace('.', '').split(' ')
+        for c in splits:
+            idx, intval = c.split(':')
+            appeared.append(int(idx))
+            if self.use_gt_label and self.bg_code != int(idx):
+                code[int(idx)] = target
+            else:
+                code[int(idx)] = int(intval)
+
+        example = {
+            'pixel_values': img,
+            'captions': cap,
+            'codes': code,
+            'labels': target,
+            'appeared': appeared
+        }
+
+        return example

dreamcreature/dino.py

@@ -0,0 +1,42 @@
+import torch.nn as nn
+from transformers import AutoModel, AutoImageProcessor
+
+
+class DINO(nn.Module):
+    NAMES = {
+        'dino': 'facebook/dino-vits16',
+        'dinov2': 'facebook/dinov2-base'
+    }
+
+    def __init__(self, name='facebook/dinov2-base', **kwargs):
+        super().__init__()
+
+        self.model = AutoModel.from_pretrained(name)
+        self.processor = AutoImageProcessor.from_pretrained(name)
+
+    def forward(self, image):
+        vit_output = self.model(image,
+                                output_hidden_states=True,
+                                return_dict=True)
+
+        outputs = {}
+        for i in range(1, len(vit_output.hidden_states)):
+            outputs[f'block{i}'] = vit_output.hidden_states[i][:, 0]  # get cls only
+        outputs['feats'] = outputs[f'block{i}']
+        return outputs
+
+    def preprocess(self, image, size=None):
+        inputs = self.processor(images=image, return_tensors="pt", size=size)
+        return inputs['pixel_values']
+
+    def get_feat_maps(self, image, index=-1):
+        vit_output = self.model(image,
+                                output_hidden_states=True,
+                                return_dict=True)
+
+        last_hidden_states = vit_output.hidden_states[index]
+
+        B, T, C = last_hidden_states.size()
+        HW = int((T - 1) ** 0.5)
+
+        return last_hidden_states[:, 1:, :].reshape(B, HW, HW, C).permute(0, 3, 1, 2)  # (B, C, H, W)

dreamcreature/kmeans_segmentation.py

@@ -0,0 +1,188 @@
+import os
+from typing import List
+
+import numpy as np
+import torch
+import torchpq
+
+
+class KMeansSegmentation:
+    FOREGROUND = 'foreground_background'
+    COARSE = 'coarse_kmeans'
+    FINE = 'fine_kmeans'
+
+    def __init__(self, path, foreground_idx=0, background_code=7, M=8, K=256):
+        if not os.path.exists(path):
+            raise FileNotFoundError(f'please train {path}')
+        kmeans = torch.load(path)
+
+        self.foreground_idx = foreground_idx
+        self.kmeans = kmeans
+        self.background_code = background_code
+        self.M = M
+        self.K = K
+
+        self.fg: torchpq.clustering.KMeans = kmeans[KMeansSegmentation.FOREGROUND]
+        self.coarse: torchpq.clustering.KMeans = kmeans[KMeansSegmentation.COARSE]
+        self.fine: List[torchpq.clustering.KMeans] = kmeans[KMeansSegmentation.FINE]
+
+    def obtain_fine_feats(self, prompt, filter_idxs=[]):
+        if isinstance(prompt, str):
+            code = np.zeros((self.M,), dtype=int)
+            splits = prompt.strip().split(' ')
+            for s in splits:
+                m, k = s.split(':')
+                code[int(m)] = int(k)
+        else:
+            code = prompt
+
+        fine_feats = []
+        for m in range(self.M):
+            fine_feats.append(self.fine[m].centroids.cpu().t()[code[m]])
+        fine_feats = torch.stack(fine_feats, dim=0)
+
+        if len(filter_idxs) != 0:
+            new_fine_feats = []
+
+            for m in range(self.M):
+                if m not in filter_idxs:
+                    new_fine_feats.append(fine_feats[m])
+
+            fine_feats = torch.stack(new_fine_feats, dim=0)
+
+        return fine_feats
+
+    def get_segmask(self, feat_map, with_appeared_tokens=False):
+        N, C, H, W = feat_map.size()
+        query = feat_map.cuda().reshape(N, C, H * W).permute(0, 2, 1)  # (N, H*W, C)
+
+        fg_labels = self.fg.predict(query.reshape(N * H * W, C).t().contiguous())  # (N*H*W)
+        fg_labels = fg_labels.reshape(N, H * W)
+
+        fg_idx = self.foreground_idx
+        bg_idx = 1 - self.foreground_idx
+
+        nobg = []
+        bgmean = []
+
+        for i in range(N):
+            bgnorm_mean = query[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)
+
+            if fg_idx == 0:
+                bg_mask = fg_labels[i]
+            else:
+                bg_mask = 1 - fg_labels[i]
+
+            bg_mask = bg_mask.unsqueeze(1)
+            nobg.append(query[i] * (1 - bg_mask) + (-1 * bg_mask))
+            bgmean.append(bgnorm_mean)
+
+        nobg = torch.stack(nobg, dim=0)  # (B, H*W, C)
+        coarse_labels = self.coarse.predict(nobg.reshape(N * H * W, 768).t().contiguous())
+        coarse_labels = coarse_labels.reshape(N, H, W)
+
+        segmasks = []
+        for m in range(self.M):
+            mask = (coarse_labels == m).float()  # (N, H, W)
+            segmasks.append(mask)
+        segmasks = torch.stack(segmasks, dim=1)  # (N, M, H, W)
+
+        if with_appeared_tokens:
+            appeared_tokens = []
+            for i in range(N):
+                appeared_tokens.append(torch.unique(coarse_labels[i].reshape(-1)).tolist())
+            return segmasks, appeared_tokens
+
+        return segmasks
+
+    def predict(self, feat_map, disable=True, filter_idxs=[]):
+        # feat_map: (B, C, H, W)
+
+        N, C, H, W = feat_map.size()
+        query = feat_map.reshape(N, C, H * W).permute(0, 2, 1)  # (N, H*W, C)
+
+        fg_labels = self.fg.predict(query.reshape(N * H * W, C).t().contiguous().cuda()).cpu()  # (N*H*W)
+        fg_labels = fg_labels.reshape(N, H * W)
+
+        fg_idx = self.foreground_idx
+        bg_idx = 1 - self.foreground_idx
+
+        nobg = []
+        bgmean = []
+
+        for i in range(N):
+            bgnorm_mean = query[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)
+
+            if fg_idx == 0:
+                bg_mask = fg_labels[i]
+            else:
+                bg_mask = 1 - fg_labels[i]
+
+            bg_mask = bg_mask.unsqueeze(1)
+            nobg.append(query[i] * (1 - bg_mask) + (-1 * bg_mask))
+            bgmean.append(bgnorm_mean)
+
+        nobg = torch.stack(nobg, dim=0)  # (B, H*W, C)
+        bgmean = torch.cat(bgmean, dim=0)
+
+        coarse_labels = self.coarse.predict(nobg.reshape(N * H * W, 768).t().contiguous().cuda()).cpu()
+        coarse_labels = coarse_labels.reshape(N, H * W)
+
+        from tqdm.auto import tqdm
+
+        fgmean = []
+        M = self.M
+
+        locs = np.zeros((N, M, 2))
+
+        for i in tqdm(range(N), disable=disable):
+            mean_feats = []
+            for m in range(M):
+                coarse_mask = coarse_labels[i] == m
+                if coarse_mask.sum().item() == 0:
+                    m_mean_feats = torch.zeros(1, C)
+                else:
+                    locs[i, m] = (coarse_mask.reshape(H, W).nonzero().float().add(0.5).mean(dim=0) / H).cpu().numpy()
+                    m_mean_feats = query[i][coarse_mask].mean(dim=0, keepdim=True)  # (H*W,C) -> (1,C)
+
+                mean_feats.append(m_mean_feats)
+
+            mean_feats = torch.cat(mean_feats, dim=0)
+            fgmean.append(mean_feats)
+
+        fgmean = torch.stack(fgmean, dim=0)  # (N, M, C)
+        final_labels = torch.ones(N, M) * self.K
+
+        for m in range(M):
+            fine_kmeans = self.fine[m]
+
+            if m == self.background_code:
+                fine_labels = fine_kmeans.predict(bgmean.t().contiguous().cuda()).cpu()
+                final_labels[:, m] = fine_labels
+            else:
+                fine_inp = fgmean[:, m].reshape(N, C)
+                is_zero = fine_inp.sum(dim=1) == 0
+                fine_labels = fine_kmeans.predict(fine_inp.t().contiguous().cuda()).cpu()
+                fine_labels[is_zero] = self.K
+
+                final_labels[:, m] = fine_labels
+
+        fgmean[:, self.background_code] = bgmean
+        fine_prompts = []
+
+        for fine_label in final_labels:
+            prompt_dict = {k: int(v) for k, v in enumerate(list(fine_label))}
+            if len(filter_idxs) != 0:
+                for i in filter_idxs:
+                    del prompt_dict[i]
+            prompt = ' '.join([f'{k}:{v}' for k, v in prompt_dict.items() if v != self.K])
+            fine_prompts.append(prompt)
+
+        return {
+            'features': fgmean,
+            'fg_labels': fg_labels,
+            'coarse_labels': coarse_labels,
+            'fine_labels': final_labels,
+            'fine_prompts': fine_prompts,
+            'location': locs
+        }

dreamcreature/loss.py

@@ -0,0 +1,73 @@
+import torch
+import torch.nn.functional as F
+from diffusers import UNet2DConditionModel
+from diffusers.models.attention_processor import Attention
+
+from dreamcreature.dino import DINO
+from dreamcreature.kmeans_segmentation import KMeansSegmentation
+
+
+def dreamcreature_loss(batch,
+                       unet: UNet2DConditionModel,
+                       dino: DINO,
+                       seg: KMeansSegmentation,
+                       placeholder_token_ids,
+                       accelerator):
+    attn_probs = {}
+
+    for name, module in unet.named_modules():
+        if isinstance(module, Attention) and module.attn_probs is not None:
+            a = module.attn_probs.mean(dim=1)  # (B,Head,H,W,77) -> (B,H,W,77)
+            attn_probs[name] = a
+
+    avg_attn_map = []
+    for name in attn_probs:
+        avg_attn_map.append(attn_probs[name])
+
+    avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0)  # (L,B,H,W,77) -> (B,H,W,77)
+    B, H, W, seq_length = avg_attn_map.size()
+    located_attn_map = []
+
+    # locate the attn map
+    for i, placeholder_token_id in enumerate(placeholder_token_ids):
+        for bi in range(B):
+            if "input_ids" in batch:
+                learnable_idx = (batch["input_ids"][bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
+            else:
+                learnable_idx = (batch["input_ids_one"][bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
+
+            if len(learnable_idx) != 0:  # only assign if found
+                if len(learnable_idx) == 1:
+                    offset_learnable_idx = learnable_idx
+                else:  # if there is two and above.
+                    raise NotImplementedError
+
+                located_map = avg_attn_map[bi, :, :, offset_learnable_idx]
+                located_attn_map.append(located_map)
+            else:
+                located_attn_map.append(torch.zeros(H, W, 1).to(accelerator.device))
+
+    M = len(placeholder_token_ids)
+    located_attn_map = torch.stack(located_attn_map, dim=0).reshape(M, B, H, W).transpose(0, 1)  # (B, M, 16, 16)
+
+    raw_images = batch['raw_images']
+    dino_input = dino.preprocess(raw_images, size=224)
+    with torch.no_grad():
+        dino_ft = dino.get_feat_maps(dino_input)
+        segmasks, appeared_tokens = seg.get_segmask(dino_ft, True)  # (B, M, H, W)
+        segmasks = segmasks.to(located_attn_map.dtype)
+        if H != 16:  # for res 1024
+            segmasks = F.interpolate(segmasks, (H, W), mode='nearest')
+
+        masks = []
+        for i, appeared in enumerate(appeared_tokens):
+            mask = (segmasks[i, appeared].sum(dim=0) > 0).float()  # (A, H, W) -> (H, W)
+            masks.append(mask)
+        masks = torch.stack(masks, dim=0)  # (B, H, W)
+        batch['masks'] = masks
+
+    norm_map = located_attn_map / located_attn_map.sum(dim=1, keepdim=True).clamp(min=1e-6)
+    # if norm_map is assigned manually, means the sub-concept token is not found, hence no gradient will be backprop
+    attn_loss = F.binary_cross_entropy(norm_map.clamp(min=0, max=1),
+                                       segmasks.clamp(min=0, max=1))
+    return attn_loss, located_attn_map.detach().max()

dreamcreature/mapper.py

@@ -0,0 +1,75 @@
+from typing import Optional
+
+import torch
+from torch import nn
+
+
+class TokenMapper(nn.Module):
+    def __init__(self,
+                 num_parts,
+                 num_k_per_part,
+                 out_dims,
+                 projection_nlayers=1,
+                 projection_activation=nn.ReLU(),
+                 with_pe=True):
+        super().__init__()
+
+        self.num_parts = num_parts
+        self.num_k_per_part = num_k_per_part
+        self.with_pe = with_pe
+        self.out_dims = out_dims
+
+        self.embedding = nn.Embedding((self.num_k_per_part + 1) * num_parts, out_dims)
+        if with_pe:
+            self.pe = nn.Parameter(torch.randn(num_parts, out_dims))
+        else:
+            self.register_buffer('pe', torch.zeros(num_parts, out_dims))
+
+        if projection_nlayers == 0:
+            self.projection = nn.Identity()
+        else:
+            projections = []
+            for i in range(projection_nlayers - 1):
+                projections.append(nn.Linear(out_dims, out_dims))
+                projections.append(projection_activation)
+
+            projections.append(nn.Linear(out_dims, out_dims))
+            self.projection = nn.Sequential(*projections)
+
+    def get_all_embeddings(self, no_projection=False, no_pe=False):
+        idx = torch.arange(self.num_parts * (self.num_k_per_part + 1)).long().to(self.embedding.weight.device)
+        idx = idx.reshape(self.num_parts, self.num_k_per_part + 1)
+        emb = self.embedding(idx)  # (K, N, d)
+
+        if not no_pe:
+            emb_pe = emb + self.pe.unsqueeze(1)
+        else:
+            emb_pe = emb
+
+        if not no_projection:
+            projected = self.projection(emb_pe)
+        else:
+            projected = emb_pe
+
+        return projected
+
+    def forward(self, hashes, index: Optional[torch.Tensor] = None):
+        B = hashes.size(0)
+
+        # 0, 257, 514, ...
+        if index is None:
+            offset = torch.arange(self.num_parts, device=hashes.device) * (self.num_k_per_part + 1)
+            hashes = self.embedding(hashes.long() + offset.reshape(1, -1))  # (B, N, d)
+        else:
+            offset = index.reshape(-1) * (self.num_k_per_part + 1)
+            hashes = self.embedding(hashes.long() + offset.reshape(B, -1).long())  # (B, N, d)
+
+        if index is not None:
+            pe = self.pe[index.reshape(-1)]  # index must be equal size
+            pe = pe.reshape(B, -1, self.out_dims)
+            hashes = hashes + pe
+        else:
+            hashes = hashes + self.pe.unsqueeze(0).repeat(B, 1, 1)
+        projected = self.projection(hashes)
+
+        return projected

