add app

app.py

@@ -0,0 +1,201 @@
+import os
+from PIL import Image
+import json
+import random
+
+import cv2
+import einops
+import gradio as gr
+import numpy as np
+import torch
+
+from pytorch_lightning import seed_everything
+from annotator.util import resize_image, HWC3
+from torch.nn.functional import threshold, normalize, interpolate
+from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
+from transformers import SegformerFeatureExtractor, SegformerForSemanticSegmentation
+from einops import rearrange, repeat
+
+import argparse
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+parseargs = argparse.ArgumentParser()
+parseargs.add_argument('--pretrained_model', type=str, default='runwayml/stable-diffusion-v1-5')
+parseargs.add_argument('--controlnet', type=str, default='controlnet')
+parseargs.add_argument('--precision', type=str, default='fp32')
+args = parseargs.parse_args()
+pretrained_model = args.pretrained_model
+
+# Check for different hardware architectures
+if torch.cuda.is_available():
+    device = "cuda"
+    # Check for xformers
+    try:
+        import xformers
+
+        enable_xformers = True
+    except ImportError:
+        enable_xformers = False
+elif torch.backends.mps.is_available():
+    device = "mps"
+else:
+    device = "cpu"
+
+print(f"Using device: {device}")
+
+# Load models
+if args.precision == 'fp32':
+    torch_dtype = torch.float32
+elif args.precision == 'fp16':
+    torch_dtype = torch.float16
+elif args.precision == 'bf16':
+    torch_dtype = torch.bfloat16
+else:
+    raise ValueError(f"Invalid precision: {args.precision}")
+
+controlnet = ControlNetModel.from_pretrained(args.controlnet, torch_dtype=torch_dtype)
+pipe = StableDiffusionControlNetPipeline.from_pretrained(
+    args.pretrained_model, controlnet=controlnet, torch_dtype=torch_dtype
+)
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to(device)
+
+# Apply optimizations based on hardware
+if device == "cuda":
+    pipe = pipe.to(device)
+    if enable_xformers:
+        pipe.enable_xformers_memory_efficient_attention()
+        print("xformers optimization enabled")
+elif device == "mps":
+    pipe = pipe.to(device)
+    pipe.enable_attention_slicing()
+    print("Attention slicing enabled for Apple Silicon")
+else:
+    # CPU-specific optimizations
+    pipe = pipe.to(device)
+    # pipe.enable_sequential_cpu_offload()
+    # pipe.enable_attention_slicing()
+
+feature_extractor = SegformerFeatureExtractor.from_pretrained("matei-dorian/segformer-b5-finetuned-human-parsing")
+segmodel = SegformerForSemanticSegmentation.from_pretrained("matei-dorian/segformer-b5-finetuned-human-parsing")
+
+
+def LGB_TO_RGB(gray_image, rgb_image):
+    # gray_image [H, W, 3]
+    # rgb_image [H, W, 3]
+
+    print("gray_image shape: ", gray_image.shape)
+    print("rgb_image shape: ", rgb_image.shape)
+
+    gray_image = cv2.cvtColor(gray_image, cv2.COLOR_RGB2GRAY)
+    lab_image = cv2.cvtColor(rgb_image, cv2.COLOR_RGB2LAB)
+    lab_image[:, :, 0] = gray_image[:, :]
+
+    return cv2.cvtColor(lab_image, cv2.COLOR_LAB2RGB)
+
+
+@torch.inference_mode()
+def process(input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, ddim_steps, strength,
+            guidance_scale, seed, eta, threshold, save_memory=False):
+    with torch.no_grad():
+        img = resize_image(input_image, image_resolution)
+        H, W, C = img.shape
+        print("img shape: ", img.shape)
+        if C == 3:
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
+        control = torch.from_numpy(img).to(device).float()
+        control = control / 255.0
+        control = rearrange(control, 'h w c -> 1 c h w')
+        # control = repeat(control, 'b c h w -> b c h w', b=num_samples)
+        # control = rearrange(control, 'b h w c -> b c h w')
+
+        if a_prompt:
+            prompt = prompt + ', ' + a_prompt
+
+        if seed == -1:
+            seed = random.randint(0, 65535)
+        seed_everything(seed)
+
+        generator = torch.Generator(device=device).manual_seed(seed)
+        # Generate images
+        output = pipe(
+            num_images_per_prompt=num_samples,
+            prompt=prompt,
+            image=control.to(device),
+            negative_prompt=n_prompt,
+            num_inference_steps=ddim_steps,
+            guidance_scale=guidance_scale,
+            generator=generator,
+            eta=eta,
+            strength=strength,
+            output_type='np',
+
+        ).images
+
+        # output = einops.rearrange(output, 'b c h w -> b h w c')
+        output = (output * 127.5 + 127.5).clip(0, 255).astype(np.uint8)
+
+        results = [output[i] for i in range(num_samples)]
+        results = [LGB_TO_RGB(img, result) for result in results]
+
+        # results의 각 이미지를 mask로 변환
+        masks = []
+        for result in results:
+            inputs = feature_extractor(images=result, return_tensors="pt")
+            outputs = segmodel(**inputs)
+            logits = outputs.logits
+            logits = logits.squeeze(0)
+            thresholded = torch.zeros_like(logits)
+            thresholded[logits > threshold] = 1
+            mask = thresholded[1:, :, :].sum(dim=0)
+            mask = mask.unsqueeze(0).unsqueeze(0)
+            mask = interpolate(mask, size=(H, W), mode='bilinear')
+            mask = mask.detach().numpy()
+            mask = np.squeeze(mask)
+            mask = np.where(mask > threshold, 1, 0)
+            masks.append(mask)
+
+        # results의 각 이미지를 mask를 이용해 mask가 0인 부분은 img 즉 흑백 이미지로 변환.
+        # img를 channel이 3인 rgb 이미지로 변환
+        final = [img * (1 - mask[:, :, None]) + result * mask[:, :, None] for result, mask in zip(results, masks)]
+
+    # mask to 255 img
+
+    mask_img = [mask * 255 for mask in masks]
+    return [img] + results + mask_img + final
+
+
+block = gr.Blocks().queue()
+with block:
+    with gr.Row():
+        gr.Markdown("## Control Stable Diffusion with Gray Image")
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(sources=['upload'], type="numpy")
+            prompt = gr.Textbox(label="Prompt")
+            run_button = gr.Button(value="Run")
+            with gr.Accordion("Advanced options", open=False):
+                num_samples = gr.Slider(label="Images", minimum=1, maximum=1, value=1, step=1, visible=False)
+                # num_samples = 1
+                image_resolution = gr.Slider(label="Image Resolution", minimum=256, maximum=768, value=512, step=64)
+                strength = gr.Slider(label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01)
+                # guess_mode = gr.Checkbox(label='Guess Mode', value=False)
+                ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=20, value=20, step=1)
+                scale = gr.Slider(label="Guidance Scale", minimum=0.1, maximum=30.0, value=1.0, step=0.1)
+                threshold = gr.Slider(label="Segmentation Threshold", minimum=0.1, maximum=0.9, value=0.5, step=0.05)
+                seed = gr.Slider(label="Seed", minimum=-1, maximum=2147483647, value=-1, step=1)
+                eta = gr.Number(label="eta (DDIM)", value=0.0)
+                a_prompt = gr.Textbox(label="Added Prompt", value='best quality, extremely detailed')
+                n_prompt = gr.Textbox(label="Negative Prompt",
+                                      value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality')
+        with gr.Column():
+            # result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto')
+            result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery")
+    ips = [input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, ddim_steps, strength, scale, seed,
+           eta, threshold]
+    run_button.click(fn=process, inputs=ips, outputs=[result_gallery], concurrency_limit=4)
+
+block.queue(max_size=100)
+block.launch(share=True)

ciff_dataset.py

@@ -0,0 +1,214 @@
+import os
+import random
+from concurrent.futures import ProcessPoolExecutor
+from pathlib import Path
+import json
+from PIL import Image
+import numpy as np
+import argparse
+from tqdm import tqdm
+
+# 인자 파싱
+parser = argparse.ArgumentParser(description="Dataset creation for image colorization")
+parser.add_argument("--source_dir", type=str, required=True, help="Source directory")
+parser.add_argument(
+    "--target_dir", type=str, required=True, help="Target directory for the dataset"
+)
+parser.add_argument(
+    "--resolution", type=int, default=512, help="Resolution for the dataset"
+)
+args = parser.parse_args()
+
+# 경로 설정
+root_dir = Path("E:/datasets")
+target_dir = root_dir / args.target_dir
+source_dir = root_dir / args.source_dir
+target_images_dir = target_dir / "images"
+target_conditioning_dir = target_dir / "conditioning_images"
+metadata_file = target_dir / "metadata.jsonl"
+
+# 디렉토리 생성
+target_dir.mkdir(parents=True, exist_ok=True)
+target_images_dir.mkdir(exist_ok=True)
+target_conditioning_dir.mkdir(exist_ok=True)
+
+# 프롬프트 목록
+prompts = [
+    "a color image, realistic style, photo",
+    "a color image, high resolution, realistic, painting",
+    "a color image, high resolution, realistic, photo",
+    "very good quality, absurd, photo, color, 4k image",
+    "high resolution, color, photo, realistic",
+    "high resolution, color, photo, realistic, 4k image",
+    "a color image, high resolution, realistic, 4k image",
+    "color, high resolution, photo, realistic",
+    "512x512, color, photo, realistic",
+]
+
+
+def process_image(image_path):
+    try:
+        # 이미지 로드 및 크롭
+        with Image.open(image_path) as img:
+            # 이미지 크기 확인
+            width, height = img.size
+            size = min(width, height)
+            left = (width - size) // 2
+            top = (height - size) // 2
+            right = left + size
+            bottom = top + size
+
+            # 크롭 및 리사이즈
+            img_cropped = img.crop((left, top, right, bottom)).resize(
+                (args.resolution, args.resolution), Image.LANCZOS
+            )
+
+            # 그레이스케일 변환
+            img_gray = img_cropped.convert("L")
+
+            # 파일명 생성
+            filename = image_path.stem + ".jpg"
+
+            # 이미지 저장
+            img_cropped.save(target_images_dir / filename)
+            img_gray.save(target_conditioning_dir / filename)
+
+            # 메타데이터 생성
+            metadata = {
+                "image": str(filename),
+                "text": random.choice(prompts),
+                "conditioning_image": str(filename),
+            }
+
+            return metadata
+    except Exception as e:
+        print(f"Error processing {image_path}: {e}")
+        return None
+
+
+def generate_dataset_loader(target_dir):
+    # 대상 디렉토리의 이름을 가져옵니다
+    dir_name = target_dir.name
+
+    # 클래스 이름을 생성합니다 (예: ciff_dataset -> CiffDataset)
+    class_name = ''.join(word.capitalize() for word in dir_name.split('_'))
+
+    # 파일 이름을 생성합니다
+    file_name = f"{dir_name}.py"
+
+    # 파일 경로를 생성합니다
+    file_path = target_dir / file_name
+
+    # 데이터셋 로더 코드를 생성합니다
+    code = f'''
+import pandas as pd
+from pathlib import Path
+import datasets
+import os
+
+_VERSION = datasets.Version("0.0.2")