dreamcreature/pipeline.py

@@ -0,0 +1,771 @@
+import os
+import re
+
+from diffusers.loaders import AttnProcsLayers
+from diffusers.models.attention_processor import Attention
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import *
+from omegaconf import OmegaConf
+
+from dreamcreature.attn_processor import LoRAAttnProcessorCustom
+from dreamcreature.mapper import TokenMapper
+from dreamcreature.text_encoder import CustomCLIPTextModel
+from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
+from utils import add_tokens, get_attn_processors
+
+
+def setup_attn_processor(unet, **kwargs):
+    lora_attn_procs = {}
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+
+        lora_attn_procs[name] = LoRAAttnProcessorCustom(
+            hidden_size=hidden_size,
+            cross_attention_dim=cross_attention_dim,
+            rank=kwargs['rank'],
+        )
+
+    unet.set_attn_processor(lora_attn_procs)
+
+
+def load_attn_processor(unet, filename):
+    lora_layers = AttnProcsLayers(get_attn_processors(unet))
+    lora_layers.load_state_dict(torch.load(filename))
+
+
+def convert_prompt_re(prompt: str):
+    pattern = r"<(\d+):(\d+)>"
+    result = prompt
+    offset = 0
+
+    ints = []
+    parts_i = []
+
+    for match in re.finditer(pattern, prompt):
+        i = int(match.group(1))
+        b = int(match.group(2))
+
+        replacement = f"<part>_{i}"
+        start, end = match.span()
+
+        # Adjust the start and end positions based on the offset from previous replacements
+        start += offset
+        end += offset
+
+        # Replace the matched text with the replacement
+        result = result[:start] + replacement + result[end:]
+
+        # Update the offset for the next replacement
+        offset += len(replacement) - (end - start)
+
+        parts_i.append(i)
+        ints.append(b)
+
+    result = result.strip()
+
+    if len(ints) == 0:
+        return result, None, None
+
+    ints = torch.tensor(ints)  # (nparts,)
+    return result, ints, parts_i
+
+
+def convert_prompt(prompt: str, replace_token: bool = False, v='v1'):
+    r"""
+    Parameters:
+        prompt (`str`):
+            The prompt to guide the image generation.
+
+    Returns:
+        `str`: The converted prompt
+    """
+    if v == 're':
+        return convert_prompt_re(prompt)
+
+    if ':' not in prompt:
+        return prompt, None, None
+
+    splits = prompt.replace('.', '').strip().split(' ')
+    # v1: a photo of a 0:1 1:24 ...
+    # v2: a photo of a <0:1> <1:24> ...
+    ints = []
+    noncode_start = ''
+    noncode_end = ''
+    parts = ''
+    parts_i = []
+    split_tokens = []
+    for b in splits:
+        if ':' not in b:
+            split_tokens.append(b)
+            continue
+
+        if v == 'v1':
+            i, b = b.strip().split(':')
+            has_comma = ',' in b
+            if has_comma:
+                b = b[:-1]
+            intb = int(b)
+            parts += f'<part>_{i} '
+            split_tokens.append(f'<part>_{i}')
+            if has_comma:
+                split_tokens.append(',')
+        else:
+            if b[0] == '<':
+                if '>' not in b:  # no closing >, ignore
+                    split_tokens.append(b)
+                    continue
+
+                i, b = b[1:].strip().split(':')
+                token_to_add = ''
+                if b[-1] in [',', '.']:
+                    token_to_add = b[-1]
+                    b = b[:-1]
+
+                if b[-1] == '>':
+                    b = b[:-1]
+                else:  # not >, just search for the first >
+                    for ci, char in enumerate(b):
+                        if char == '>':
+                            token_to_add = b[ci + 1:] + token_to_add
+                            b = b[:ci]  # skip >
+                            break
+            else:  # has : but not start with <
+                split_tokens.append(b)
+                continue
+
+            intb = abs(int(b))  # just force negative one to positive
+
+            parts += f'<part>_{i} '
+            split_tokens.append(f'<part>_{i}')
+            if len(token_to_add) != 0:
+                split_tokens.append(token_to_add)
+
+        try:
+            int(i)
+        except:
+            raise ValueError(f'cannot cast `part` properly, please make sure input is correct')
+
+        parts_i.append(int(i))
+        ints.append(intb)
+
+    ints = torch.tensor(ints)  # (nparts,)
+
+    if replace_token:
+        new_caption = f'{noncode_start.strip()} <part> {noncode_end.strip()}'
+    else:
+        new_caption = ' '.join(split_tokens)
+
+    new_caption = new_caption.strip()
+
+    return new_caption, ints, parts_i
+
+
+class DreamCreatureSDPipeline(StableDiffusionPipeline):
+    def _maybe_convert_prompt(self, prompt: str, tokenizer: MultiTokenCLIPTokenizer):
+        r"""
+        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+        Parameters:
+            prompt (`str`):
+                The prompt to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str`: The converted prompt
+        """
+        if hasattr(self, 'replace_token'):
+            replace_token = self.replace_token
+        else:
+            replace_token = True
+
+        if hasattr(self, 'v'):
+            v = self.v
+        else:
+            v = 'v1'
+
+        new_caption, code, parts_i = convert_prompt(prompt, replace_token, v)
+        if hasattr(self, 'num_k_per_part'):
+            if code is not None and any(code >= self.num_k_per_part):
+                raise ValueError(f'`id` cannot more than {self.num_k_per_part}')
+
+        if hasattr(self, 'verbose') and self.verbose:
+            print(new_caption)
+
+        return new_caption, code, parts_i
+
+    def compute_prompt_embeddings(self, prompts, device):
+        # textual inversion: procecss multi-vector tokens if necessary
+        if not isinstance(prompts, List):
+            prompts = [prompts]
+
+        prompt_embeds_concat = []
+        for prompt in prompts:
+            prompt, code, parts_i = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+            if hasattr(self, 'replace_token'):
+                replace_token = self.replace_token
+            else:
+                replace_token = True
+
+            text_inputs = self.tokenizer(
+                prompt,
+                replace_token=replace_token,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            if hasattr(self, 'verbose') and self.verbose:
+                print(text_input_ids)
+
+            untruncated_ids = self.tokenizer(prompt,
+                                             replace_token=replace_token,
+                                             padding="longest",
+                                             return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            if code is None:
+                modified_hs = None
+            else:
+                placeholder_token_ids = self.placeholder_token_ids
+                placeholder_token_ids = [placeholder_token_ids[i] for i in parts_i]  # follow the order of prompt's i
+                mapper_outputs = self.simple_mapper(code.unsqueeze(0).to(device), torch.tensor(parts_i).to(device))
+                modified_hs = self.text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids.to(device),
+                                                                                          None,
+                                                                                          mapper_outputs,
+                                                                                          placeholder_token_ids)
+
+            prompt_embeds = self.text_encoder(
+                text_input_ids.to(device),
+                attention_mask=attention_mask,
+                hidden_states=modified_hs
+            )
+            prompt_embeds = prompt_embeds[0]
+            prompt_embeds_concat.append(prompt_embeds)
+
+        if len(prompt_embeds_concat) == 1:
+            return prompt_embeds_concat[0]
+        else:
+            return torch.cat(prompt_embeds_concat, dim=0)
+
+    def encode_prompt(
+            self,
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt=None,
+            prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+            lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self.compute_prompt_embeddings(prompt, device)
+
+        # if self.text_encoder is not None:
+        #     prompt_embeds_dtype = self.text_encoder.dtype
+        # elif self.unet is not None:
+        #     prompt_embeds_dtype = self.unet.dtype
+        # else:
+        #     prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds_dtype = self.unet.dtype  # should be unet only because this is unet's condition input
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt] * batch_size
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            negative_prompt_embeds = []
+            for u_tokens in uncond_tokens:
+                negative_prompt_embeds.append(self.compute_prompt_embeddings(u_tokens, device))
+            negative_prompt_embeds = torch.cat(negative_prompt_embeds, dim=0)
+
+            # if isinstance(self, TextualInversionLoaderMixin):
+            #     uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+            # max_length = prompt_embeds.shape[1]
+            # uncond_input = self.tokenizer(
+            #     uncond_tokens,
+            #     padding="max_length",
+            #     max_length=max_length,
+            #     truncation=True,
+            #     return_tensors="pt",
+            # )
+            #
+            # if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+            #     attention_mask = uncond_input.attention_mask.to(device)
+            # else:
+            #     attention_mask = None
+            #
+            # negative_prompt_embeds = self.text_encoder(
+            #     uncond_input.input_ids.to(device),
+            #     attention_mask=attention_mask,
+            # )
+            # negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    @torch.no_grad()
+    def obtain_attention_map(self, image, prompt, timesteps):
+        prompt, codes, index = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+        if hasattr(self, 'replace_token'):
+            replace_token = self.replace_token
+        else:
+            replace_token = True
+
+        text_inputs = self.tokenizer(
+            prompt,
+            replace_token=replace_token,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        input_ids = text_inputs.input_ids
+
+        placeholder_token_ids = self.placeholder_token_ids
+        placeholder_token_ids = [placeholder_token_ids[i] for i in index]
+
+        # forward an image, denoise it and obtain the attention map
+        device = self._execution_device
+
+        latents = self.vae.encode(image.to(device, dtype=self.weight_dtype)).latent_dist.sample()
+        latents = latents * self.vae.config.scaling_factor
+
+        # bsz = latents.shape[0]
+        # Sample a random timestep for each image
+        # timesteps = torch.randint(0, self.noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
+        timesteps = timesteps.long().to(latents.device)
+
+        # Add noise to the latents according to the noise magnitude at each timestep
+        # (this is the forward diffusion process)
+        noise = torch.randn_like(latents)
+        noisy_latents = self.noise_scheduler.add_noise(latents, noise, timesteps)
+
+        mapper_outputs = self.simple_mapper(codes.unsqueeze(0).to(device), torch.tensor(index).to(device))
+        # print(mapper_outputs.size(), batch["input_ids"].size())
+        modified_hs = self.text_encoder.text_model.forward_embeddings_with_mapper(input_ids.to(device),
+                                                                                  None,
+                                                                                  mapper_outputs,
+                                                                                  placeholder_token_ids)
+        encoder_hidden_states = self.text_encoder(input_ids, hidden_states=modified_hs)[0]
+        model_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states.to(dtype=self.weight_dtype)).sample
+
+        attn_probs = {}
+
+        for name, module in self.unet.named_modules():
+            if isinstance(module, Attention) and module.attn_probs is not None:
+                a = module.attn_probs[0].mean(dim=0)  # (2,Head,H,W,77)->(H,W,77)
+                attn_probs[name] = a
+
+        avg_attn_map = []
+        for name in attn_probs:
+            avg_attn_map.append(attn_probs[name])
+        avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0)  # (5,B,H,W,77) -> (B,H,W,77)
+
+        return attn_probs, avg_attn_map, input_ids
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+            self,
+            prompt: Union[str, List[str]] = None,
+            height: Optional[int] = None,
+            width: Optional[int] = None,
+            num_inference_steps: int = 50,
+            guidance_scale: float = 7.5,
+            negative_prompt: Optional[Union[str, List[str]]] = None,
+            num_images_per_prompt: Optional[int] = 1,
+            eta: float = 0.0,
+            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+            latents: Optional[torch.FloatTensor] = None,
+            prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+            output_type: Optional[str] = "pil",
+            return_dict: bool = True,
+            callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+            callback_steps: int = 1,
+            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+            guidance_rescale: float = 0.0,
+            get_attention_map: bool = False
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that calls every `callback_steps` steps during inference. The function is called with the
+                following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function is called. If not specified, the callback is called at
+                every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.7):
+                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+                using zero terminal SNR.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+        )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            if get_attention_map:
+                attn_maps = {}  # each t one attn map
+
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    return_dict=False,
+                )[0]
+                if get_attention_map:
+                    attn_probs = {}
+
+                    for name, module in self.unet.named_modules():
+                        if isinstance(module, Attention) and module.attn_probs is not None:
+                            a = module.attn_probs[1].mean(dim=0)  # (2,Head,H,W,77)->(H,W,77)
+                            attn_probs[name] = a
+
+                    attn_maps[i] = attn_probs
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if do_classifier_free_guidance and guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        callback(i, t, latents)
+
+        if get_attention_map:
+            output_maps = {}
+            for name in attn_probs.keys():
+                timeavg_maps = []
+                for i in attn_maps.keys():
+                    timeavg_maps.append(attn_maps[i][name])
+                timeavg_maps = torch.stack(timeavg_maps, dim=0).mean(dim=0)
+                output_maps[name] = timeavg_maps
+
+            avg_attn_map = []
+            for name in attn_probs:
+                avg_attn_map.append(attn_probs[name])
+            avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0)  # (5,B,H,W,77) -> (B,H,W,77)
+            output_maps['avg'] = avg_attn_map
+
+            del attn_maps
+            del attn_probs
+            self.attn_maps = output_maps
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # Offload last model to CPU
+        if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+            self.final_offload_hook.offload()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+
+def create_args(output_dir, num_parts=8, num_k_per_part=256):
+    args = OmegaConf.create({
+        'pretrained_model_name_or_path': 'runwayml/stable-diffusion-v1-5',
+        'num_parts': num_parts,
+        'num_k_per_part': num_k_per_part,
+        'revision': None,
+        'variant': None,
+        'rank': 4,
+        'projection_nlayers': 1,
+        'output_dir': output_dir
+    })
+    folders = sorted(os.listdir(args.output_dir))
+    cps = [int(f.split('-')[1]) for f in folders if 'checkpoint' in f and '.ipynb' not in f]
+    maxcp = max(cps)
+
+    args.maxcp = maxcp
+    args.unet_path = None
+    return args
+
+
+def load_pipeline(args, weight_dtype=torch.float16, device=torch.device('cuda')):
+    tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
+    )
+    text_encoder = CustomCLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+    )
+    unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
+    unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
+        unet_path, subfolder="unet", revision=args.revision
+    )
+    pipeline = DreamCreatureSDPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        unet=unet,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        revision=args.revision,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.num_k_per_part = args.num_k_per_part
+    pipeline.num_parts = args.num_parts
+    placeholder_token = "<part>"
+    initializer_token = None
+    placeholder_token_ids = add_tokens(tokenizer,
+                                       text_encoder,
+                                       placeholder_token,
+                                       args.num_parts,
+                                       initializer_token)
+    pipeline.placeholder_token_ids = placeholder_token_ids
+    pipeline.simple_mapper = TokenMapper(args.num_parts,
+                                         args.num_k_per_part,
+                                         768,
+                                         args.projection_nlayers)
+    pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + f'/checkpoint-{args.maxcp}/pytorch_model_1.bin',
+                                                      map_location='cpu'))
+    pipeline.simple_mapper.to(device)
+    pipeline.replace_token = False
+    pipeline = pipeline.to(device)
+
+    # load attention processors
+    # pipeline.unet.load_attn_procs(args.output_dir, use_safetensors=not args.custom_diffusion)
+    setup_attn_processor(pipeline.unet, rank=args.rank)
+    load_attn_processor(pipeline.unet, args.output_dir + f'/checkpoint-{args.maxcp}/pytorch_model.bin')
+    return pipeline