+
+_DESCRIPTION = "TODO"
+_HOMEPAGE = "TODO"
+_LICENSE = "TODO"
+_CITATION = "TODO"
+
+_FEATURES = datasets.Features(
+    {{
+        "image": datasets.Image(),
+        "conditioning_image": datasets.Image(),
+        "text": datasets.Value("string"),
+    }}
+)
+
+_DEFAULT_CONFIG = datasets.BuilderConfig(name="default", version=_VERSION)
+
+
+class {class_name}(datasets.GeneratorBasedBuilder):
+    BUILDER_CONFIGS = [_DEFAULT_CONFIG]
+    DEFAULT_CONFIG_NAME = "default"
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=_FEATURES,
+            supervised_keys=None,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager):
+        base_path = Path(dl_manager._base_path)
+        metadata_path = base_path / "metadata.jsonl"
+        images_dir = base_path / "images"
+        conditioning_images_dir = base_path / "conditioning_images"
+
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={{
+                    "metadata_path": metadata_path,
+                    "images_dir": images_dir,
+                    "conditioning_images_dir": conditioning_images_dir,
+                }},
+            ),
+        ]
+
+    def _generate_examples(self, metadata_path, images_dir, conditioning_images_dir):
+        metadata = pd.read_json(metadata_path, lines=True)
+
+        for idx, row in metadata.iterrows():
+            text = row["text"]
+
+            image_path = os.path.join(images_dir, row["image"])
+            image = open(image_path, "rb").read()
+
+            conditioning_image_path = os.path.join(conditioning_images_dir, row["conditioning_image"])
+            conditioning_image = open(conditioning_image_path, "rb").read()
+
+            yield idx, {{
+                "text": text,
+                "image": {{
+                    "path": image_path,
+                    "bytes": image,
+                }},
+                "conditioning_image": {{
+                    "path": conditioning_image_path,
+                    "bytes": conditioning_image,
+                }},
+            }}
+'''
+
+    # 파일을 생성하고 코드를 작성합니다
+    with open(file_path, 'w') as f:
+        f.write(code)
+
+    print(f"데이터셋 로더 파일이 생성되었습니다: {file_path}")
+
+
+def main():
+    # 이미지 파일 목록 가져오기
+    image_files = list(source_dir.glob("*"))
+
+    # 프로세스 수 설정 (CPU 코어 수 - 1)
+    num_workers = (3 * os.cpu_count()) // 4
+
+    # 멀티프로세싱 실행
+    with ProcessPoolExecutor(max_workers=num_workers) as executor:
+        results = list(tqdm(executor.map(process_image, image_files), total=len(image_files), desc="Processing images"))
+
+    # 메타데이터 저장
+    with open(metadata_file, "w") as f:
+        for metadata in results:
+            if metadata:
+                json.dump(metadata, f)
+                f.write("\n")
+
+    # 데이터셋 로더 파일 생성
+    generate_dataset_loader(target_dir)
+
+
+if __name__ == "__main__":
+    main()
+    print(f"Dataset creation completed. Output directory: {target_dir}")

train_controlnet.py

@@ -0,0 +1,1239 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 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
+
+import argparse
+import contextlib
+import gc
+import logging
+import math
+import os
+import random
+import shutil
+from functools import partial
+from pathlib import Path
+
+import accelerate
+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 datasets import load_dataset, Features, Value, Image
+from huggingface_hub import create_repo, upload_folder
+from packaging import version
+from PIL import Image
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import AutoTokenizer, PretrainedConfig
+
+import diffusers
+from diffusers import (
+    AutoencoderKL,
+    ControlNetModel,
+    DDPMScheduler,
+    StableDiffusionControlNetPipeline,
+    UNet2DConditionModel,
+    UniPCMultistepScheduler,
+)
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.torch_utils import is_compiled_module
+
+
+if is_wandb_available():
+    import wandb
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.30.0.dev0")
+
+logger = get_logger(__name__)
+
+
+def image_grid(imgs, rows, cols):
+    assert len(imgs) == rows * cols
+
+    w, h = imgs[0].size
+    grid = Image.new("RGB", size=(cols * w, rows * h))
+
+    for i, img in enumerate(imgs):
+        grid.paste(img, box=(i % cols * w, i // cols * h))
+    return grid
+
+
+def log_validation(
+    vae, text_encoder, tokenizer, unet, controlnet, args, accelerator, weight_dtype, step, is_final_validation=False
+):
+    logger.info("Running validation... ")
+
+    if not is_final_validation:
+        controlnet = accelerator.unwrap_model(controlnet)
+    else:
+        controlnet = ControlNetModel.from_pretrained(args.output_dir, torch_dtype=weight_dtype)
+
+    pipeline = StableDiffusionControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        vae=vae,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        unet=unet,
+        controlnet=controlnet,
+        safety_checker=None,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    if args.enable_xformers_memory_efficient_attention:
+        pipeline.enable_xformers_memory_efficient_attention()
+
+    if args.seed is None:
+        generator = None
+    else:
+        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
+
+    # if args.validation_image is folder, get all images in the folder
+    if len(args.validation_image) == 1 and os.path.isdir(args.validation_image[0]):
+        logger.info(f"Loading images from {args.validation_image[0]}")
+        dir_path = args.validation_image[0]
+        validation_images = [os.path.join(dir_path, f) for f in os.listdir(dir_path) if os.path.isfile(os.path.join(dir_path, f))]
+        logger.info(f"Found {len(validation_images)} images")
+    else:
+        validation_images = args.validation_image
+
+
+    if len(validation_images) == len(args.validation_prompt):
+        validation_prompts = args.validation_prompt
+    elif len(validation_images) == 1:
+        validation_images = validation_images * len(args.validation_prompt)
+        validation_prompts = args.validation_prompt
+    elif len(args.validation_prompt) == 1:
+        validation_prompts = args.validation_prompt * len(validation_images)
+    else:
+        raise ValueError(
+            "number of `args.validation_image` and `args.validation_prompt` should be checked in `parse_args`"
+        )
+
+    image_logs = []
+    inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast("cuda")
+
+    for validation_prompt, validation_image in zip(validation_prompts, validation_images):
+        validation_image = Image.open(validation_image).convert("RGB")
+        # Resize
+        validation_image = transforms.Resize(args.resolution)(validation_image)
+
+        images = []
+
+        for _ in range(args.num_validation_images):
+            with inference_ctx:
+                image = pipeline(
+                    validation_prompt, validation_image, num_inference_steps=20, generator=generator
+                ).images[0]
+
+            images.append(image)
+
+        image_logs.append(
+            {"validation_image": validation_image, "images": images, "validation_prompt": validation_prompt}
+        )
+
+    tracker_key = "test" if is_final_validation else "validation"
+    for tracker in accelerator.trackers:
+        if tracker.name == "tensorboard":
+            for log in image_logs:
+                images = log["images"]
+                validation_prompt = log["validation_prompt"]
+                validation_image = log["validation_image"]
+
+                formatted_images = []
+
+                formatted_images.append(np.asarray(validation_image))
+
+                for image in images:
+                    formatted_images.append(np.asarray(image))
+
+                formatted_images = np.stack(formatted_images)
+
+                tracker.writer.add_images(validation_prompt, formatted_images, step, dataformats="NHWC")
+        elif tracker.name == "wandb":
+            formatted_images = []
+
+            for log in image_logs:
+                images = log["images"]
+                validation_prompt = log["validation_prompt"]
+                validation_image = log["validation_image"]
+
+                formatted_images.append(wandb.Image(validation_image, caption="Controlnet conditioning"))
+
+                for image in images:
+                    image = wandb.Image(image, caption=validation_prompt)
+                    formatted_images.append(image)
+
+            tracker.log({tracker_key: formatted_images})
+        else:
+            logger.warning(f"image logging not implemented for {tracker.name}")
+
+        del pipeline
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        return image_logs
+
+
+def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
+    text_encoder_config = PretrainedConfig.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="text_encoder",
+        revision=revision,
+    )
+    model_class = text_encoder_config.architectures[0]
+
+    if model_class == "CLIPTextModel":
+        from transformers import CLIPTextModel
+
+        return CLIPTextModel
+    elif model_class == "RobertaSeriesModelWithTransformation":
+        from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation
+
+        return RobertaSeriesModelWithTransformation
+    else:
+        raise ValueError(f"{model_class} is not supported.")
+
+
+def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):
+    img_str = ""
+    if image_logs is not None:
+        img_str = "You can find some example images below.\n\n"
+        for i, log in enumerate(image_logs):
+            images = log["images"]
+            validation_prompt = log["validation_prompt"]
+            validation_image = log["validation_image"]
+            validation_image.save(os.path.join(repo_folder, "image_control.png"))
+            img_str += f"prompt: {validation_prompt}\n"
+            images = [validation_image] + images
+            image_grid(images, 1, len(images)).save(os.path.join(repo_folder, f"images_{i}.png"))
+            img_str += f"![images_{i})](./images_{i}.png)\n"
+
+    model_description = f"""
+# controlnet-{repo_id}
+
+These are controlnet weights trained on {base_model} with new type of conditioning.
+{img_str}
+"""
+    model_card = load_or_create_model_card(
+        repo_id_or_path=repo_id,
+        from_training=True,
+        license="creativeml-openrail-m",
+        base_model=base_model,
+        model_description=model_description,
+        inference=True,
+    )
+
+    tags = [
+        "stable-diffusion",
+        "stable-diffusion-diffusers",
+        "text-to-image",
+        "diffusers",
+        "controlnet",
+        "diffusers-training",
+    ]
+    model_card = populate_model_card(model_card, tags=tags)
+
+    model_card.save(os.path.join(repo_folder, "README.md"))
+
+
+def parse_args(input_args=None):
+    parser = argparse.ArgumentParser(description="Simple example of a ControlNet 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(
+        "--controlnet_model_name_or_path",
+        type=str,
+        default=None,
+        help="Path to pretrained controlnet model or model identifier from huggingface.co/models."