dreamcreature/pipeline_xl.py

@@ -0,0 +1,895 @@
+import os
+
+from diffusers.models.attention_processor import Attention
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import *
+from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import (
+    StableDiffusionXLPipeline,
+    StableDiffusionXLPipelineOutput,
+    XLA_AVAILABLE,
+    Tuple,
+    StableDiffusionXLLoraLoaderMixin,
+    PipelineImageInput
+)
+from omegaconf import OmegaConf
+
+from dreamcreature.attn_processor import AttnProcessorCustom
+from dreamcreature.mapper import TokenMapper
+from dreamcreature.pipeline import convert_prompt
+from dreamcreature.text_encoder import CustomCLIPTextModel, CustomCLIPTextModelWithProjection
+from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
+from utils import add_tokens
+
+
+def init_for_pipeline(args):
+    tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer",
+        revision=args.revision,
+        use_fast=False,
+    )
+    tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer_2",
+        revision=args.revision,
+        use_fast=False,
+    )
+    text_encoder_cls_one = CustomCLIPTextModel
+    text_encoder_cls_two = CustomCLIPTextModelWithProjection
+    text_encoder_one = text_encoder_cls_one.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    text_encoder_two = text_encoder_cls_two.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
+    )
+
+    OUT_DIMS = 768 + 1280  # 2048
+    simple_mapper = TokenMapper(args.num_parts,
+                                args.num_k_per_part,
+                                OUT_DIMS,
+                                args.projection_nlayers)
+    return text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper
+
+
+class DreamCreatureSDXLPipeline(StableDiffusionXLPipeline):
+    def _maybe_convert_prompt(self, prompt: str, tokenizer: MultiTokenCLIPTokenizer):
+        r"""
+        Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+        to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+        is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+        inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+        Parameters:
+            prompt (`str`):
+                The prompt to guide the image generation.
+            tokenizer (`PreTrainedTokenizer`):
+                The tokenizer responsible for encoding the prompt into input tokens.
+
+        Returns:
+            `str`: The converted prompt
+        """
+        if hasattr(self, 'replace_token'):
+            replace_token = self.replace_token
+        else:
+            replace_token = True
+
+        new_caption, code, parts_i = convert_prompt(prompt, replace_token)
+
+        if hasattr(self, 'verbose') and self.verbose:
+            print(new_caption)
+
+        return new_caption, code, parts_i
+
+    def compute_prompt_embeddings(self, prompts, text_encoder, tokenizer, device, index):
+        # textual inversion: procecss multi-vector tokens if necessary
+        if not isinstance(prompts, List):
+            prompts = [prompts]
+
+        prompt_embeds_concat = []
+        pooled_prompt_embeds_concat = []
+        for prompt in prompts:
+            prompt, code, parts_i = self.maybe_convert_prompt(prompt, tokenizer)
+
+            if hasattr(self, 'replace_token'):
+                replace_token = self.replace_token
+            else:
+                replace_token = True
+
+            text_inputs = tokenizer(
+                prompt,
+                replace_token=replace_token,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            if hasattr(self, 'verbose') and self.verbose:
+                print(text_input_ids)
+
+            untruncated_ids = tokenizer(prompt,
+                                        replace_token=replace_token,
+                                        padding="longest",
+                                        return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+            ):
+                removed_text = tokenizer.batch_decode(
+                    untruncated_ids[:, tokenizer.model_max_length - 1: -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(text_encoder.config, "use_attention_mask") and text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            if code is None:
+                modified_hs = None
+            else:
+                placeholder_token_ids = self.placeholder_token_ids
+                placeholder_token_ids = [placeholder_token_ids[i] for i in parts_i]  # follow the order of prompt's i
+                mapper_outputs = self.simple_mapper(code.unsqueeze(0).to(device), torch.tensor(parts_i).to(device))
+                if index == 0:  # first encoder
+                    mapper_outputs = mapper_outputs[..., :768]
+                else:
+                    mapper_outputs = mapper_outputs[..., 768:]
+                modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids.to(device),
+                                                                                     None,
+                                                                                     mapper_outputs,
+                                                                                     placeholder_token_ids)
+
+            prompt_embeds = text_encoder(
+                text_input_ids.to(device),
+                output_hidden_states=True,
+                attention_mask=attention_mask,
+                hidden_states=modified_hs
+            )
+            # We are only ALWAYS interested in the pooled output of the final text encoder
+            pooled_prompt_embeds = prompt_embeds[0]
+            prompt_embeds = prompt_embeds.hidden_states[-2]
+
+            pooled_prompt_embeds_concat.append(pooled_prompt_embeds)
+            prompt_embeds_concat.append(prompt_embeds)
+
+        if len(prompt_embeds_concat) == 1:
+            return prompt_embeds_concat[0], pooled_prompt_embeds_concat[0]
+        else:
+            return torch.cat(prompt_embeds_concat, dim=0), torch.cat(pooled_prompt_embeds_concat, dim=0)
+
+    def encode_prompt(
+            self,
+            prompt: str,
+            prompt_2: Optional[str] = None,
+            device: Optional[torch.device] = None,
+            num_images_per_prompt: int = 1,
+            do_classifier_free_guidance: bool = True,
+            negative_prompt: Optional[str] = None,
+            negative_prompt_2: Optional[str] = None,
+            prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+            lora_scale: Optional[float] = None,
+            clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder, lora_scale)
+
+            if self.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: procecss multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            index = 0
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                prompt_embeds, pooled_prompt_embeds = self.compute_prompt_embeddings(prompt,
+                                                                                     text_encoder,
+                                                                                     tokenizer,
+                                                                                     device,
+                                                                                     index)
+                prompt_embeds_list.append(prompt_embeds)
+                index += 1
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            index = 0
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                negative_prompt_embeds, negative_pooled_prompt_embeds = self.compute_prompt_embeddings(negative_prompt,
+                                                                                                       text_encoder,
+                                                                                                       tokenizer,
+                                                                                                       device,
+                                                                                                       index)
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+                index += 1
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if self.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if self.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+            self,
+            prompt: Union[str, List[str]] = None,
+            prompt_2: Optional[Union[str, List[str]]] = None,
+            height: Optional[int] = None,
+            width: Optional[int] = None,
+            num_inference_steps: int = 50,
+            timesteps: List[int] = None,
+            denoising_end: Optional[float] = None,
+            guidance_scale: float = 5.0,
+            negative_prompt: Optional[Union[str, List[str]]] = None,
+            negative_prompt_2: Optional[Union[str, List[str]]] = None,
+            num_images_per_prompt: Optional[int] = 1,
+            eta: float = 0.0,
+            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+            latents: Optional[torch.FloatTensor] = None,
+            prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+            ip_adapter_image: Optional[PipelineImageInput] = None,
+            output_type: Optional[str] = "pil",
+            return_dict: bool = True,
+            cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+            guidance_rescale: float = 0.0,
+            original_size: Optional[Tuple[int, int]] = None,
+            crops_coords_top_left: Tuple[int, int] = (0, 0),
+            target_size: Optional[Tuple[int, int]] = None,
+            negative_original_size: Optional[Tuple[int, int]] = None,
+            negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+            negative_target_size: Optional[Tuple[int, int]] = None,
+            clip_skip: Optional[int] = None,
+            callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+            callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+            get_attention_map: bool = False,
+            **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if self.do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+                image_embeds = image_embeds.to(device)
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8.1 Apply denoising_end
+        if (
+                self.denoising_end is not None
+                and isinstance(self.denoising_end, float)
+                and self.denoising_end > 0
+                and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        # 9. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            if get_attention_map:
+                attn_maps = {}  # each t one attn map
+
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                if ip_adapter_image is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+                if get_attention_map:
+                    attn_probs = {}
+
+                    for name, module in self.unet.named_modules():
+                        if isinstance(module, Attention) and module.attn_probs is not None:
+                            # take 1 because we are taking the noise_pred_text
+                            a = module.attn_probs[1].mean(dim=0)  # (2,Head,H,W,77)->(H,W,77)
+                            attn_probs[name] = a
+
+                    attn_maps[i] = attn_probs
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+                if XLA_AVAILABLE:
+                    import torch_xla.core.xla_model as xm
+                    xm.mark_step()
+
+        if get_attention_map:
+            output_maps = {}
+            for name in attn_probs.keys():
+                timeavg_maps = []
+                for i in attn_maps.keys():
+                    timeavg_maps.append(attn_maps[i][name])
+                timeavg_maps = torch.stack(timeavg_maps, dim=0).mean(dim=0)
+                output_maps[name] = timeavg_maps
+
+            avg_attn_map = []
+            for name in attn_probs:
+                avg_attn_map.append(attn_probs[name])
+            avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0)  # (5,B,H,W,77) -> (B,H,W,77)
+            output_maps['avg'] = avg_attn_map
+
+            del attn_maps
+            del attn_probs
+            self.attn_maps = output_maps
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)
+
+
+def create_args_xl(output_dir, num_parts=8, num_k_per_part=256):
+    args = OmegaConf.create({
+        'pretrained_model_name_or_path': 'stabilityai/stable-diffusion-xl-base-1.0',
+        'num_parts': num_parts,
+        'num_k_per_part': num_k_per_part,
+        'revision': None,
+        'variant': None,
+        'rank': 4,
+        'projection_nlayers': 1,
+        'output_dir': output_dir
+    })
+    folders = sorted(os.listdir(args.output_dir))
+    cps = [int(f.split('-')[1]) for f in folders if 'checkpoint' in f and '.ipynb' not in f]
+    maxcp = max(cps)
+
+    args.maxcp = maxcp
+    return args
+
+
+def load_pipeline_xl(args, weight_dtype=torch.float16, device=torch.device('cuda')):
+    text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper = init_for_pipeline(args)
+    placeholder_token = "<part>"
+    initializer_token = None
+    placeholder_token_ids_one = add_tokens(tokenizer_one,
+                                           text_encoder_one,
+                                           placeholder_token,
+                                           args.num_parts,
+                                           initializer_token)
+    placeholder_token_ids_two = add_tokens(tokenizer_two,
+                                           text_encoder_two,
+                                           placeholder_token,
+                                           args.num_parts,
+                                           initializer_token)
+
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+    pipeline = DreamCreatureSDXLPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        vae=vae,
+        tokenizer=tokenizer_one,
+        tokenizer_2=tokenizer_two,
+        text_encoder=text_encoder_one,
+        text_encoder_2=text_encoder_two,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.placeholder_token_ids = placeholder_token_ids_one
+    pipeline.replace_token = False
+    pipeline.simple_mapper = simple_mapper
+    pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + f'/checkpoint-{args.maxcp}/hash_mapper.pth',
+                                                      map_location='cpu'))
+
+    pipeline.simple_mapper.to(device)
+    pipeline = pipeline.to(device)
+
+    # load attention processors
+    pipeline.load_lora_weights(args.output_dir + f'/checkpoint-{args.maxcp}')
+
+    def setup_attn_processors(unet, attn_size):
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            attn_procs[name] = AttnProcessorCustom(attn_size)
+        unet.set_attn_processor(attn_procs)
+
+    pipeline = pipeline.to(weight_dtype)
+    setup_attn_processors(pipeline.unet, 16)
+
+    return pipeline

dreamcreature/text_encoder.py

@@ -0,0 +1,201 @@
+from collections import defaultdict
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+from transformers.models.clip.modeling_clip import (CLIPTextConfig,
+                                                    CLIPTextModel,
+                                                    CLIPTextModelWithProjection,
+                                                    CLIPTextModelOutput)
+from transformers.models.clip.modeling_clip import (CLIPTextTransformer,
+                                                    _prepare_4d_attention_mask,
+                                                    _create_4d_causal_attention_mask)
+
+
+class CustomCLIPTextModel(CLIPTextModel):
+    """ Modification of CLIPTextModel to use our NeTI mapper for computing the embeddings of the concept. """
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CustomCLIPTextTransformer(config)
+        self.post_init()
+
+    def forward(self, input_ids: Optional[torch.Tensor] = None,
+                attention_mask: Optional[torch.Tensor] = None,
+                position_ids: Optional[torch.Tensor] = None,
+                output_attentions: Optional[bool] = None,
+                output_hidden_states: Optional[bool] = None,
+                return_dict: Optional[bool] = None,
+                hidden_states: Optional[torch.Tensor] = None) -> Union[Tuple, BaseModelOutputWithPooling]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.text_model.forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            hidden_states=hidden_states
+        )
+
+
+class CustomCLIPTextTransformer(CLIPTextTransformer):
+    def forward_embeddings(self, input_ids, position_ids, inputs_embeds):
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids, inputs_embeds=inputs_embeds)
+        return hidden_states
+
+    def forward_embeddings_with_mapper(self, input_ids, position_ids, mapper_outputs, placeholder_token_ids):
+        inputs_embeds = self.embeddings.token_embedding(input_ids)
+        dtype = inputs_embeds.dtype
+
+        offset = defaultdict(int)
+        for i, placeholder_token_id in enumerate(placeholder_token_ids):
+            # Overwrite the index of the placeholder token with the mapper output for each entry in the batch
+            # learnable_idxs = (input_ids == placeholder_token_id).nonzero(as_tuple=True)[1]
+            # inputs_embeds[torch.arange(input_ids.shape[0]), learnable_idxs] = mapper_outputs[:, i].to(dtype)
+
+            for bi in range(input_ids.shape[0]):
+                learnable_idx = (input_ids[bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
+
+                if len(learnable_idx) != 0:  # only assign if found
+                    if len(learnable_idx) == 1:
+                        offset_learnable_idx = learnable_idx
+                    else:  # if there is two and above.
+                        start = offset[(bi, placeholder_token_id)]
+                        offset_learnable_idx = learnable_idx[start:start + 1]
+                        offset[(bi, placeholder_token_id)] += 1
+
+                    # print(i, offset_learnable_idx)
+
+                    # before = inputs_embeds[bi, learnable_idx]
+                    inputs_embeds[bi, offset_learnable_idx] = mapper_outputs[bi, i].to(dtype)
+                    # after = inputs_embeds[bi, learnable_idx]
+
+        return self.forward_embeddings(input_ids, position_ids, inputs_embeds)
+
+    def forward(
+            self,
+            input_ids: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            hidden_states: Optional[torch.Tensor] = None
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is None:
+            raise ValueError("You have to specify either input_ids")
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+        if hidden_states is None:
+            hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = _create_4d_causal_attention_mask(
+            input_shape, hidden_states.dtype, device=hidden_states.device
+        )
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
+
+        # # bsz, seq_len = input_shape
+        # # CLIP's text model uses causal mask, prepare it here.
+        # # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        # causal_attention_mask = _make_causal_mask(input_shape, hidden_states.dtype, device=hidden_states.device)
+        # # causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(
+        # #     hidden_states.device
+        # # )
+        #
+        # # expand attention_mask
+        # if attention_mask is not None:
+        #     # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        #     attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+        pooled_output = last_hidden_state[
+            torch.arange(last_hidden_state.shape[0], device=input_ids.device), input_ids.to(torch.int).argmax(dim=-1)
+        ]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class CustomCLIPTextModelWithProjection(CLIPTextModelWithProjection):
+    """ Modification of CLIPTextModel to use our NeTI mapper for computing the embeddings of the concept. """
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CustomCLIPTextTransformer(config)
+        self.post_init()
+
+    def forward(self, input_ids: Optional[torch.Tensor] = None,
+                attention_mask: Optional[torch.Tensor] = None,
+                position_ids: Optional[torch.Tensor] = None,
+                output_attentions: Optional[bool] = None,
+                output_hidden_states: Optional[bool] = None,
+                return_dict: Optional[bool] = None,
+                hidden_states: Optional[torch.Tensor] = None) -> Union[Tuple, CLIPTextModelOutput]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            hidden_states=hidden_states
+        )
+
+        pooled_output = text_outputs[1]
+
+        text_embeds = self.text_projection(pooled_output)
+
+        if not return_dict:
+            outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return CLIPTextModelOutput(
+            text_embeds=text_embeds,
+            last_hidden_state=text_outputs.last_hidden_state,
+            hidden_states=text_outputs.hidden_states,
+            attentions=text_outputs.attentions,
+        )

dreamcreature/tokenizer.py

@@ -0,0 +1,109 @@
+"""
+The main idea for this code is to provide a way for users to not need to bother with the hassle of multiple tokens for a concept by typing
+a photo of <concept>_0 <concept>_1 ... and so on
+and instead just do
+a photo of <concept>
+which gets translated to the above. This needs to work for both inference and training.
+For inference,
+the tokenizer encodes the text. So, we would want logic for our tokenizer to replace the placeholder token with
+it's underlying vectors
+For training,
+we would want to abstract away some logic like
+1. Adding tokens
+2. Updating gradient mask
+3. Saving embeddings
+to our Util class here.
+so
+TODO:
+1. have tokenizer keep track of concept, multiconcept pairs and replace during encode call x
+2. have mechanism for adding tokens x
+3. have mech for saving emebeddings x
+4. get mask to update x
+5. Loading tokens from embedding x
+6. Integrate to training x
+7. Test
+"""
+import copy
+import random
+
+from transformers import CLIPTokenizer
+
+
+class MultiTokenCLIPTokenizer(CLIPTokenizer):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.token_map = {}
+
+    def try_adding_tokens(self, placeholder_token, *args, **kwargs):
+        num_added_tokens = super().add_tokens(placeholder_token, *args, **kwargs)
+        if num_added_tokens == 0:
+            raise ValueError(
+                f"The tokenizer already contains the token {placeholder_token}. Please pass a different"
+                " `placeholder_token` that is not already in the tokenizer."
+            )
+
+    def add_placeholder_tokens(self, placeholder_token, *args, num_vec_per_token=1, **kwargs):
+        output = []
+        # if num_vec_per_token == 1:
+        #     self.try_adding_tokens(placeholder_token, *args, **kwargs)
+        #     output.append(placeholder_token)
+        # else:
+        output = []
+        for i in range(num_vec_per_token):
+            ith_token = placeholder_token + f"_{i}"
+            self.try_adding_tokens(ith_token, *args, **kwargs)
+            output.append(ith_token)
+        # handle cases where there is a new placeholder token that contains the current placeholder token but is larger
+        for token in self.token_map:
+            if token in placeholder_token:
+                raise ValueError(
+                    f"The tokenizer already has placeholder token {token} that can get confused with"
+                    f" {placeholder_token}keep placeholder tokens independent"
+                )
+        self.token_map[placeholder_token] = output
+
+    def replace_placeholder_tokens_in_text(self, text, vector_shuffle=False, prop_tokens_to_load=1.0):
+        """
+        Here, we replace the placeholder tokens in text recorded in token_map so that the text_encoder
+        can encode them
+        vector_shuffle was inspired by https://github.com/rinongal/textual_inversion/pull/119
+        where shuffling tokens were found to force the model to learn the concepts more descriptively.
+        """
+        if isinstance(text, list):
+            output = []
+            for i in range(len(text)):
+                output.append(self.replace_placeholder_tokens_in_text(text[i], vector_shuffle=vector_shuffle))
+            return output
+        for placeholder_token in self.token_map:
+            if placeholder_token in text:
+                tokens = self.token_map[placeholder_token]
+                tokens = tokens[: 1 + int(len(tokens) * prop_tokens_to_load)]
+                if vector_shuffle:
+                    tokens = copy.copy(tokens)
+                    random.shuffle(tokens)
+                text = text.replace(placeholder_token, " ".join(tokens))  # <part>_0 <part>_1 -> <part>
+        return text
+
+    def __call__(self, text, *args, vector_shuffle=False, prop_tokens_to_load=1.0, replace_token=True, **kwargs):
+        if replace_token:
+            return super().__call__(
+                self.replace_placeholder_tokens_in_text(
+                    text, vector_shuffle=vector_shuffle, prop_tokens_to_load=prop_tokens_to_load
+                ),
+                *args,
+                **kwargs,
+            )
+        else:
+            return super().__call__(text, *args, **kwargs)
+
+    def encode(self, text, *args, vector_shuffle=False, prop_tokens_to_load=1.0, replace_token=True, **kwargs):
+        if replace_token:
+            return super().encode(
+                self.replace_placeholder_tokens_in_text(
+                    text, vector_shuffle=vector_shuffle, prop_tokens_to_load=prop_tokens_to_load
+                ),
+                *args,
+                **kwargs,
+            )
+        else:
+            return super().encoder(text, *args, **kwargs)