+        " If not specified controlnet weights are initialized from unet.",
+    )
+    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(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="controlnet-model",
+        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(
+        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=1)
+    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. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
+            "In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
+            "Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
+            "See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
+            "instructions."
+        ),
+    )
+    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=5e-6,
+        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(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
+    )
+    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument(
+        "--dataset_num_workers",
+        type=int,
+        default=1,
+        help=(
+            "Number of subprocesses to use for data loading."
+        ),
+    )
+    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(
+        "--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(
+        "--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(
+        "--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(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument(
+        "--set_grads_to_none",
+        action="store_true",
+        help=(
+            "Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
+            " behaviors, so disable this argument if it causes any problems. More info:"
+            " https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
+        ),
+    )
+    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 the target image."
+    )
+    parser.add_argument(
+        "--conditioning_image_column",
+        type=str,
+        default="conditioning_image",
+        help="The column of the dataset containing the controlnet conditioning 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(
+        "--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(
+        "--proportion_empty_prompts",
+        type=float,
+        default=0,
+        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
+    )
+    parser.add_argument(
+        "--validation_prompt",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of prompts evaluated every `--validation_steps` and logged to `--report_to`."
+            " Provide either a matching number of `--validation_image`s, a single `--validation_image`"
+            " to be used with all prompts, or a single prompt that will be used with all `--validation_image`s."
+        ),
+    )
+    parser.add_argument(
+        "--validation_image",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of paths to the controlnet conditioning image be evaluated every `--validation_steps`"
+            " and logged to `--report_to`. Provide either a matching number of `--validation_prompt`s, a"
+            " a single `--validation_prompt` to be used with all `--validation_image`s, or a single"
+            " `--validation_image` that will be used with all `--validation_prompt`s."
+        ),
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images to be generated for each `--validation_image`, `--validation_prompt` pair",
+    )
+    parser.add_argument(
+        "--validation_steps",
+        type=int,
+        default=100,
+        help=(
+            "Run validation every X steps. Validation consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`"
+            " and logging the images."
+        ),
+    )
+    parser.add_argument(
+        "--tracker_project_name",
+        type=str,
+        default="train_controlnet",
+        help=(
+            "The `project_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+    parser.add_argument(
+        "--tracker_run_name",
+        type=str,
+        default=None,
+        help=(
+            "The `run_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Specify either `--dataset_name` or `--train_data_dir`")
+
+    if args.dataset_name is not None and args.train_data_dir is not None:
+        raise ValueError("Specify only one of `--dataset_name` or `--train_data_dir`")
+
+    if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
+        raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
+
+    if args.validation_prompt is not None and args.validation_image is None:
+        raise ValueError("`--validation_image` must be set if `--validation_prompt` is set")
+
+    if args.validation_prompt is None and args.validation_image is not None:
+        raise ValueError("`--validation_prompt` must be set if `--validation_image` is set")
+
+    if (
+        args.validation_image is not None
+        and args.validation_prompt is not None
+        and len(args.validation_image) != 1
+        and len(args.validation_prompt) != 1
+        and len(args.validation_image) != len(args.validation_prompt)
+    ):
+        raise ValueError(
+            "Must provide either 1 `--validation_image`, 1 `--validation_prompt`,"
+            " or the same number of `--validation_prompt`s and `--validation_image`s"
+        )
+
+    if args.resolution % 8 != 0:
+        raise ValueError(
+            "`--resolution` must be divisible by 8 for consistently sized encoded images between the VAE and the controlnet encoder."
+        )
+
+    return args
+
+def preprocess_train(examples, image_column, conditioning_image_column, image_transforms, conditioning_image_transforms, tokenize_caption):
+    images = [image.convert("RGB") for image in examples[image_column]]
+    images = [image_transforms(image) for image in images]
+
+    conditioning_images = [image.convert("RGB") for image in examples[conditioning_image_column]]
+    conditioning_images = [conditioning_image_transforms(image) for image in conditioning_images]
+
+    examples["pixel_values"] = images
+    examples["conditioning_pixel_values"] = conditioning_images
+    examples["input_ids"] = tokenize_caption
+
+    return examples
+
+def make_train_dataset(args, tokenizer, accelerator):
+    # Get the datasets: you can either provide your own training and evaluation files (see below)
+    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
+
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        dataset = load_dataset(
+            args.dataset_name,
+            args.dataset_config_name,
+            cache_dir=args.cache_dir,
+        )
+    else:
+        # Get train_data_dir's last folder name
+        if args.train_data_dir is not None:
+            # For optimal performance
+            dataset = load_dataset(
+                args.train_data_dir,
+                cache_dir=args.cache_dir,
+                num_proc=args.dataset_num_workers,
+                # streaming=True,
+                trust_remote_code=True,
+            )
+
+        # See more about loading custom images at
+        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    column_names = dataset["train"].column_names
+
+    # 6. Get the column names for input/target.