gradio_demo_cub200.py

@@ -0,0 +1,166 @@
+import argparse
+import gc
+import os
+import re
+import shutil
+
+import gradio as gr
+import requests
+import torch
+
+from dreamcreature.pipeline import create_args, load_pipeline
+
+
+def download_file(url, local_path):
+    if os.path.exists(local_path):
+        return
+
+    with requests.get(url, stream=True) as r:
+        with open(local_path, 'wb') as f:
+            shutil.copyfileobj(r.raw, f)
+
+    # Example usage
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name', default='dreamcreature-sd1.5-cub200')
+parser.add_argument('--checkpoint', default='checkpoint-74900')
+opt = parser.parse_args()
+
+model_name = opt.model_name
+checkpoint_name = opt.checkpoint
+
+repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
+file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
+local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
+os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
+download_file(file_url, local_path)
+
+file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
+local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
+download_file(file_url, local_path)
+
+OUTPUT_DIR = model_name
+
+args = create_args(OUTPUT_DIR)
+if 'dpo' in OUTPUT_DIR:
+    args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
+
+pipe = load_pipeline(args, torch.float16, 'cuda')
+pipe = pipe.to(torch.float16)
+
+pipe.verbose = True
+pipe.v = 're'
+pipe.num_k_per_part = 200
+
+MAPPING = {
+    'body': 0,
+    'tail': 1,
+    'head': 2,
+    'wing': 4,
+    'leg': 6
+}
+
+ID2NAME = open('data/cub200_2011/class_names.txt').readlines()
+ID2NAME = [line.strip() for line in ID2NAME]
+
+
+def process_text(text):
+    pattern = r"<([^:>]+):(\d+)>"
+    result = text
+    offset = 0
+
+    part2id = []
+
+    for match in re.finditer(pattern, text):
+        key = match.group(1)
+        clsid = int(match.group(2))
+        clsid = min(max(clsid, 1), 200)  # must be 1~200
+
+        replacement = f"<{MAPPING[key]}:{clsid - 1}>"
+        start, end = match.span()
+
+        # Adjust the start and end positions based on the offset from previous replacements
+        start += offset
+        end += offset
+
+        # Replace the matched text with the replacement
+        result = result[:start] + replacement + result[end:]
+
+        # Update the offset for the next replacement
+        offset += len(replacement) - (end - start)
+
+        part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
+
+    return result, part2id
+
+
+def generate_images(prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+
+    try:
+        prompt, part2id = process_text(prompt)
+        negative_prompt, _ = process_text(negative_prompt)
+
+        images = pipe(prompt,
+                      negative_prompt=negative_prompt, generator=generator,
+                      num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                      num_images_per_prompt=num_images).images
+    except Exception as e:
+        raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
+                       f"The error message: {e}")
+    finally:
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    return images, '; '.join(part2id)
+
+
+with gr.Blocks(title="DreamCreature") as demo:
+    with gr.Row():
+        gr.Markdown(
+            """
+            # DreamCreature (CUB-200-2011)
+            To create your own creature, you can type:
+
+            `"a photo of a <head:id> <wing:id> bird"` where `id` ranges from 1~200 (200 classes corresponding to CUB-200-2011)
+
+            For instance `"a photo of a <head:17> <wing:18> bird"` using head of `cardinal (17)` and wing of `spotted catbird (18)`
+
+            Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/cub200_2011/class_names.txt
+
+            You can also try any prompt you like such as:
+
+            Sub-concept transfer: `"a photo of a <wing:17> cat"`
+
+            Inspiring design: `"a photo of a <head:101> <wing:191> teddy bear"`
+
+            (Experimental) You can also use two parts together such as:
+
+            `"a photo of a <head:17> <head:18> bird"` mixing head of `cardinal (17)` and `spotted catbird (18)`
+
+            The current available parts are: `head`, `body`, `wing`, `tail`, and `leg`
+
+            """)
+    with gr.Column():
+        with gr.Row():
+            with gr.Group():
+                prompt = gr.Textbox(label="Prompt", value="a photo of a <head:101> <wing:191> teddy bear")
+                negative_prompt = gr.Textbox(label="Negative Prompt",
+                                             value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+                num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
+                guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+                num_images = gr.Slider(minimum=1, maximum=4, step=1, value=4, label="Number of Images")
+                seed = gr.Number(label="Seed", value=777881414)
+                button = gr.Button()
+
+            with gr.Column():
+                output_images = gr.Gallery(columns=4, label='Output')
+                markdown_labels = gr.Markdown("")
+
+    button.click(fn=generate_images,
+                 inputs=[prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
+                         seed], outputs=[output_images, markdown_labels], show_progress=True)
+
+demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')

gradio_demo_dog.py

@@ -0,0 +1,166 @@
+import argparse
+import gc
+import os
+import re
+import shutil
+
+import gradio as gr
+import requests
+import torch
+
+from dreamcreature.pipeline import create_args, load_pipeline
+
+
+def download_file(url, local_path):
+    if os.path.exists(local_path):
+        return
+
+    with requests.get(url, stream=True) as r:
+        with open(local_path, 'wb') as f:
+            shutil.copyfileobj(r.raw, f)
+
+    # Example usage
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--model_name', default='dreamcreature-sd1.5-dog')
+parser.add_argument('--checkpoint', default='checkpoint-150000')
+opt = parser.parse_args()
+
+model_name = opt.model_name
+checkpoint_name = opt.checkpoint
+
+repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
+file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
+local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
+os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
+download_file(file_url, local_path)
+
+file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
+local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
+download_file(file_url, local_path)
+
+OUTPUT_DIR = model_name
+
+args = create_args(OUTPUT_DIR)
+if 'dpo' in OUTPUT_DIR:
+    args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
+
+pipe = load_pipeline(args, torch.float16, 'cuda')
+pipe = pipe.to(torch.float16)
+
+pipe.verbose = True
+pipe.v = 're'
+pipe.num_k_per_part = 120
+
+MAPPING = {
+    'eye': 0,
+    'neck': 2,
+    'ear': 3,
+    'body': 4,
+    'leg': 5,
+    'nose': 6,
+    'forehead': 7
+}
+
+ID2NAME = open('data/dogs/class_names.txt').readlines()
+ID2NAME = [line.strip() for line in ID2NAME]
+
+
+def process_text(text):
+    pattern = r"<([^:>]+):(\d+)>"
+    result = text
+    offset = 0
+
+    part2id = []
+
+    for match in re.finditer(pattern, text):
+        key = match.group(1)
+        clsid = int(match.group(2))
+        clsid = min(max(clsid, 1), 200)  # must be 1~200
+
+        replacement = f"<{MAPPING[key]}:{clsid - 1}>"
+        start, end = match.span()
+
+        # Adjust the start and end positions based on the offset from previous replacements
+        start += offset
+        end += offset
+
+        # Replace the matched text with the replacement
+        result = result[:start] + replacement + result[end:]
+
+        # Update the offset for the next replacement
+        offset += len(replacement) - (end - start)
+
+        part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
+
+    return result, part2id
+
+
+def generate_images(prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
+    generator = torch.Generator(device='cuda')
+    generator = generator.manual_seed(int(seed))
+
+    try:
+        prompt, part2id = process_text(prompt)
+        negative_prompt, _ = process_text(negative_prompt)
+
+        images = pipe(prompt,
+                      negative_prompt=negative_prompt, generator=generator,
+                      num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
+                      num_images_per_prompt=num_images).images
+    except Exception as e:
+        raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
+                       f"The error message: {e}")
+    finally:
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    return images, '; '.join(part2id)
+
+
+with gr.Blocks(title="DreamCreature") as demo:
+    with gr.Row():
+        gr.Markdown(
+            """
+            # DreamCreature (Stanford Dogs)
+            To create your own creature, you can type:
+            
+            `"a photo of a <nose:id> <ear:id> dog"` where `id` ranges from 0~119 (120 classes corresponding to Stanford Dogs)
+            
+            For instance `"a photo of a <nose:2> <ear:112> dog"` using head of `maltese dog (2)` and wing of `cardigan (112)`
+            
+            Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/dogs/class_names.txt
+
+            Sub-concept transfer: `"a photo of a <ear:112> cat"`
+
+            Inspiring design: `"a photo of a <eye:38> <body:38> teddy bear"`
+            
+            (Experimental) You can also use two parts together such as:
+            
+            `"a photo of a <nose:1> <nose:112> dog"` mixing head of `maltese dog (2)` and `spotted cardigan (112)`
+            
+            The current available parts are: `eye`, `neck`, `ear`, `body`, `leg`, `nose` and `forehead`
+            
+            """)
+    with gr.Column():
+        with gr.Row():
+            with gr.Group():
+                prompt = gr.Textbox(label="Prompt", value="a photo of a <eye:37> <body:37> teddy bear")
+                negative_prompt = gr.Textbox(label="Negative Prompt",
+                                             value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
+                num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
+                guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
+                num_images = gr.Slider(minimum=1, maximum=4, step=1, value=1, label="Number of Images")
+                seed = gr.Number(label="Seed", value=777881414)
+                button = gr.Button()
+
+            with gr.Column():
+                output_images = gr.Gallery(columns=4, label='Output')
+                markdown_labels = gr.Markdown("")
+
+    button.click(fn=generate_images,
+                 inputs=[prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
+                         seed], outputs=[output_images, markdown_labels], show_progress=True)
+
+demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')

requirements.txt

@@ -0,0 +1,12 @@
+torch
+torchvision
+git+https://github.com/huggingface/diffusers
+transformers
+torchpq
+omegaconf
+scikit-learn
+faiss-cpu
+tqdm
+accelerate
+gradio
+huggingface_hub

run_sd_sup.sh

@@ -0,0 +1,12 @@
+python train_dreamcreature_sd.py \
+     --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
+     --train_data_dir=data/cub200_2011 \
+     --resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
+     --num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
+     --lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sd15-cub200-sup" \
+     --validation_prompt="a photo of a 0:16 1:16 2:16 4:16 6:16" \
+     --num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
+     --code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
+     --use_templates --vector_shuffle --snr_gamma=5 \
+     --attn_loss=0.01 --use_gt_label --bg_code=7 \
+     --resume_from_checkpoint="latest" --mixed_precision="fp16"

run_sd_unsup.sh

@@ -0,0 +1,12 @@
+python train_dreamcreature_sd.py \
+     --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
+     --train_data_dir=data/cub200_2011 \
+     --resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
+     --num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
+     --lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sd15-cub200-unsup" \
+     --validation_prompt="a photo of a 0:16 1:16 2:16 4:16 6:16" \
+     --num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
+     --code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
+     --use_templates --vector_shuffle --snr_gamma=5 \
+     --attn_loss=0.01 \
+     --resume_from_checkpoint="latest"

run_sdxl_sup.sh

@@ -0,0 +1,13 @@
+python train_dreamcreature_sdxl.py \
+     --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
+     --scheduler_steps 1000 \
+     --train_data_dir=data/cub200_2011 \
+     --resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
+     --num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
+     --lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sdxlbase-cub200-sup" \
+     --validation_prompt="a photo of a 0:1 2:1 3:1 4:1 5:1 6:1 7:1" \
+     --num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
+     --code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
+     --use_templates --vector_shuffle --snr_gamma=5 \
+     --attn_loss=0.1 --use_gt_label --bg_code=7 \
+     --resume_from_checkpoint="latest"

run_sdxl_unsup.sh

@@ -0,0 +1,13 @@
+python train_dreamcreature_sdxl.py \
+     --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
+     --scheduler_steps 1000 \
+     --train_data_dir=data/cub200_2011 \
+     --resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
+     --num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
+     --lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sdxlbase-cub200-unsup" \
+     --validation_prompt="a photo of a 0:1 2:1 3:1 4:1 5:1 6:1 7:1" \
+     --num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
+     --code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
+     --use_templates --vector_shuffle --snr_gamma=5 \
+     --attn_loss=0.1 \
+     --resume_from_checkpoint="latest"