+    if args.image_column is None:
+        image_column = column_names[0]
+        logger.info(f"image column defaulting to {image_column}")
+    else:
+        image_column = args.image_column
+        if image_column not in column_names:
+            raise ValueError(
+                f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.caption_column is None:
+        caption_column = column_names[1]
+        logger.info(f"caption column defaulting to {caption_column}")
+    else:
+        caption_column = args.caption_column
+        if caption_column not in column_names:
+            raise ValueError(
+                f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    if args.conditioning_image_column is None:
+        conditioning_image_column = column_names[2]
+        logger.info(f"conditioning image column defaulting to {conditioning_image_column}")
+    else:
+        conditioning_image_column = args.conditioning_image_column
+        if conditioning_image_column not in column_names:
+            raise ValueError(
+                f"`--conditioning_image_column` value '{args.conditioning_image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+            )
+
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if random.random() < args.proportion_empty_prompts:
+                captions.append("")
+            elif isinstance(caption, str):
+                captions.append(caption)
+            elif isinstance(caption, (list, np.ndarray)):
+                # take a random caption if there are multiple
+                captions.append(random.choice(caption) if is_train else caption[0])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        inputs = tokenizer(
+            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+    # Tokenize captions
+    # tokenize_caption = tokenizer(
+    #     [""], max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
+    # ).input_ids
+
+    image_transforms = transforms.Compose(
+        [
+            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+
+    conditioning_image_transforms = transforms.Compose(
+        [
+            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.CenterCrop(args.resolution),
+            transforms.ToTensor(),
+        ]
+    )
+
+    batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+    tokenize_caption = tokenizer(
+        [""] * batch_size, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
+    ).input_ids
+
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
+        # Set the training transforms
+        logger.info("Applying preprocessing to the training dataset...")
+        train_dataset = dataset["train"].with_transform(partial(preprocess_train, image_column=image_column, conditioning_image_column=conditioning_image_column, image_transforms=image_transforms, conditioning_image_transforms=conditioning_image_transforms, tokenize_caption=tokenize_caption))
+        logger.info("Preprocessing applied to the training dataset.")
+
+    return train_dataset
+
+
+def collate_fn(examples):
+    pixel_values = torch.stack([example["pixel_values"] for example in examples])
+    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    conditioning_pixel_values = torch.stack([example["conditioning_pixel_values"] for example in examples])
+    conditioning_pixel_values = conditioning_pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    input_ids = torch.stack([example["input_ids"] for example in examples])
+
+    return {
+        "pixel_values": pixel_values,
+        "conditioning_pixel_values": conditioning_pixel_values,
+        "input_ids": input_ids,
+    }
+
+
+def main(args):
+    if args.report_to == "wandb" and args.hub_token is not None:
+        raise ValueError(
+            "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
+            " Please use `huggingface-cli login` to authenticate with the Hub."
+        )
+    os.environ["WANDB__SERVICE_WAIT"] = "300"
+
+    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,
+    )
+
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
+    # 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)
+    logger.info(f"Training/evaluation parameters {args}")
+    if accelerator.is_local_main_process:
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        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 tokenizer
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+            use_fast=False,
+        )
+
+    # import correct text encoder class
+    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
+
+    # Load scheduler and models
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    text_encoder = text_encoder_cls.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    vae = AutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
+    )
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
+    )
+
+    if args.controlnet_model_name_or_path:
+        logger.info("Loading existing controlnet weights")
+        controlnet = ControlNetModel.from_pretrained(args.controlnet_model_name_or_path)
+    else:
+        logger.info("Initializing controlnet weights from unet")
+        controlnet = ControlNetModel.from_unet(unet)
+
+    # Taken from [Sayak Paul's Diffusers PR #6511](https://github.com/huggingface/diffusers/pull/6511/files)
+    def unwrap_model(model):
+        model = accelerator.unwrap_model(model)
+        model = model._orig_mod if is_compiled_module(model) else model
+        return model
+
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
+        # 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:
+                i = len(weights) - 1
+
+                while len(weights) > 0:
+                    weights.pop()
+                    model = models[i]
+
+                    sub_dir = "controlnet"
+                    model.save_pretrained(os.path.join(output_dir, sub_dir))
+
+                    i -= 1
+
+        def load_model_hook(models, input_dir):
+            while len(models) > 0:
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                # load diffusers style into model
+                load_model = ControlNetModel.from_pretrained(input_dir, subfolder="controlnet")
+                model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
+    vae.requires_grad_(False)
+    unet.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    controlnet.train()
+
+    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.warning(
+                    "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()
+            controlnet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    if args.gradient_checkpointing:
+        controlnet.enable_gradient_checkpointing()
+
+    # Check that all trainable models are in full precision
+    low_precision_error_string = (
+        " Please make sure to always have all model weights in full float32 precision when starting training - even if"
+        " doing mixed precision training, copy of the weights should still be float32."
+    )
+
+    if unwrap_model(controlnet).dtype != torch.float32:
+        raise ValueError(
+            f"Controlnet loaded as datatype {unwrap_model(controlnet).dtype}. {low_precision_error_string}"
+        )
+
+    # 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
+
+    # Optimizer creation
+    params_to_optimize = controlnet.parameters()
+    optimizer = optimizer_class(
+        params_to_optimize,
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    logger.info("Loading the training dataset")
+    train_dataset = make_train_dataset(args, tokenizer, accelerator)
+
+    logger.info("Creating the training dataloader")
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=False,
+        collate_fn=collate_fn,
+        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,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+    )
+
+    # Prepare everything with our `accelerator`.