train_dreamcreature_sd.py

@@ -0,0 +1,1122 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Fine-tuning script for Stable Diffusion for text2image with support for LoRA."""
+
+import argparse
+import copy
+import logging
+import math
+import os
+import random
+import shutil
+from pathlib import Path
+
+import datasets
+import diffusers
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import ProjectConfiguration, set_seed
+from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
+from diffusers.loaders import AttnProcsLayers
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import compute_snr
+from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils.import_utils import is_xformers_available
+from huggingface_hub import create_repo, upload_folder
+from packaging import version
+from torchvision import transforms
+from torchvision.transforms import InterpolationMode
+from tqdm.auto import tqdm
+
+from dreamcreature.attn_processor import LoRAAttnProcessorCustom
+from dreamcreature.dataset import DreamCreatureDataset
+from dreamcreature.dino import DINO
+from dreamcreature.kmeans_segmentation import KMeansSegmentation
+from dreamcreature.loss import dreamcreature_loss
+from dreamcreature.mapper import TokenMapper
+from dreamcreature.pipeline import DreamCreatureSDPipeline
+from dreamcreature.text_encoder import CustomCLIPTextModel
+from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
+from utils import add_tokens, tokenize_prompt, get_attn_processors
+
+imagenet_templates = [
+    "a photo of a {}",
+    "a rendering of a {}",
+    "a cropped photo of the {}",
+    "the photo of a {}",
+    "a photo of a clean {}",
+    "a photo of a dirty {}",
+    "a dark photo of the {}",
+    "a photo of my {}",
+    "a photo of the cool {}",
+    "a close-up photo of a {}",
+    "a bright photo of the {}",
+    "a cropped photo of a {}",
+    "a photo of the {}",
+    "a good photo of the {}",
+    "a photo of one {}",
+    "a close-up photo of the {}",
+    "a rendition of the {}",
+    "a photo of the clean {}",
+    "a rendition of a {}",
+    "a photo of a nice {}",
+    "a good photo of a {}",
+    "a photo of the nice {}",
+    "a photo of the small {}",
+    "a photo of the weird {}",
+    "a photo of the large {}",
+    "a photo of a cool {}",
+    "a photo of a small {}",
+]
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.21.0.dev0")
+
+logger = get_logger(__name__, log_level="INFO")
+
+
+def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
+    img_str = ""
+    for i, image in enumerate(images):
+        image.save(os.path.join(repo_folder, f"image_{i}.png"))
+        img_str += f"![img_{i}](./image_{i}.png)\n"
+
+    yaml = f"""
+---
+license: creativeml-openrail-m
+base_model: {base_model}
+tags:
+- stable-diffusion
+- stable-diffusion-diffusers
+- text-to-image
+- diffusers
+- lora
+inference: true
+---
+    """
+    model_card = f"""
+# LoRA text2image fine-tuning - {repo_id}
+These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
+{img_str}
+"""
+    with open(os.path.join(repo_folder, "README.md"), "w") as f:
+        f.write(yaml + model_card)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help="Revision of pretrained model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default="text",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference."
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images that should be generated during validation with `validation_prompt`.",
+    )
+    parser.add_argument(
+        "--validation_epochs",
+        type=int,
+        default=1,
+        help=(
+            "Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`."
+        ),
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-model-finetuned-lora",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
+            " cropped. The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--snr_gamma",
+        type=float,
+        default=None,
+        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
+             "More details here: https://arxiv.org/abs/2303.09556.",
+    )
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=(
+            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
+        ),
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--prediction_type",
+        type=str,
+        default=None,
+        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
+    )
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=500,
+        help=(
+            "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
+            " training using `--resume_from_checkpoint`."
+        ),
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help=(
+            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
+            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
+        ),
+    )
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=4,
+        help=("The dimension of the LoRA update matrices."),
+    )
+
+    parser.add_argument('--filename', default='train.txt')
+    parser.add_argument('--code_filename', default='train_caps_better_m8_k256.txt')
+    parser.add_argument('--repeat', default=1, type=int)
+
+    parser.add_argument('--scheduler_steps', default=1000, type=int, help='scheduler step, if turbo, set to 4')
+    parser.add_argument('--num_parts', type=int, default=4, help="Number of parts")
+    parser.add_argument('--num_k_per_part', type=int, default=256, help='Number of k')
+
+    parser.add_argument('--mapper_lr_scale', default=1, type=float)
+    parser.add_argument('--mapper_lr', default=0.0001, type=float)
+    parser.add_argument('--attn_loss', default=0, type=float)
+    parser.add_argument('--projection_nlayers', default=3, type=int)
+
+    parser.add_argument('--masked_training', action='store_true')
+    parser.add_argument('--drop_tokens', action='store_true')
+    parser.add_argument('--drop_rate', type=float, default=0.5)
+    parser.add_argument('--drop_counts', default='half')
+
+    parser.add_argument('--class_name', default='')
+    parser.add_argument('--no_pe', action='store_true')
+    parser.add_argument('--vector_shuffle', action='store_true')
+    parser.add_argument('--use_templates', action='store_true')
+
+    parser.add_argument('--use_gt_label', action='store_true')
+    parser.add_argument('--bg_code', default=7, type=int)  # for gt_label
+    parser.add_argument('--fg_idx', default=0, type=int)  # for gt_label
+
+    parser.add_argument('--filter_class', default=None, type=int, help='debugging purpose')
+
+    parser.add_argument('--unet_path', default=None)
+
+    args = parser.parse_args()
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Need either a dataset name or a training folder.")
+
+    return args
+
+
+def collate_fn(args, tokenizer, placeholder_token):
+    train_resizecrop = transforms.Compose([
+        transforms.Resize(int(args.resolution), InterpolationMode.BILINEAR),
+        transforms.RandomCrop(args.resolution),
+    ])
+
+    train_transforms = transforms.Compose(
+        [
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    def f(examples):
+        raw_images = [train_resizecrop(example["pixel_values"]) for example in examples]
+
+        pixel_values = torch.stack([train_transforms(image) for image in raw_images])
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+        captions = []
+        appeared_tokens = []
+
+        for i in range(len(examples)):
+            if args.use_templates and random.random() <= 0.5:  # 50% using templates
+                if args.class_name != '':
+                    caption = random.choice(imagenet_templates).format(f'{placeholder_token} {args.class_name}')
+                else:
+                    caption = random.choice(imagenet_templates).format(placeholder_token)
+            else:
+                if args.class_name != '':
+                    caption = f'{placeholder_token} {args.class_name}'
+                else:
+                    caption = placeholder_token
+
+            tokens = tokenizer.token_map[placeholder_token][:args.num_parts]
+            tokens = [tokens[a] for a in examples[i]['appeared']]
+
+            if args.vector_shuffle or args.drop_tokens:
+                tokens = copy.copy(tokens)
+                random.shuffle(tokens)
+
+            if args.drop_tokens and random.random() < args.drop_rate and len(tokens) >= 2:
+                # randomly drop half of the tokens
+                if args.drop_counts == 'half':
+                    tokens = tokens[:len(tokens) // 2]
+                else:
+                    tokens = tokens[:int(args.drop_counts)]
+
+            appeared = [int(t.split('_')[1]) for t in tokens]  # <part>_i
+            appeared_tokens.append(appeared)
+
+            caption = caption.replace(placeholder_token, ' '.join(tokens))
+            captions.append(caption)
+
+        input_ids = tokenize_prompt(tokenizer, captions)
+        # input_ids = inputs.input_ids.repeat(len(examples), 1)  # (1, 77) -> (B, 77)
+
+        codes = torch.stack([example["codes"] for example in examples])
+
+        return {"pixel_values": pixel_values,
+                "raw_images": raw_images,
+                "appeared_tokens": appeared_tokens,
+                "input_ids": input_ids,
+                "codes": codes}
+
+    return f
+
+
+def setup_attn_processor(unet, **kwargs):
+    lora_attn_procs = {}
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+
+        lora_attn_procs[name] = LoRAAttnProcessorCustom(
+            hidden_size=hidden_size,
+            cross_attention_dim=cross_attention_dim,
+            rank=kwargs['rank'],
+        )
+
+    unet.set_attn_processor(lora_attn_procs)
+
+
+def load_attn_processor(unet, filename):
+    logger.info(f'Load attn processors from {filename}')
+    lora_layers = AttnProcsLayers(get_attn_processors(unet))
+    lora_layers.load_state_dict(torch.load(filename))
+
+
+def main():
+    args = parse_args()
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+    )
+    if args.report_to == "wandb":
+        if not is_wandb_available():
+            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
+        import wandb
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=False)
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+        if args.push_to_hub:
+            repo_id = create_repo(
+                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
+            ).repo_id
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
+    )
+
+    OUT_DIMS = 1024 if 'stabilityai/stable-diffusion-2-1' in args.pretrained_model_name_or_path else 768
+
+    text_encoder = CustomCLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+    )
+    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision)
+
+    unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
+    unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
+        unet_path, subfolder="unet", revision=args.revision
+    )
+
+    dino = DINO()
+    seg = KMeansSegmentation(args.train_data_dir + '/pretrained_kmeans.pth',
+                             args.fg_idx,
+                             args.bg_code,
+                             args.num_parts,
+                             args.num_k_per_part)
+
+    simple_mapper = TokenMapper(args.num_parts,
+                                args.num_k_per_part,
+                                OUT_DIMS,
+                                args.projection_nlayers)
+    # initialize placeholder token
+    placeholder_token = "<part>"
+    initializer_token = None
+    placeholder_token_ids = add_tokens(tokenizer,
+                                       text_encoder,
+                                       placeholder_token,
+                                       args.num_parts,
+                                       initializer_token)
+
+    # freeze parameters of models to save more memory
+    unet.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    dino.requires_grad_(False)
+
+    # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
+    # as these weights are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move unet, vae and text_encoder to device and cast to weight_dtype
+    unet.to(accelerator.device, dtype=weight_dtype)
+    vae.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # now we will add new LoRA weights to the attention layers
+    # It's important to realize here how many attention weights will be added and of which sizes
+    # The sizes of the attention layers consist only of two different variables:
+    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
+    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
+
+    # Let's first see how many attention processors we will have to set.
+    # For Stable Diffusion, it should be equal to:
+    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
+    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
+    # - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18
+    # => 32 layers
+
+    # Set correct lora layers
+    setup_attn_processor(unet, rank=args.rank)
+
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = version.parse(xformers.__version__)
+            if xformers_version == version.parse("0.0.16"):
+                logger.warn(
+                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                )
+            unet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    lora_layers = AttnProcsLayers(get_attn_processors(unet))
+
+    # Enable TF32 for faster training on Ampere GPUs,
+    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
+    if args.allow_tf32:
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if args.scale_lr:
+        args.learning_rate = (
+                args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Initialize the optimizer
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError(
+                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
+            )
+
+        optimizer_cls = bnb.optim.AdamW8bit
+    else:
+        optimizer_cls = torch.optim.AdamW
+
+    extra_params = list(simple_mapper.parameters())
+    mapper_lr = args.learning_rate * args.mapper_lr_scale if args.learning_rate != 0 else args.mapper_lr
+
+    optimizer = optimizer_cls(
+        [{'params': lora_layers.parameters()},
+         {'params': extra_params, 'lr': mapper_lr}],
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    train_dataset = DreamCreatureDataset(args.train_data_dir,
+                                         args.filename,
+                                         code_filename=args.code_filename,
+                                         num_parts=args.num_parts,
+                                         num_k_per_part=args.num_k_per_part,
+                                         use_gt_label=args.use_gt_label,
+                                         bg_code=args.bg_code,
+                                         repeat=args.repeat)
+
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            train_dataset.set_max_samples(args.max_train_samples, args.seed)
+
+    # DataLoaders creation:
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn(args, tokenizer, placeholder_token),
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
+    )
+
+    # Prepare everything with our `accelerator`.
+    lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        lora_layers, optimizer, train_dataloader, lr_scheduler
+    )
+    simple_mapper = accelerator.prepare(simple_mapper)
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        accelerator.init_trackers("text2image-fine-tune", config=vars(args))
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+    if args.resume_from_checkpoint:
+        if args.resume_from_checkpoint != "latest":
+            path = os.path.basename(args.resume_from_checkpoint)
+        else:
+            # Get the most recent checkpoint
+            dirs = os.listdir(args.output_dir)
+            dirs = [d for d in dirs if d.startswith("checkpoint")]
+            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+            path = dirs[-1] if len(dirs) > 0 else None
+
+        if path is None:
+            accelerator.print(
+                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
+            )
+            args.resume_from_checkpoint = None
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            global_step = int(path.split("-")[1])
+
+            resume_global_step = global_step * args.gradient_accumulation_steps
+            first_epoch = global_step // num_update_steps_per_epoch
+            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
+
+    # Only show the progress bar once on each machine.
+    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process,
+                        bar_format="{l_bar}{bar:10}{r_bar}{bar:-10b}")
+    progress_bar.set_description("Steps")
+
+    print(simple_mapper)
+
+    for epoch in range(first_epoch, args.num_train_epochs):
+        unet.train()
+        train_loss = 0.0
+        train_attn_loss = 0.0
+        train_diff_loss = 0.0
+        for step, batch in enumerate(train_dataloader):
+            # Skip steps until we reach the resumed step
+            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
+                if step % args.gradient_accumulation_steps == 0:
+                    progress_bar.update(1)
+                continue
+
+            with accelerator.accumulate(unet, simple_mapper):
+                # Convert images to latent space
+                latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
+                latents = latents * vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+                if args.noise_offset:
+                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
+                    noise += args.noise_offset * torch.randn(
+                        (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
+                    )
+
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+
+                # Get the text embedding for conditioning
+                mapper_outputs = simple_mapper(batch['codes'])
+                # print(mapper_outputs.size(), batch["input_ids"].size())
+                modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(batch["input_ids"],
+                                                                                     None,
+                                                                                     mapper_outputs,
+                                                                                     placeholder_token_ids)
+                # print(modified_hs.size())
+                encoder_hidden_states = text_encoder(batch["input_ids"], hidden_states=modified_hs)[0]
+
+                # Get the target for loss depending on the prediction type
+                if args.prediction_type is not None:
+                    # set prediction_type of scheduler if defined
+                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
+
+                if noise_scheduler.config.prediction_type == "epsilon":
+                    target = noise
+                elif noise_scheduler.config.prediction_type == "v_prediction":
+                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
+                else:
+                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+                # Predict the noise residual and compute loss
+                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
+
+                if args.snr_gamma is None:
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
+                    attn_loss, max_attn = dreamcreature_loss(batch,
+                                                             unet,
+                                                             dino,
+                                                             seg,
+                                                             placeholder_token_ids,
+                                                             accelerator)
+                    if args.masked_training:
+                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
+                        loss_image_mask = F.interpolate(masks.float(),
+                                                        size=target.shape[-2:],
+                                                        mode='bilinear') * torch.ones_like(target)
+                        loss = loss * loss_image_mask
+                        loss = loss.sum() / loss_image_mask.sum()
+                    else:
+                        loss = loss.mean()
+                else:
+                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
+                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
+                    # This is discussed in Section 4.2 of the same paper.
+                    snr = compute_snr(noise_scheduler, timesteps)
+                    if noise_scheduler.config.prediction_type == "v_prediction":
+                        # Velocity objective requires that we add one to SNR values before we divide by them.
+                        snr = snr + 1
+                    mse_loss_weights = (
+                            torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
+                    )
+
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
+                    attn_loss, max_attn = dreamcreature_loss(batch,
+                                                             unet,
+                                                             dino,
+                                                             seg,
+                                                             placeholder_token_ids,
+                                                             accelerator)
+                    if args.masked_training:
+                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
+                        loss_image_mask = F.interpolate(masks.float(),
+                                                        size=target.shape[-2:],
+                                                        mode='bilinear') * torch.ones_like(target)
+                        loss = loss * loss_image_mask
+                        loss = loss.sum(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
+                        loss = loss.sum() / loss_image_mask.sum()
+                    else:
+                        loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
+                        loss = loss.mean()
+
+                diff_loss = loss.clone().detach()
+                avg_diff_loss = accelerator.gather(diff_loss.repeat(args.train_batch_size)).mean()
+                train_diff_loss += avg_diff_loss.item() / args.gradient_accumulation_steps
+
+                avg_attn_loss = accelerator.gather(attn_loss.repeat(args.train_batch_size)).mean()
+                train_attn_loss += avg_attn_loss.item() / args.gradient_accumulation_steps
+
+                loss += args.attn_loss * attn_loss
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
+                train_loss += avg_loss.item() / args.gradient_accumulation_steps
+
+                # Backpropagate
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    params_to_clip = list(lora_layers.parameters()) + list(simple_mapper.parameters())
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+                accelerator.log({"train_loss": train_loss,
+                                 "diff_loss": train_diff_loss,
+                                 "attn_loss": train_attn_loss,
+                                 "mapper_norm": mapper_outputs.detach().norm().item(),
+                                 "max_attn": max_attn.item()
+                                 }, step=global_step)
+                train_loss = 0.0
+                train_attn_loss = 0.0
+                train_diff_loss = 0.0
+
+                if global_step % args.checkpointing_steps == 0:
+                    if accelerator.is_main_process:
+                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
+                        if args.checkpoints_total_limit is not None:
+                            checkpoints = os.listdir(args.output_dir)
+                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
+                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
+
+                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
+                            if len(checkpoints) >= args.checkpoints_total_limit:
+                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
+                                removing_checkpoints = checkpoints[0:num_to_remove]
+
+                                logger.info(
+                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
+                                )
+                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
+
+                                for removing_checkpoint in removing_checkpoints:
+                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
+                                    shutil.rmtree(removing_checkpoint)
+
+                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                        accelerator.save_state(save_path)
+                        logger.info(f"Saved state to {save_path}")
+
+            logs = {"step_loss": diff_loss.detach().item(),
+                    "attn_loss": attn_loss.detach().item(),
+                    "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+        if accelerator.is_main_process:
+            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
+                logger.info(
+                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+                    f" {args.validation_prompt}."
+                )
+                pipeline = DreamCreatureSDPipeline.from_pretrained(
+                    args.pretrained_model_name_or_path,
+                    unet=accelerator.unwrap_model(unet),
+                    text_encoder=accelerator.unwrap_model(text_encoder),
+                    tokenizer=tokenizer,
+                    revision=args.revision,
+                    torch_dtype=weight_dtype,
+                )
+                pipeline.placeholder_token_ids = placeholder_token_ids
+                pipeline.simple_mapper = accelerator.unwrap_model(simple_mapper)
+                pipeline.replace_token = False
+
+                pipeline = pipeline.to(accelerator.device)
+                pipeline.set_progress_bar_config(disable=True)
+
+                # run inference
+                generator = torch.Generator(device=accelerator.device)
+                if args.seed is not None:
+                    generator = generator.manual_seed(args.seed)
+                images = []
+                for _ in range(args.num_validation_images):
+                    images.append(
+                        pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
+                    )
+
+                for tracker in accelerator.trackers:
+                    if tracker.name == "tensorboard":
+                        np_images = np.stack([np.asarray(img) for img in images])
+                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
+                    if tracker.name == "wandb":
+                        tracker.log(
+                            {
+                                "validation": [
+                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
+                                    for i, image in enumerate(images)
+                                ]
+                            }
+                        )
+
+                del pipeline
+                torch.cuda.empty_cache()
+
+    # Save the lora layers
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        # unet = unet.to(torch.float32)
+        # unet.save_attn_procs(args.output_dir, safe_serialization=not args.custom_diffusion)
+
+        torch.save(lora_layers.to(torch.float32).state_dict(), args.output_dir + '/lora_layers.pth')
+        torch.save(simple_mapper.to(torch.float32).state_dict(), args.output_dir + '/hash_mapper.pth')
+
+        if args.push_to_hub:
+            save_model_card(
+                repo_id,
+                images=images,
+                base_model=args.pretrained_model_name_or_path,
+                dataset_name=args.dataset_name,
+                repo_folder=args.output_dir,
+            )
+            upload_folder(
+                repo_id=repo_id,
+                folder_path=args.output_dir,
+                commit_message="End of training",
+                ignore_patterns=["step_*", "epoch_*"],
+            )
+
+        del unet
+
+    # Final inference
+    # Load previous pipeline
+    tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
+    )
+    text_encoder = CustomCLIPTextModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
+    )
+    unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
+    unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
+        unet_path, subfolder="unet", revision=args.revision
+    )
+    pipeline = DreamCreatureSDPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        unet=unet,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        revision=args.revision,
+        torch_dtype=weight_dtype,
+    )
+    placeholder_token = "<part>"
+    initializer_token = None
+    placeholder_token_ids = add_tokens(tokenizer,
+                                       text_encoder,
+                                       placeholder_token,
+                                       args.num_parts,
+                                       initializer_token)
+    pipeline.placeholder_token_ids = placeholder_token_ids
+    pipeline.simple_mapper = TokenMapper(args.num_parts,
+                                         args.num_k_per_part,
+                                         OUT_DIMS,
+                                         args.projection_nlayers)
+    pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + '/hash_mapper.pth', map_location='cpu'))
+    pipeline.simple_mapper.to(accelerator.device)
+
+    pipeline = pipeline.to(accelerator.device)
+
+    # load attention processors
+    # pipeline.unet.load_attn_procs(args.output_dir, use_safetensors=not args.custom_diffusion)
+    setup_attn_processor(pipeline.unet, rank=args.rank)
+    load_attn_processor(pipeline.unet, args.output_dir + '/lora_layers.pth')
+
+    # run inference
+    pipeline.replace_token = False
+    generator = torch.Generator(device=accelerator.device)
+    if args.seed is not None:
+        generator = generator.manual_seed(args.seed)
+    images = []
+    for _ in range(args.num_validation_images):
+        images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
+
+    if accelerator.is_main_process:
+        for tracker in accelerator.trackers:
+            if len(images) != 0:
+                if tracker.name == "tensorboard":
+                    np_images = np.stack([np.asarray(img) for img in images])
+                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
+                if tracker.name == "wandb":
+                    tracker.log(
+                        {
+                            "test": [
+                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
+                                for i, image in enumerate(images)
+                            ]
+                        }
+                    )
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    main()

train_dreamcreature_sdxl.py

@@ -0,0 +1,1539 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
+
+import argparse
+import copy
+import itertools
+import logging
+import math
+import os
+import random
+import shutil
+from pathlib import Path
+from typing import Dict
+
+import datasets
+import diffusers
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
+from diffusers import (
+    AutoencoderKL,
+    DDPMScheduler,
+    StableDiffusionXLPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.loaders import LoraLoaderMixin
+from diffusers.models.lora import LoRALinearLayer
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import compute_snr
+from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils.import_utils import is_xformers_available
+from huggingface_hub import create_repo, upload_folder
+from packaging import version
+from torchvision import transforms
+from torchvision.transforms.functional import crop
+from tqdm.auto import tqdm
+from transformers import PretrainedConfig
+
+from dreamcreature.attn_processor import AttnProcessorCustom
+from dreamcreature.dataset import DreamCreatureDataset
+from dreamcreature.dino import DINO
+from dreamcreature.kmeans_segmentation import KMeansSegmentation
+from dreamcreature.loss import dreamcreature_loss
+from dreamcreature.mapper import TokenMapper
+from dreamcreature.pipeline_xl import DreamCreatureSDXLPipeline
+from dreamcreature.text_encoder import CustomCLIPTextModel, CustomCLIPTextModelWithProjection
+from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
+from utils import add_tokens, tokenize_prompt, get_attn_processors
+
+IMAGENET_TEMPLATES = [
+    "a photo of a {}",
+    "a rendering of a {}",
+    "a cropped photo of the {}",
+    "the photo of a {}",
+    "a photo of a clean {}",
+    "a photo of a dirty {}",
+    "a dark photo of the {}",
+    "a photo of my {}",
+    "a photo of the cool {}",
+    "a close-up photo of a {}",
+    "a bright photo of the {}",
+    "a cropped photo of a {}",
+    "a photo of the {}",
+    "a good photo of the {}",
+    "a photo of one {}",
+    "a close-up photo of the {}",
+    "a rendition of the {}",
+    "a photo of the clean {}",
+    "a rendition of a {}",
+    "a photo of a nice {}",
+    "a good photo of a {}",
+    "a photo of the nice {}",
+    "a photo of the small {}",
+    "a photo of the weird {}",
+    "a photo of the large {}",
+    "a photo of a cool {}",
+    "a photo of a small {}",
+]
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.25.0.dev0")
+
+logger = get_logger(__name__)
+
+
+# TODO: This function should be removed once training scripts are rewritten in PEFT
+def text_encoder_lora_state_dict(text_encoder):
+    state_dict = {}
+
+    def text_encoder_attn_modules(text_encoder):
+        from transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+        attn_modules = []
+
+        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+                name = f"text_model.encoder.layers.{i}.self_attn"
+                mod = layer.self_attn
+                attn_modules.append((name, mod))
+
+        return attn_modules
+
+    for name, module in text_encoder_attn_modules(text_encoder):
+        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+    return state_dict
+
+
+def save_model_card(
+        repo_id: str,
+        images=None,
+        base_model=str,
+        dataset_name=str,
+        train_text_encoder=False,
+        repo_folder=None,
+        vae_path=None,
+):
+    img_str = ""
+    for i, image in enumerate(images):
+        image.save(os.path.join(repo_folder, f"image_{i}.png"))
+        img_str += f"![img_{i}](./image_{i}.png)\n"
+
+    yaml = f"""
+---
+license: creativeml-openrail-m
+base_model: {base_model}
+dataset: {dataset_name}
+tags:
+- stable-diffusion-xl
+- stable-diffusion-xl-diffusers
+- text-to-image
+- diffusers
+- lora
+inference: true
+---
+    """
+    model_card = f"""
+# LoRA text2image fine-tuning - {repo_id}
+
+These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
+{img_str}
+
+LoRA for the text encoder was enabled: {train_text_encoder}.
+
+Special VAE used for training: {vae_path}.
+"""
+    with open(os.path.join(repo_folder, "README.md"), "w") as f:
+        f.write(yaml + model_card)
+
+
+def import_model_class_from_model_name_or_path(
+        pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
+):
+    text_encoder_config = PretrainedConfig.from_pretrained(
+        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
+    )
+    model_class = text_encoder_config.architectures[0]
+
+    if model_class == "CLIPTextModel":
+        from transformers import CLIPTextModel
+
+        return CLIPTextModel
+    elif model_class == "CLIPTextModelWithProjection":
+        from transformers import CLIPTextModelWithProjection
+
+        return CLIPTextModelWithProjection
+    else:
+        raise ValueError(f"{model_class} is not supported.")
+
+
+def parse_args(input_args=None):
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--pretrained_vae_model_name_or_path",
+        type=str,
+        default=None,
+        help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help="Revision of pretrained model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--variant",
+        type=str,
+        default=None,
+        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--train_data_dir",
+        type=str,
+        default=None,
+        help=(
+            "A folder containing the training data. Folder contents must follow the structure described in"
+            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
+            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
+        ),
+    )
+    parser.add_argument(
+        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default="text",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--validation_prompt",
+        type=str,
+        default=None,
+        help="A prompt that is used during validation to verify that the model is learning.",
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images that should be generated during validation with `validation_prompt`.",
+    )
+    parser.add_argument(
+        "--validation_epochs",
+        type=int,
+        default=1,
+        help=(
+            "Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`."
+        ),
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sd-model-finetuned-lora",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=1024,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
+            " cropped. The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_text_encoder",
+        action="store_true",
+        help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=100)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=500,
+        help=(
+            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
+            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
+            " training using `--resume_from_checkpoint`."
+        ),
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help=(
+            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
+            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
+        ),
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--snr_gamma",
+        type=float,
+        default=None,
+        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
+             "More details here: https://arxiv.org/abs/2303.09556.",
+    )
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=(
+            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
+        ),
+    )
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
+    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--prediction_type",
+        type=str,
+        default=None,
+        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
+    )
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=4,
+        help=("The dimension of the LoRA update matrices."),
+    )
+
+    parser.add_argument('--filename', default='train.txt')
+    parser.add_argument('--code_filename', default='train_caps_better_m8_k256.txt')
+    parser.add_argument('--repeat', default=1, type=int)
+
+    parser.add_argument('--scheduler_steps', default=1000, type=int, help='scheduler step, if turbo, set to 4')
+    parser.add_argument('--num_parts', type=int, default=4, help="Number of parts")
+    parser.add_argument('--num_k_per_part', type=int, default=256, help='Number of k')
+
+    parser.add_argument('--mapper_lr_scale', default=1, type=float)
+    parser.add_argument('--mapper_lr', default=0.0001, type=float)
+    parser.add_argument('--attn_loss', default=0, type=float)
+    parser.add_argument('--projection_nlayers', default=3, type=int)
+
+    parser.add_argument('--masked_training', action='store_true')
+    parser.add_argument('--drop_tokens', action='store_true')
+    parser.add_argument('--drop_rate', type=float, default=0.5)
+    parser.add_argument('--drop_counts', default='half')
+
+    parser.add_argument('--class_name', default='')
+    parser.add_argument('--no_pe', action='store_true')
+    parser.add_argument('--vector_shuffle', action='store_true')
+
+    parser.add_argument('--use_gt_label', action='store_true')
+    parser.add_argument('--bg_code', default=7, type=int)  # for gt_label
+    parser.add_argument('--fg_idx', default=0, type=int)
+    parser.add_argument('--use_templates', action='store_true')
+
+    parser.add_argument('--filter_class', default=None, type=int, help='debugging purpose')
+
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Need either a dataset name or a training folder.")
+
+    return args
+
+
+def unet_attn_processors_state_dict(unet) -> Dict[str, torch.tensor]:
+    """
+    Returns:
+        a state dict containing just the attention processor parameters.
+    """
+    attn_processors = get_attn_processors(unet)
+
+    attn_processors_state_dict = {}
+
+    for attn_processor_key, attn_processor in attn_processors.items():
+        for parameter_key, parameter in attn_processor.state_dict().items():
+            attn_processors_state_dict[f"{attn_processor_key}.{parameter_key}"] = parameter
+
+    return attn_processors_state_dict
+
+
+def encode_prompt(text_encoders, text_input_ids_list, placeholder_token_ids, mapper_outputs):
+    prompt_embeds_list = []
+
+    for i, text_encoder in enumerate(text_encoders):
+        text_input_ids = text_input_ids_list[i]
+
+        modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids,
+                                                                             None,
+                                                                             mapper_outputs[i],
+                                                                             placeholder_token_ids)
+
+        prompt_embeds = text_encoder(text_input_ids,
+                                     hidden_states=modified_hs,
+                                     output_hidden_states=True)
+        # prompt_embeds = text_encoder(
+        #     text_input_ids.to(text_encoder.device),
+        #     output_hidden_states=True,
+        # )
+
+        # We are only ALWAYS interested in the pooled output of the final text encoder
+        pooled_prompt_embeds = prompt_embeds[0]
+        prompt_embeds = prompt_embeds.hidden_states[-2]
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
+        prompt_embeds_list.append(prompt_embeds)
+
+    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
+    return prompt_embeds, pooled_prompt_embeds
+
+
+def collate_fn(args, tokenizer_one, tokenizer_two, placeholder_token):
+    # Preprocessing the datasets.
+    train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR)
+    train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
+    train_flip = transforms.RandomHorizontalFlip(p=1.0)
+    train_transforms = transforms.Compose(
+        [
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    def f(examples):
+        # image aug
+        original_sizes = []
+        all_images = []
+        crop_top_lefts = []
+        captions = []
+        raw_images = []
+        appeared_tokens = []
+        codes = []
+        for i in range(len(examples)):
+            ##### original sdxl process #####
+            image = examples[i]['pixel_values'].convert('RGB')
+            original_sizes.append((image.height, image.width))
+            image = train_resize(image)
+            if args.center_crop:
+                y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
+                x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
+                image = train_crop(image)
+            else:
+                y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
+                image = crop(image, y1, x1, h, w)
+            if args.random_flip and random.random() < 0.5:
+                # flip
+                x1 = image.width - x1
+                image = train_flip(image)
+            crop_top_left = (y1, x1)
+            crop_top_lefts.append(crop_top_left)
+            raw_images.append(image)
+            image = train_transforms(image)
+            all_images.append(image)
+
+            ##### dreamcreature caption #####
+            if args.use_templates and random.random() <= 0.5:  # 50% using templates
+                if args.class_name != '':
+                    caption = random.choice(IMAGENET_TEMPLATES).format(f'{placeholder_token} {args.class_name}')
+                else:
+                    caption = random.choice(IMAGENET_TEMPLATES).format(placeholder_token)
+            else:
+                if args.class_name != '':
+                    caption = f'{placeholder_token} {args.class_name}'
+                else:
+                    caption = placeholder_token
+
+            tokens = tokenizer_one.token_map[placeholder_token][:args.num_parts]
+            tokens = [tokens[a] for a in examples[i]['appeared']]
+
+            if args.vector_shuffle or args.drop_tokens:
+                tokens = copy.copy(tokens)
+                random.shuffle(tokens)
+
+            if args.drop_tokens and random.random() < args.drop_rate and len(tokens) >= 2:
+                # randomly drop half of the tokens
+                if args.drop_counts == 'half':
+                    tokens = tokens[:len(tokens) // 2]
+                else:
+                    tokens = tokens[:int(args.drop_counts)]
+
+            caption = caption.replace(placeholder_token, ' '.join(tokens))
+            captions.append(caption)
+
+            appeared = [int(t.split('_')[1]) for t in tokens]  # <part>_i
+            # examples[i]['appeared'] = appeared
+
+            appeared_tokens.append(appeared)
+
+            code = examples[i]['codes']
+            codes.append(code)
+
+        tokens_one = tokenize_prompt(tokenizer_one, captions)
+        tokens_two = tokenize_prompt(tokenizer_two, captions)
+
+        ##### start stacking #####
+        pixel_values = torch.stack([image for image in all_images])
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+        original_sizes = [s for s in original_sizes]
+        crop_top_lefts = [c for c in crop_top_lefts]
+        input_ids_one = torch.stack([t for t in tokens_one])
+        input_ids_two = torch.stack([t for t in tokens_two])
+
+        codes = torch.stack(codes, dim=0)
+
+        collate_output = {
+            "original_sizes": original_sizes,
+            "crop_top_lefts": crop_top_lefts,
+            "pixel_values": pixel_values,
+            "input_ids_one": input_ids_one,
+            "input_ids_two": input_ids_two,
+            "raw_images": raw_images,
+            "appeared_tokens": appeared_tokens,
+            "codes": codes
+        }
+
+        return collate_output
+
+    return f
+
+
+def setup_attn_processors(unet, args):
+    attn_size = args.resolution // 32
+    attn_procs = {}
+    for name in unet.attn_processors.keys():
+        attn_procs[name] = AttnProcessorCustom(attn_size)
+    unet.set_attn_processor(attn_procs)
+
+
+def init_for_pipeline(args):
+    tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer",
+        revision=args.revision,
+        use_fast=False,
+    )
+    tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer_2",
+        revision=args.revision,
+        use_fast=False,
+    )
+    text_encoder_cls_one = CustomCLIPTextModel
+    text_encoder_cls_two = CustomCLIPTextModelWithProjection
+    text_encoder_one = text_encoder_cls_one.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    text_encoder_two = text_encoder_cls_two.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
+    )
+
+    OUT_DIMS = 768 + 1280  # 2048
+    simple_mapper = TokenMapper(args.num_parts,
+                                args.num_k_per_part,
+                                OUT_DIMS,
+                                args.projection_nlayers)
+    return text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper
+
+
+def main(args):
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+        kwargs_handlers=[kwargs],
+    )
+
+    if args.report_to == "wandb":
+        if not is_wandb_available():
+            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
+        import wandb
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=False)
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+        if args.push_to_hub:
+            repo_id = create_repo(
+                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
+            ).repo_id
+
+    # Load the tokenizers (replace AutoTokenizer with the custom MultiTokenCLIPTokenizer)
+    tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer",
+        revision=args.revision,
+        use_fast=False,
+    )
+    tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer_2",
+        revision=args.revision,
+        use_fast=False,
+    )
+    # import correct text encoder classes
+    # text_encoder_cls_one = import_model_class_from_model_name_or_path(
+    #     args.pretrained_model_name_or_path, args.revision
+    # )
+    # text_encoder_cls_two = import_model_class_from_model_name_or_path(
+    #     args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
+    # )
+    text_encoder_cls_one = CustomCLIPTextModel
+    text_encoder_cls_two = CustomCLIPTextModelWithProjection
+
+    # Load scheduler and models
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path,
+                                                    subfolder="scheduler",
+                                                    num_train_steps=args.scheduler_steps)
+    text_encoder_one = text_encoder_cls_one.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    text_encoder_two = text_encoder_cls_two.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
+    )
+    vae_path = (
+        args.pretrained_model_name_or_path
+        if args.pretrained_vae_model_name_or_path is None
+        else args.pretrained_vae_model_name_or_path
+    )
+    vae = AutoencoderKL.from_pretrained(
+        vae_path,
+        subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
+        revision=args.revision,
+        variant=args.variant,
+    )
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
+    )
+
+    ##### dreamcreature init #####
+    OUT_DIMS = 768 + 1280  # 2048
+
+    dino = DINO()
+    seg = KMeansSegmentation(args.train_data_dir + '/pretrained_kmeans.pth',
+                             args.fg_idx,
+                             args.bg_code,
+                             args.num_parts,
+                             args.num_k_per_part)
+
+    simple_mapper = TokenMapper(args.num_parts,
+                                args.num_k_per_part,
+                                OUT_DIMS,
+                                args.projection_nlayers)
+
+    # We only train the additional adapter LoRA layers
+    vae.requires_grad_(False)
+    text_encoder_one.requires_grad_(False)
+    text_encoder_two.requires_grad_(False)
+    unet.requires_grad_(False)
+    dino.requires_grad_(False)
+
+    ##### dreamcreature, add sub-concepts token ids ####
+    placeholder_token = "<part>"
+    initializer_token = None
+    placeholder_token_ids_one = add_tokens(tokenizer_one,
+                                           text_encoder_one,
+                                           placeholder_token,
+                                           args.num_parts,
+                                           initializer_token)
+    placeholder_token_ids_two = add_tokens(tokenizer_two,
+                                           text_encoder_two,
+                                           placeholder_token,
+                                           args.num_parts,
+                                           initializer_token)
+
+    # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
+    # as these weights are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    # Move unet, vae and text_encoder to device and cast to weight_dtype
+    # The VAE is in float32 to avoid NaN losses.
+    unet.to(accelerator.device, dtype=weight_dtype)
+    if args.pretrained_vae_model_name_or_path is None:
+        vae.to(accelerator.device, dtype=torch.float32)
+    else:
+        vae.to(accelerator.device, dtype=weight_dtype)
+    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
+    text_encoder_two.to(accelerator.device, dtype=weight_dtype)
+    simple_mapper.to(accelerator.device)
+
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            import xformers
+
+            xformers_version = version.parse(xformers.__version__)
+            if xformers_version == version.parse("0.0.16"):
+                logger.warn(
+                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
+                )
+            unet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    # now we will add new LoRA weights to the attention layers
+    # Set correct lora layers
+    unet_lora_parameters = []
+    for attn_processor_name, attn_processor in unet.attn_processors.items():
+        # Parse the attention module.
+        attn_module = unet
+        for n in attn_processor_name.split(".")[:-1]:
+            attn_module = getattr(attn_module, n)
+
+        # Set the `lora_layer` attribute of the attention-related matrices.
+        attn_module.to_q.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_k.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_v.set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
+            )
+        )
+        attn_module.to_out[0].set_lora_layer(
+            LoRALinearLayer(
+                in_features=attn_module.to_out[0].in_features,
+                out_features=attn_module.to_out[0].out_features,
+                rank=args.rank,
+            )
+        )
+
+        # Accumulate the LoRA params to optimize.
+        unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
+        unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
+
+    setup_attn_processors(unet, args)
+
+    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
+    # So, instead, we monkey-patch the forward calls of its attention-blocks.
+    if args.train_text_encoder:
+        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
+        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_one, dtype=torch.float32, rank=args.rank
+        )
+        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
+            text_encoder_two, dtype=torch.float32, rank=args.rank
+        )
+
+    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+    def save_model_hook(models, weights, output_dir):
+        if accelerator.is_main_process:
+            # there are only two options here. Either are just the unet attn processor layers
+            # or there are the unet and text encoder atten layers
+            unet_lora_layers_to_save = None
+            text_encoder_one_lora_layers_to_save = None
+            text_encoder_two_lora_layers_to_save = None