+    controlnet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        controlnet, optimizer, train_dataloader, lr_scheduler
+    )
+
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models 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 vae, unet and text_encoder to device and cast to weight_dtype
+    vae.to(accelerator.device, dtype=weight_dtype)
+    unet.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # 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:
+        tracker_config = dict(vars(args))
+
+        # tensorboard cannot handle list types for config
+        tracker_config.pop("validation_prompt")
+        tracker_config.pop("validation_image")
+
+        logger.info(f"Init trackers: {args.tracker_project_name}")
+        accelerator.init_trackers(args.tracker_project_name, config=tracker_config)
+        # if args.tracker_run_name is not None:
+        #     accelerator.trackers[-1].run.name = args.tracker_run_name
+
+    # 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 batches each epoch = {len(train_dataloader)}")
+    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,
+    )
+
+    image_logs = None
+    for epoch in range(first_epoch, args.num_train_epochs):
+        for step, batch in enumerate(train_dataloader):
+            with accelerator.accumulate(controlnet):
+                # 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)
+                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
+                encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]
+
+                controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)
+
+                down_block_res_samples, mid_block_res_sample = controlnet(
+                    noisy_latents,
+                    timesteps,
+                    encoder_hidden_states=encoder_hidden_states,
+                    controlnet_cond=controlnet_image,
+                    return_dict=False,
+                )
+
+                # Predict the noise residual
+                model_pred = unet(
+                    noisy_latents,
+                    timesteps,
+                    encoder_hidden_states=encoder_hidden_states,
+                    down_block_additional_residuals=[
+                        sample.to(dtype=weight_dtype) for sample in down_block_res_samples
+                    ],
+                    mid_block_additional_residual=mid_block_res_sample.to(dtype=weight_dtype),
+                    return_dict=False,
+                )[0]
+
+                # Get the target for loss depending on the 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}")
+                loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    params_to_clip = controlnet.parameters()
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad(set_to_none=args.set_grads_to_none)
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+                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}")
+
+                    if args.validation_prompt is not None and global_step % args.validation_steps == 0:
+                        image_logs = log_validation(
+                            vae,
+                            text_encoder,
+                            tokenizer,
+                            unet,
+                            controlnet,
+                            args,
+                            accelerator,
+                            weight_dtype,
+                            global_step,
+                        )
+
+            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+
+            if global_step >= args.max_train_steps:
+                break
+
+    # Create the pipeline using using the trained modules and save it.
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        controlnet = unwrap_model(controlnet)
+        controlnet.save_pretrained(args.output_dir)
+
+        # Run a final round of validation.
+        image_logs = None
+        if args.validation_prompt is not None:
+            image_logs = log_validation(
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                unet=unet,
+                controlnet=None,
+                args=args,
+                accelerator=accelerator,
+                weight_dtype=weight_dtype,
+                step=global_step,
+                is_final_validation=True,
+            )
+
+        if args.push_to_hub:
+            save_model_card(
+                repo_id,
+                image_logs=image_logs,
+                base_model=args.pretrained_model_name_or_path,
+                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_*"],
+            )
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

train_multi_open.py

@@ -0,0 +1,1192 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 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
+
+import argparse
+import contextlib
+import gc
+import logging
+import math
+import os
+import random
+import shutil
+from pathlib import Path
+
+import accelerate
+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 datasets import load_dataset, Features, Value, Image, IterableDataset
+from datasets.iterable_dataset import ExamplesIterable
+from torch.utils.data import get_worker_info
+from multiprocessing import Pool, Queue, cpu_count
+from threading import Thread
+from itertools import cycle
+from huggingface_hub import create_repo, upload_folder
+from packaging import version
+from PIL import Image
+from torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import AutoTokenizer, PretrainedConfig
+
+import diffusers
+from diffusers import (
+    AutoencoderKL,
+    ControlNetModel,
+    DDPMScheduler,
+    StableDiffusionControlNetPipeline,
+    UNet2DConditionModel,
+    UniPCMultistepScheduler,
+)
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.torch_utils import is_compiled_module
+
+
+if is_wandb_available():
+    import wandb
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.30.0.dev0")
+
+logger = get_logger(__name__)
+
+
+def image_grid(imgs, rows, cols):
+    assert len(imgs) == rows * cols
+
+    w, h = imgs[0].size
+    grid = Image.new("RGB", size=(cols * w, rows * h))
+
+    for i, img in enumerate(imgs):
+        grid.paste(img, box=(i % cols * w, i // cols * h))
+    return grid
+
+
+def log_validation(
+    vae, text_encoder, tokenizer, unet, controlnet, args, accelerator, weight_dtype, step, is_final_validation=False
+):
+    logger.info("Running validation... ")
+
+    if not is_final_validation:
+        controlnet = accelerator.unwrap_model(controlnet)
+    else:
+        controlnet = ControlNetModel.from_pretrained(args.output_dir, torch_dtype=weight_dtype)
+
+    pipeline = StableDiffusionControlNetPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        vae=vae,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        unet=unet,
+        controlnet=controlnet,
+        safety_checker=None,
+        revision=args.revision,
+        variant=args.variant,
+        torch_dtype=weight_dtype,