+            mapper_to_save = None
+
+            for model in models:
+                if isinstance(model, type(accelerator.unwrap_model(unet))):
+                    unet_lora_layers_to_save = unet_attn_processors_state_dict(model)
+                elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
+                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
+                    text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
+                elif isinstance(model, TokenMapper):
+                    mapper_to_save = model.state_dict()
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
+
+                # make sure to pop weight so that corresponding model is not saved again
+                weights.pop()
+
+            StableDiffusionXLPipeline.save_lora_weights(
+                output_dir,
+                unet_lora_layers=unet_lora_layers_to_save,
+                text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
+                text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
+            )
+            torch.save(mapper_to_save, output_dir + '/hash_mapper.pth')
+
+    def load_model_hook(models, input_dir):
+        unet_ = None
+        text_encoder_one_ = None
+        text_encoder_two_ = None
+        mapper_ = None
+
+        while len(models) > 0:
+            model = models.pop()
+
+            if isinstance(model, type(accelerator.unwrap_model(unet))):
+                unet_ = model
+            elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
+                text_encoder_one_ = model
+            elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
+                text_encoder_two_ = model
+            elif isinstance(model, TokenMapper):
+                mapper_ = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
+
+        lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
+        LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alphas=network_alphas, unet=unet_)
+
+        text_encoder_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder." in k}
+        LoraLoaderMixin.load_lora_into_text_encoder(
+            text_encoder_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_one_
+        )
+
+        text_encoder_2_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder_2." in k}
+        LoraLoaderMixin.load_lora_into_text_encoder(
+            text_encoder_2_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_two_
+        )
+        mapper_.load_state_dict(torch.load(input_dir + '/hash_mapper.pth'))
+
+    accelerator.register_save_state_pre_hook(save_model_hook)
+    accelerator.register_load_state_pre_hook(load_model_hook)
+
+    # Enable TF32 for faster training on Ampere GPUs,
+    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
+    if args.allow_tf32:
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if args.scale_lr:
+        args.learning_rate = (
+                args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
+    if args.use_8bit_adam:
+        try:
+            import bitsandbytes as bnb
+        except ImportError:
+            raise ImportError(
+                "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
+            )
+
+        optimizer_class = bnb.optim.AdamW8bit
+    else:
+        optimizer_class = torch.optim.AdamW
+
+    extra_params = list(simple_mapper.parameters())
+    mapper_lr = args.learning_rate * args.mapper_lr_scale if args.learning_rate != 0 else args.mapper_lr
+
+    # Optimizer creation
+    params_to_optimize = (
+        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
+        if args.train_text_encoder
+        else unet_lora_parameters
+    )
+    optimizer = optimizer_class(
+        [{'params': params_to_optimize},
+         {'params': extra_params, 'lr': mapper_lr}],
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    # create
+    train_dataset = DreamCreatureDataset(args.train_data_dir,
+                                         args.filename,
+                                         code_filename=args.code_filename,
+                                         num_parts=args.num_parts,
+                                         num_k_per_part=args.num_k_per_part,
+                                         repeat=args.repeat,
+                                         use_gt_label=args.use_gt_label,
+                                         bg_code=args.bg_code)
+
+    with accelerator.main_process_first():
+        if args.filter_class is not None:
+            train_dataset.filter_by_class(args.filter_class)
+            print('selected', len(train_dataset))
+        if args.max_train_samples is not None:
+            train_dataset.set_max_samples(args.max_train_samples, args.seed)
+
+    # DataLoaders creation:
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=True,
+        collate_fn=collate_fn(args, tokenizer_one, tokenizer_two, placeholder_token),
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+    )
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
+        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+    )
+
+    # Prepare everything with our `accelerator`.
+    if args.train_text_encoder:
+        unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
+        )
+    else:
+        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, optimizer, train_dataloader, lr_scheduler
+        )
+    simple_mapper = accelerator.prepare(simple_mapper)
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        accelerator.init_trackers("text2image-fine-tune", config=vars(args))
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+    if args.resume_from_checkpoint:
+        if args.resume_from_checkpoint != "latest":
+            path = os.path.basename(args.resume_from_checkpoint)
+        else:
+            # Get the most recent checkpoint
+            dirs = os.listdir(args.output_dir)
+            dirs = [d for d in dirs if d.startswith("checkpoint")]
+            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+            path = dirs[-1] if len(dirs) > 0 else None
+
+        if path is None:
+            accelerator.print(
+                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
+            )
+            args.resume_from_checkpoint = None
+            initial_global_step = 0
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            global_step = int(path.split("-")[1])
+
+            initial_global_step = global_step
+            first_epoch = global_step // num_update_steps_per_epoch
+
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
+
+    for epoch in range(first_epoch, args.num_train_epochs):
+        unet.train()
+        if args.train_text_encoder:
+            text_encoder_one.train()
+            text_encoder_two.train()
+        train_loss = 0.0
+        train_diff_loss = 0.0
+        train_attn_loss = 0.0
+        for step, batch in enumerate(train_dataloader):
+            with accelerator.accumulate(unet, simple_mapper):
+                # Convert images to latent space
+                if args.pretrained_vae_model_name_or_path is not None:
+                    pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
+                else:
+                    pixel_values = batch["pixel_values"]
+
+                model_input = vae.encode(pixel_values).latent_dist.sample()
+                model_input = model_input * vae.config.scaling_factor
+                if args.pretrained_vae_model_name_or_path is None:
+                    model_input = model_input.to(weight_dtype)
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(model_input)
+                if args.noise_offset:
+                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
+                    noise += args.noise_offset * torch.randn(
+                        (model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
+                    )
+
+                bsz = model_input.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(
+                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
+                )
+                timesteps = timesteps.long()
+
+                # Add noise to the model input according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
+
+                # time ids
+                def compute_time_ids(original_size, crops_coords_top_left):
+                    # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
+                    target_size = (args.resolution, args.resolution)
+                    add_time_ids = list(original_size + crops_coords_top_left + target_size)
+                    add_time_ids = torch.tensor([add_time_ids])
+                    add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
+                    return add_time_ids
+
+                add_time_ids = torch.cat(
+                    [compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])]
+                )
+
+                # Predict the noise residual
+                unet_added_conditions = {"time_ids": add_time_ids}
+                # prompt_embeds, pooled_prompt_embeds = encode_prompt(
+                #     text_encoders=[text_encoder_one, text_encoder_two],
+                #     tokenizers=None,
+                #     prompt=None,
+                #     text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
+                # )
+                mapper_outputs = simple_mapper(batch['codes'])
+                prompt_embeds, pooled_prompt_embeds = encode_prompt(
+                    text_encoders=[text_encoder_one, text_encoder_two],
+                    text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
+                    placeholder_token_ids=placeholder_token_ids_one,
+                    mapper_outputs=[mapper_outputs[..., :768], mapper_outputs[..., 768:]]
+                )
+
+                unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
+                model_pred = unet(
+                    noisy_model_input, timesteps, prompt_embeds, added_cond_kwargs=unet_added_conditions
+                ).sample
+
+                # Get the target for loss depending on the prediction type
+                if args.prediction_type is not None:
+                    # set prediction_type of scheduler if defined
+                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
+
+                if noise_scheduler.config.prediction_type == "epsilon":
+                    target = noise
+                elif noise_scheduler.config.prediction_type == "v_prediction":
+                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
+                else:
+                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+                if args.snr_gamma is None:
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
+                    attn_loss, max_attn = dreamcreature_loss(batch,
+                                                             unet,
+                                                             dino,
+                                                             seg,
+                                                             placeholder_token_ids_one,
+                                                             accelerator)
+                    if args.masked_training:
+                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
+                        loss_image_mask = F.interpolate(masks.float(),
+                                                        size=target.shape[-2:],
+                                                        mode='bilinear') * torch.ones_like(target)
+                        loss = loss * loss_image_mask
+                        loss = loss.sum() / loss_image_mask.sum()
+                    else:
+                        loss = loss.mean()
+                else:
+                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
+                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
+                    # This is discussed in Section 4.2 of the same paper.
+                    snr = compute_snr(noise_scheduler, timesteps)
+                    if noise_scheduler.config.prediction_type == "v_prediction":
+                        # Velocity objective requires that we add one to SNR values before we divide by them.
+                        snr = snr + 1
+                    mse_loss_weights = (
+                            torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
+                    )
+
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
+                    attn_loss, max_attn = dreamcreature_loss(batch,
+                                                             unet,
+                                                             dino,
+                                                             seg,
+                                                             placeholder_token_ids_one,
+                                                             accelerator)
+                    if args.masked_training:
+                        masks = batch['masks'].unsqueeze(1).to(accelerator.device)
+                        loss_image_mask = F.interpolate(masks.float(),
+                                                        size=target.shape[-2:],
+                                                        mode='bilinear') * torch.ones_like(target)
+                        loss = loss * loss_image_mask
+                        loss = loss.sum(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
+                        loss = loss.sum() / loss_image_mask.sum()
+                    else:
+                        loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
+                        loss = loss.mean()
+
+                diff_loss = loss.clone().detach()
+                avg_diff_loss = accelerator.gather(diff_loss.repeat(args.train_batch_size)).mean()
+                train_diff_loss += avg_diff_loss.item() / args.gradient_accumulation_steps
+
+                avg_attn_loss = accelerator.gather(attn_loss.repeat(args.train_batch_size)).mean()
+                train_attn_loss += avg_attn_loss.item() / args.gradient_accumulation_steps
+
+                loss += args.attn_loss * attn_loss
+
+                # Gather the losses across all processes for logging (if we use distributed training).
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
+                train_loss += avg_loss.item() / args.gradient_accumulation_steps
+
+                # Backpropagate
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    params_to_clip = (
+                        itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
+                        if args.train_text_encoder
+                        else unet_lora_parameters
+                    )
+                    params_to_clip = list(params_to_clip) + extra_params
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+                accelerator.log({"train_loss": train_loss,
+                                 "diff_loss": train_diff_loss,
+                                 "attn_loss": train_attn_loss,
+                                 "max_attn": max_attn.item()
+                                 }, step=global_step)
+                train_loss = 0.0
+                train_attn_loss = 0.0
+                train_diff_loss = 0.0
+
+                if accelerator.is_main_process:
+                    if global_step % args.checkpointing_steps == 0:
+                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
+                        if args.checkpoints_total_limit is not None:
+                            checkpoints = os.listdir(args.output_dir)
+                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
+                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
+
+                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
+                            if len(checkpoints) >= args.checkpoints_total_limit:
+                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
+                                removing_checkpoints = checkpoints[0:num_to_remove]
+
+                                logger.info(
+                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
+                                )
+                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
+
+                                for removing_checkpoint in removing_checkpoints:
+                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
+                                    shutil.rmtree(removing_checkpoint)
+
+                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                        accelerator.save_state(save_path)
+                        logger.info(f"Saved state to {save_path}")
+
+            logs = {"step_loss": diff_loss.detach().item(),
+                    "attn_loss": attn_loss.detach().item(),
+                    "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+        if accelerator.is_main_process:
+            # todo: change pipeline
+            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
+                logger.info(
+                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+                    f" {args.validation_prompt}."
+                )
+                # create pipeline
+                pipeline = DreamCreatureSDXLPipeline.from_pretrained(
+                    args.pretrained_model_name_or_path,
+                    vae=vae,
+                    tokenizer=tokenizer_one,
+                    tokenizer_2=tokenizer_two,
+                    text_encoder=accelerator.unwrap_model(text_encoder_one),
+                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
+                    unet=accelerator.unwrap_model(unet),
+                    revision=args.revision,
+                    variant=args.variant,
+                    torch_dtype=weight_dtype,
+                )
+                pipeline.placeholder_token_ids = placeholder_token_ids_one
+                pipeline.simple_mapper = accelerator.unwrap_model(simple_mapper)
+                pipeline.replace_token = False
+
+                pipeline = pipeline.to(accelerator.device)
+                pipeline.set_progress_bar_config(disable=True)
+
+                # run inference
+                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+                pipeline_args = {"prompt": args.validation_prompt}
+
+                num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
+                gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0
+
+                images = [
+                    pipeline(**pipeline_args, num_inference_steps=num_steps, guidance_scale=gs,
+                             generator=generator, height=args.resolution, width=args.resolution).images[0]
+                    for _ in range(args.num_validation_images)
+                ]
+
+                for tracker in accelerator.trackers:
+                    if tracker.name == "tensorboard":
+                        np_images = np.stack([np.asarray(img) for img in images])
+                        tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
+                    if tracker.name == "wandb":
+                        tracker.log(
+                            {
+                                "validation": [
+                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
+                                    for i, image in enumerate(images)
+                                ]
+                            }
+                        )
+
+                del pipeline
+                torch.cuda.empty_cache()
+
+    # Save the lora layers
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        unet = accelerator.unwrap_model(unet)
+        unet_lora_layers = unet_attn_processors_state_dict(unet)
+
+        if args.train_text_encoder:
+            text_encoder_one = accelerator.unwrap_model(text_encoder_one)
+            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one)
+            text_encoder_two = accelerator.unwrap_model(text_encoder_two)
+            text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two)
+        else:
+            text_encoder_lora_layers = None
+            text_encoder_2_lora_layers = None
+
+        StableDiffusionXLPipeline.save_lora_weights(
+            save_directory=args.output_dir,
+            unet_lora_layers=unet_lora_layers,
+            text_encoder_lora_layers=text_encoder_lora_layers,
+            text_encoder_2_lora_layers=text_encoder_2_lora_layers,
+        )
+        torch.save(simple_mapper.to(torch.float32).state_dict(), args.output_dir + '/hash_mapper.pth')
+
+        del unet
+        del text_encoder_one
+        del text_encoder_two
+        del text_encoder_lora_layers
+        del text_encoder_2_lora_layers
+        del simple_mapper
+        torch.cuda.empty_cache()
+
+        # Final inference
+        # Load previous pipeline
+        text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper = init_for_pipeline(args)
+        pipeline = DreamCreatureSDXLPipeline.from_pretrained(
+            args.pretrained_model_name_or_path,
+            vae=vae,
+            tokenizer=tokenizer_one,
+            tokenizer_2=tokenizer_two,
+            text_encoder=text_encoder_one,
+            text_encoder_2=text_encoder_two,
+            revision=args.revision,
+            variant=args.variant,
+            torch_dtype=weight_dtype,
+        )
+        pipeline.placeholder_token_ids = placeholder_token_ids_one
+        pipeline.replace_token = False
+        pipeline.simple_mapper = simple_mapper
+        pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + '/hash_mapper.pth', map_location='cpu'))
+        pipeline.simple_mapper.to(accelerator.device)
+        setup_attn_processors(pipeline.unet, args)
+
+        pipeline = pipeline.to(accelerator.device)
+
+        # load attention processors
+        pipeline.load_lora_weights(args.output_dir)
+
+        # run inference
+        images = []
+        if args.validation_prompt and args.num_validation_images > 0:
+            num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
+            gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0
+            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+            images = [
+                pipeline(args.validation_prompt, num_inference_steps=num_steps,
+                         guidance_scale=gs, generator=generator, height=args.resolution,
+                         width=args.resolution).images[0]
+                for _ in range(args.num_validation_images)
+            ]
+
+            for tracker in accelerator.trackers:
+                if tracker.name == "tensorboard":
+                    np_images = np.stack([np.asarray(img) for img in images])
+                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
+                if tracker.name == "wandb":
+                    tracker.log(
+                        {
+                            "test": [
+                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
+                                for i, image in enumerate(images)
+                            ]
+                        }
+                    )
+
+        if args.push_to_hub:
+            save_model_card(
+                repo_id,
+                images=images,
+                base_model=args.pretrained_model_name_or_path,
+                dataset_name=args.dataset_name,
+                train_text_encoder=args.train_text_encoder,
+                repo_folder=args.output_dir,
+                vae_path=args.pretrained_vae_model_name_or_path,
+            )
+            upload_folder(
+                repo_id=repo_id,
+                folder_path=args.output_dir,
+                commit_message="End of training",
+                ignore_patterns=["step_*", "epoch_*"],
+            )
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

train_kmeans_segmentation.ipynb

@@ -0,0 +1,578 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "1c073d83-8e73-407a-a669-3a837a90aac6",
+   "metadata": {},
+   "source": [
+    "# Compute DINO features"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "901240c3-5111-4733-97ee-69891e4e7184",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import argparse\n",
+    "import math\n",
+    "import os\n",
+    "\n",
+    "import torch\n",
+    "import torchpq\n",
+    "from omegaconf import OmegaConf\n",
+    "from torch.utils.data import DataLoader\n",
+    "from sklearn.decomposition import PCA\n",
+    "from torchvision.transforms import transforms\n",
+    "from tqdm import tqdm\n",
+    "from transformers.utils import constants\n",
+    "\n",
+    "from dreamcreature.dino import DINO\n",
+    "from dreamcreature.dataset import ImageDataset\n",
+    "\n",
+    "MEAN = constants.IMAGENET_DEFAULT_MEAN\n",
+    "STD = constants.IMAGENET_DEFAULT_STD"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "2f83c248-c3d5-4fe2-a111-ecdeda648214",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "dataset_name = 'cub200_2011'\n",
+    "# dataset_name = 'dogs'\n",
+    "\n",
+    "rootdir = f'data/{dataset_name}'\n",
+    "resize = 256\n",
+    "crop = 224\n",
+    "\n",
+    "dataset = ImageDataset(rootdir,\n",
+    "                       'train.txt',\n",
+    "                       transform=transforms.Compose([\n",
+    "                           transforms.Resize(resize, interpolation=transforms.InterpolationMode.BICUBIC),\n",
+    "                           transforms.CenterCrop(crop),\n",
+    "                           transforms.ToTensor(),\n",
+    "                           transforms.Normalize(MEAN, STD)\n",
+    "                       ]))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "946a88e5-b368-4cc3-a625-fd952650036b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "dataloader = DataLoader(dataset, 32, shuffle=False, drop_last=False, num_workers=4)\n",
+    "model = DINO()\n",
+    "model.eval()\n",
+    "\n",
+    "device = torch.device('cuda')\n",
+    "model = model.to(device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f526a02a-745d-43df-94af-1a50ed438fda",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "os.makedirs(config.rootdir + '/dinov2', exist_ok=True)\n",
+    "\n",
+    "image_feats = []\n",
+    "with tqdm(dataloader, bar_format='{l_bar}{bar:10}{r_bar}') as tepoch:\n",
+    "    for i, (image, label, index) in enumerate(tepoch):\n",
+    "        image = image.to(device)\n",
+    "\n",
+    "        with torch.no_grad():\n",
+    "            output = model.get_feat_maps(image)  # (B, C, H, W)\n",
+    "\n",
+    "        B, C, H, W = output.size()\n",
+    "        output = output.reshape(B, C, H * W)\n",
+    "        image_feats.append(output.cpu())\n",
+    "\n",
+    "image_feats = torch.cat(image_feats, dim=0)  # (N, C, H*W)\n",
+    "torch.save(image_feats, rootdir + '/dinov2_image_feats.pth')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "50f9ed32-5231-47d0-864e-cfbcd8b6d732",
+   "metadata": {},
+   "source": [
+    "# Train Kmeans Segmentation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "da8b9bde-2c1a-40d5-a32c-dda98334fe17",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import random\n",
+    "import numpy as np\n",
+    "\n",
+    "dataset_name = 'cub200_2011'\n",
+    "# dataset_name = 'dogs'\n",
+    "\n",
+    "sd = torch.load(f'data/{dataset_name}/dinov2_image_feats.pth', map_location='cpu')\n",
+    "sd.size()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "67b0f5ba-e272-4aea-b17d-05a80e2ce025",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from dataset import code_to_int, int_to_caption\n",
+    "from dataset import ImageDataset\n",
+    "from torchvision.transforms import transforms\n",
+    "\n",
+    "ds = ImageDataset(f'data/{dataset_name}', transform=transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224)]))\n",
+    "train_lines = open(f'data/{dataset_name}/train.txt').readlines()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "46c57aa3-de67-4d1a-b9a7-584687e437f5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def set_seed(seed):\n",
+    "    random.seed(seed)\n",
+    "    np.random.seed(seed)\n",
+    "    torch.manual_seed(seed)\n",
+    "    torch.cuda.manual_seed(seed)\n",
+    "\n",
+    "set_seed(42)\n",
+    "    \n",
+    "n = 100  # use small training sample to avoid OOM\n",
+    "randidx = torch.randperm(len(sd))[:n]\n",
+    "randsd = sd[randidx].permute(0, 2, 1)  # (N, HW, C)\n",
+    "randsd.size()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fab9e741-ff9a-4b6a-8ee3-41bc1d43cb52",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import torchpq\n",
+    "import torch.nn.functional as F\n",
+    "import matplotlib.pyplot as plt\n",
+    "import random\n",
+    "from sklearn.decomposition import PCA\n",
+    "\n",
+    "set_seed(42)\n",
+    "\n",
+    "fg_kmeans = torchpq.clustering.KMeans(n_clusters=2,\n",
+    "                                      distance=\"cosine\",\n",
+    "                                      verbose=1,\n",
+    "                                      n_redo=5,\n",
+    "                                      max_iter=1000)\n",
+    "fg_labels = fg_kmeans.fit(randsd.reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(n, -1)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9a12550b-1c66-48a0-9bd6-a39b928cf57d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.unique(fg_labels, return_counts=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b9a3b9f0-fc73-4e44-bca8-a2a2da16df38",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for i in range(100):\n",
+    "    plt.subplot(10, 10, i+1)\n",
+    "    plt.imshow(fg_labels[i].reshape(16, 16))\n",
+    "    plt.axis('off')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6f8e6121-74ee-4539-bf61-9d0ba2198ef5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fg_idx = 0  # this have to do manual inspection, based on the visualization above\n",
+    "bg_idx = 1 - fg_idx\n",
+    "\n",
+    "randsd_bgnorm = []\n",
+    "randsd_nobg = []\n",
+    "randsd_bgmean = []\n",
+    "\n",
+    "for i in range(n):\n",
+    "    bgnorm_mean = randsd[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)\n",
+    "    \n",
+    "    if fg_idx == 0:\n",
+    "        bg_mask = fg_labels[i]\n",
+    "    else:\n",
+    "        bg_mask = 1 - fg_labels[i]\n",
+    "        \n",
+    "    bg_mask = bg_mask.unsqueeze(1)\n",
+    "    bgnorm = (randsd[i] * (1 - bg_mask)) + (bgnorm_mean * bg_mask)\n",
+    "    \n",
+    "    randsd_bgnorm.append(bgnorm)\n",
+    "    randsd_nobg.append(randsd[i] * (1 - bg_mask) + (-1 * bg_mask))\n",
+    "    randsd_bgmean.append(bgnorm_mean)\n",
+    "    \n",
+    "randsd_bgnorm = torch.stack(randsd_bgnorm, dim=0)\n",
+    "randsd_nobg = torch.stack(randsd_nobg, dim=0)\n",
+    "randsd_bgmean = torch.cat(randsd_bgmean, dim=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f90785f8-d1e4-4dc9-9e6d-c242058639a5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "set_seed(42)\n",
+    "M = 8\n",
+    "\n",
+    "coarse_kmeans = torchpq.clustering.KMeans(n_clusters=M,\n",
+    "                                          distance=\"cosine\",\n",
+    "                                          verbose=1,\n",
+    "                                          n_redo=5,\n",
+    "                                          max_iter=1000)\n",
+    "coarse_labels = coarse_kmeans.fit(randsd_nobg.reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(n, -1)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d4930216-de8a-420b-b7cd-45260299b7b9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for i in range(100):\n",
+    "    plt.subplot(10, 10, i+1)\n",
+    "    plt.imshow(coarse_labels[i].reshape(16, 16))\n",
+    "    plt.axis('off')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d5e72d5e-df24-485d-85aa-de17ba381d84",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import numpy as np\n",
+    "import matplotlib.pyplot as plt\n",
+    "\n",
+    "disp = coarse_labels[0].reshape(16, 16)\n",
+    "\n",
+    "plt.imshow(disp)\n",
+    "plt.axis('off')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b508e11d-082a-40a6-83db-4d306c0f9f00",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.unique(coarse_labels, return_counts=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8f4bf418-2297-4dc5-8bdd-15af7cf44c7a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sd_bgnorm = []\n",
+    "sd_nobg = []\n",
+    "sd_bgmean = []\n",
+    "\n",
+    "inp = sd.permute(0, 2, 1)\n",
+    "N = inp.size(0)\n",
+    "\n",
+    "sd_fg_labels = []\n",
+    "bs = 1000\n",
+    "for bidx in range(N // bs + 1):\n",
+    "    if bidx * bs >= N:\n",
+    "        break\n",
+    "        \n",
+    "    start_bidx = bidx*bs\n",
+    "    end_bidx = min((bidx+1)*bs, N)\n",
+    "    \n",
+    "    sd_fg_labels.append(fg_kmeans.predict(inp[start_bidx:end_bidx].reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(end_bidx - start_bidx, -1))\n",
+    "    \n",
+    "sd_fg_labels = torch.cat(sd_fg_labels, dim=0)\n",
+    "\n",
+    "for i in range(N):\n",
+    "    bgnorm_mean = inp[i][sd_fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)\n",
+    "    \n",
+    "    if fg_idx == 0:\n",
+    "        bg_mask = sd_fg_labels[i]\n",
+    "    else:\n",
+    "        bg_mask = 1 - sd_fg_labels[i]\n",
+    "        \n",
+    "    bg_mask = bg_mask.unsqueeze(1)\n",
+    "    bgnorm = (inp[i] * (1 - bg_mask)) + (bgnorm_mean * bg_mask)\n",
+    "    \n",
+    "    sd_bgnorm.append(bgnorm)\n",
+    "    sd_nobg.append(inp[i] * (1 - bg_mask) + (-1 * bg_mask))\n",
+    "    sd_bgmean.append(bgnorm_mean)\n",
+    "    print(i, end='\\r')\n",
+    "    \n",
+    "sd_bgnorm = torch.stack(sd_bgnorm, dim=0)\n",
+    "sd_nobg = torch.stack(sd_nobg, dim=0)\n",
+    "sd_bgmean = torch.cat(sd_bgmean, dim=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d8026046-2ac5-4d62-9d18-4cbf7e2ebdb0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sd_coarse_labels = []\n",
+    "bs = 1000\n",
+    "for bidx in range(N // bs + 1):\n",
+    "    if bidx * bs >= N:\n",
+    "        break\n",
+    "        \n",
+    "    start_bidx = bidx*bs\n",
+    "    end_bidx = min((bidx+1)*bs, N)\n",
+    "    \n",
+    "    sd_coarse_labels.append(coarse_kmeans.predict(sd_nobg[start_bidx:end_bidx].reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(end_bidx - start_bidx, -1))\n",
+    "    \n",
+    "sd_coarse_labels = torch.cat(sd_coarse_labels, dim=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "80a46c2d-6cd4-4344-b56f-4ff1fa7235d3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for i in range(100):\n",
+    "    plt.subplot(10, 10, i+1)\n",
+    "    coarse_mask = sd_coarse_labels[i].reshape(16, 16)\n",
+    "    plt.imshow(coarse_mask)\n",
+    "    plt.axis('off')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e41a997b-cccc-40c2-a983-c6092ffe69be",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.save(sd_coarse_labels.reshape(N, 16, 16).long().cpu(), f'data/{dataset_name}/coarse_mask_m8.pth')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "cc918d47-35ca-4177-b63b-a12f4f4a3d5a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.unique(sd_coarse_labels, return_counts=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9d4beb2d-e385-4ecc-9f90-287d7bc13c0d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from tqdm.auto import tqdm\n",
+    "\n",
+    "sd_fgmean = []\n",
+    "\n",
+    "inp = sd.permute(0, 2, 1)\n",
+    "N = inp.size(0)\n",
+    "M = 8\n",
+    "\n",
+    "for i in tqdm(range(N)):\n",
+    "    mean_feats = []\n",
+    "    for m in range(M):\n",
+    "        coarse_mask = sd_coarse_labels[i] == m\n",
+    "        if coarse_mask.sum().item() == 0:\n",
+    "            m_mean_feats = torch.zeros(1, 768)\n",
+    "        else:\n",
+    "            m_mean_feats = inp[i][coarse_mask].mean(dim=0, keepdim=True)\n",
+    "        \n",
+    "        mean_feats.append(m_mean_feats)\n",
+    "    \n",
+    "    mean_feats = torch.cat(mean_feats, dim=0)\n",
+    "    sd_fgmean.append(mean_feats)\n",
+    "    print(i, end='\\r')\n",
+    "    \n",
+    "sd_fgmean = torch.stack(sd_fgmean, dim=0)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5034a91a-ae6d-468b-a738-f9ec0d019d72",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "N = inp.size(0)\n",
+    "M = 8\n",
+    "K = 256\n",
+    "bgm = {'cub200_2011': 7, 'dogs': 1}[dataset_name]  # 7 for cub, 1 for dog, this means which index is background\n",
+    "\n",
+    "final_labels = torch.ones(N, M) * K\n",
+    "\n",
+    "set_seed(42)\n",
+    "\n",
+    "zero_mean_idxs = []\n",
+    "fine_feats = []\n",
+    "fine_kmeans_trained = []\n",
+    "\n",
+    "for m in range(M):\n",
+    "    fine_kmeans = torchpq.clustering.KMeans(n_clusters=K,\n",
+    "                          distance=\"cosine\",\n",
+    "                          verbose=1,\n",
+    "                          n_redo=5,\n",
+    "                          max_iter=1000)\n",
+    "    \n",
+    "    if m == bgm:\n",
+    "        fine_labels = fine_kmeans.fit(sd_bgmean.t().contiguous().cuda()).cpu()\n",
+    "        final_labels[:, m] = fine_labels\n",
+    "    else:\n",
+    "        fine_inp = sd_fgmean[:, m].reshape(-1, 768)\n",
+    "        fine_labels = fine_kmeans.fit(fine_inp.t().contiguous().cuda()).cpu()\n",
+    "        \n",
+    "        final_labels[:, m] = fine_labels\n",
+    "    \n",
+    "    fine_kmeans_trained.append(fine_kmeans)\n",
+    "    \n",
+    "    fine_feats.append(fine_kmeans.centroizds.cpu().t()[fine_labels])\n",
+    "    \n",
+    "    print('zero mean', torch.arange(K)[fine_kmeans.centroids.t().sum(dim=-1).cpu() == 0].tolist())\n",
+    "    zero_mean_idxs.append(torch.arange(K)[fine_kmeans.centroids.t().sum(dim=-1).cpu() == 0].tolist())\n",
+    "    \n",
+    "fine_feats = torch.cat(fine_feats, dim=1)\n",
+    "print(fine_feats.size())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "450bc5c1-cc41-44b3-ab6a-c0c65eb1210a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "torch.save({\n",
+    "    'foreground_background': fg_kmeans,\n",
+    "    'coarse_kmeans': coarse_kmeans,\n",
+    "    'fine_kmeans': fine_kmeans_trained,\n",
+    "}, f'data/{dataset_name}/pretrained_kmeans.pth')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9ad2f095-f99d-489e-b645-8933b6f66372",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from tqdm.auto import tqdm\n",
+    "\n",
+    "final_code_captions = []\n",
+    "counts = [[0 for _ in range(K)] for _ in range(M)]\n",
+    "\n",
+    "for i in tqdm(range(N)):\n",
+    "    m_labels = final_labels[i]  # M\n",
+    "    \n",
+    "    line = []\n",
+    "    for m in range(M):\n",
+    "        k = m_labels[m].long().item()\n",
+    "        \n",
+    "        if k not in zero_mean_idxs[m]:\n",
+    "            line.append(f'{m}:{k}')\n",
+    "            counts[m][k] += 1\n",
+    "        \n",
+    "    assert len(line) != 0, f'error at {i}'\n",
+    "    final_code_captions.append(' '.join(line))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1be7233d-d9fe-4e0a-bda9-93fc45f8e982",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import matplotlib.pyplot as plt\n",
+    "\n",
+    "for m in range(M):\n",
+    "    if max(counts[m]) == 0:\n",
+    "        continue\n",
+    "        \n",
+    "    plt.scatter(range(K), counts[m])\n",
+    "    print(m, min(counts[m]), max(counts[m]), np.mean(counts[m]))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e344ab48-fba6-4c17-a88c-e2528c0ac5cf",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "with open(f'data/{dataset_name}/train_caps_better_m{M}_k{K}.txt', 'w+') as f:\n",
+    "    for line in final_code_captions:\n",
+    "        f.write(line + '\\n')"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

utils.py

@@ -0,0 +1,127 @@
+import torch
+from diffusers.models.attention_processor import LoRAAttnProcessor
+
+
+def add_tokens(tokenizer, text_encoder, placeholder_token, num_vec_per_token=1, initializer_token=None):
+    """
+    Add tokens to the tokenizer and set the initial value of token embeddings
+    """
+    tokenizer.add_placeholder_tokens(placeholder_token, num_vec_per_token=num_vec_per_token)
+    text_encoder.resize_token_embeddings(len(tokenizer))
+    token_embeds = text_encoder.get_input_embeddings().weight.data
+    placeholder_token_ids = tokenizer.encode(placeholder_token, add_special_tokens=False)
+    if initializer_token:
+        token_ids = tokenizer.encode(initializer_token, add_special_tokens=False)
+        for i, placeholder_token_id in enumerate(placeholder_token_ids):
+            token_embeds[placeholder_token_id] = token_embeds[token_ids[i * len(token_ids) // num_vec_per_token]]
+    else:
+        for i, placeholder_token_id in enumerate(placeholder_token_ids):
+            token_embeds[placeholder_token_id] = torch.randn_like(token_embeds[placeholder_token_id])
+    return placeholder_token_ids
+
+
+def tokenize_prompt(tokenizer, prompt, replace_token=False):
+    text_inputs = tokenizer(
+        prompt,
+        replace_token=replace_token,
+        padding="max_length",
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    text_input_ids = text_inputs.input_ids
+    return text_input_ids
+
+
+def get_processor(self, return_deprecated_lora: bool = False):
+    r"""
+    Get the attention processor in use.
+
+    Args:
+        return_deprecated_lora (`bool`, *optional*, defaults to `False`):
+            Set to `True` to return the deprecated LoRA attention processor.
+
+    Returns:
+        "AttentionProcessor": The attention processor in use.
+    """
+    if not return_deprecated_lora:
+        return self.processor
+
+    # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible
+    # serialization format for LoRA Attention Processors. It should be deleted once the integration
+    # with PEFT is completed.
+    is_lora_activated = {
+        name: module.lora_layer is not None
+        for name, module in self.named_modules()
+        if hasattr(module, "lora_layer")
+    }
+
+    # 1. if no layer has a LoRA activated we can return the processor as usual
+    if not any(is_lora_activated.values()):
+        return self.processor
+
+    # If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
+    is_lora_activated.pop("add_k_proj", None)
+    is_lora_activated.pop("add_v_proj", None)
+    # 2. else it is not posssible that only some layers have LoRA activated
+    if not all(is_lora_activated.values()):
+        raise ValueError(
+            f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
+        )
+
+    # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor
+    # non_lora_processor_cls_name = self.processor.__class__.__name__
+    # lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)
+
+    hidden_size = self.inner_dim
+
+    # now create a LoRA attention processor from the LoRA layers
+    kwargs = {
+        "cross_attention_dim": self.cross_attention_dim,
+        "rank": self.to_q.lora_layer.rank,
+        "network_alpha": self.to_q.lora_layer.network_alpha,
+        "q_rank": self.to_q.lora_layer.rank,
+        "q_hidden_size": self.to_q.lora_layer.out_features,
+        "k_rank": self.to_k.lora_layer.rank,
+        "k_hidden_size": self.to_k.lora_layer.out_features,
+        "v_rank": self.to_v.lora_layer.rank,
+        "v_hidden_size": self.to_v.lora_layer.out_features,
+        "out_rank": self.to_out[0].lora_layer.rank,
+        "out_hidden_size": self.to_out[0].lora_layer.out_features,
+    }
+
+    if hasattr(self.processor, "attention_op"):
+        kwargs["attention_op"] = self.processor.attention_op
+
+    lora_processor = LoRAAttnProcessor(hidden_size, **kwargs)
+    lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
+    lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
+    lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
+    lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
+
+    return lora_processor
+
+
+def get_attn_processors(self):
+    r"""
+    Returns:
+        `dict` of attention processors: A dictionary containing all attention processors used in the model with
+        indexed by its weight name.
+    """
+    # set recursively
+    processors = {}
+
+    def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors):
+        if hasattr(module, "get_processor"):
+            processors[f"{name}.processor"] = get_processor(module, return_deprecated_lora=True)
+
+        for sub_name, child in module.named_children():
+            fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+        return processors
+
+    for name, module in self.named_children():
+        fn_recursive_add_processors(name, module, processors)
+
+    return processors
+