+    )
+    pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config)
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    if args.enable_xformers_memory_efficient_attention:
+        pipeline.enable_xformers_memory_efficient_attention()
+
+    if args.seed is None:
+        generator = None
+    else:
+        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
+
+    # if args.validation_image is folder, get all images in the folder
+    if len(args.validation_image) == 1 and os.path.isdir(args.validation_image[0]):
+        logger.info(f"Loading images from {args.validation_image[0]}")
+        dir_path = args.validation_image[0]
+        validation_images = [os.path.join(dir_path, f) for f in os.listdir(dir_path) if os.path.isfile(os.path.join(dir_path, f))]
+        logger.info(f"Found {len(validation_images)} images")
+    else:
+        validation_images = args.validation_image
+
+
+    if len(validation_images) == len(args.validation_prompt):
+        validation_prompts = args.validation_prompt
+    elif len(validation_images) == 1:
+        validation_images = validation_images * len(args.validation_prompt)
+        validation_prompts = args.validation_prompt
+    elif len(args.validation_prompt) == 1:
+        validation_prompts = args.validation_prompt * len(validation_images)
+    else:
+        raise ValueError(
+            "number of `args.validation_image` and `args.validation_prompt` should be checked in `parse_args`"
+        )
+
+    image_logs = []
+    inference_ctx = contextlib.nullcontext() if is_final_validation else torch.autocast("cuda")
+
+    for validation_prompt, validation_image in zip(validation_prompts, validation_images):
+        validation_image = Image.open(validation_image).convert("RGB")
+        # Resize
+        validation_image = transforms.Resize(args.resolution)(validation_image)
+
+        images = []
+
+        for _ in range(args.num_validation_images):
+            with inference_ctx:
+                image = pipeline(
+                    validation_prompt, validation_image, num_inference_steps=20, generator=generator
+                ).images[0]
+
+            images.append(image)
+
+        image_logs.append(
+            {"validation_image": validation_image, "images": images, "validation_prompt": validation_prompt}
+        )
+
+    tracker_key = "test" if is_final_validation else "validation"
+    for tracker in accelerator.trackers:
+        if tracker.name == "tensorboard":
+            for log in image_logs:
+                images = log["images"]
+                validation_prompt = log["validation_prompt"]
+                validation_image = log["validation_image"]
+
+                formatted_images = []
+
+                formatted_images.append(np.asarray(validation_image))
+
+                for image in images:
+                    formatted_images.append(np.asarray(image))
+
+                formatted_images = np.stack(formatted_images)
+
+                tracker.writer.add_images(validation_prompt, formatted_images, step, dataformats="NHWC")
+        elif tracker.name == "wandb":
+            formatted_images = []
+
+            for log in image_logs:
+                images = log["images"]
+                validation_prompt = log["validation_prompt"]
+                validation_image = log["validation_image"]
+
+                formatted_images.append(wandb.Image(validation_image, caption="Controlnet conditioning"))
+
+                for image in images:
+                    image = wandb.Image(image, caption=validation_prompt)
+                    formatted_images.append(image)
+
+            tracker.log({tracker_key: formatted_images})
+        else:
+            logger.warning(f"image logging not implemented for {tracker.name}")
+
+        del pipeline
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        return image_logs
+
+
+def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: str, revision: str):
+    text_encoder_config = PretrainedConfig.from_pretrained(
+        pretrained_model_name_or_path,
+        subfolder="text_encoder",
+        revision=revision,
+    )
+    model_class = text_encoder_config.architectures[0]
+
+    if model_class == "CLIPTextModel":
+        from transformers import CLIPTextModel
+
+        return CLIPTextModel
+    elif model_class == "RobertaSeriesModelWithTransformation":
+        from diffusers.pipelines.alt_diffusion.modeling_roberta_series import RobertaSeriesModelWithTransformation
+
+        return RobertaSeriesModelWithTransformation
+    else:
+        raise ValueError(f"{model_class} is not supported.")
+
+
+def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):
+    img_str = ""
+    if image_logs is not None:
+        img_str = "You can find some example images below.\n\n"
+        for i, log in enumerate(image_logs):
+            images = log["images"]
+            validation_prompt = log["validation_prompt"]
+            validation_image = log["validation_image"]
+            validation_image.save(os.path.join(repo_folder, "image_control.png"))
+            img_str += f"prompt: {validation_prompt}\n"
+            images = [validation_image] + images
+            image_grid(images, 1, len(images)).save(os.path.join(repo_folder, f"images_{i}.png"))
+            img_str += f"![images_{i})](./images_{i}.png)\n"
+
+    model_description = f"""
+# controlnet-{repo_id}
+
+These are controlnet weights trained on {base_model} with new type of conditioning.
+{img_str}
+"""
+    model_card = load_or_create_model_card(
+        repo_id_or_path=repo_id,
+        from_training=True,
+        license="creativeml-openrail-m",
+        base_model=base_model,
+        model_description=model_description,
+        inference=True,
+    )
+
+    tags = [
+        "stable-diffusion",
+        "stable-diffusion-diffusers",
+        "text-to-image",
+        "diffusers",
+        "controlnet",
+        "diffusers-training",
+    ]
+    model_card = populate_model_card(model_card, tags=tags)
+
+    model_card.save(os.path.join(repo_folder, "README.md"))
+
+
+def parse_args(input_args=None):
+    parser = argparse.ArgumentParser(description="Simple example of a ControlNet 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(
+        "--controlnet_model_name_or_path",
+        type=str,
+        default=None,
+        help="Path to pretrained controlnet model or model identifier from huggingface.co/models."
+        " If not specified controlnet weights are initialized from unet.",
+    )
+    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(
+        "--tokenizer_name",
+        type=str,
+        default=None,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="controlnet-model",
+        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(
+        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
+    )
+    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. Checkpoints can be used for resuming training via `--resume_from_checkpoint`. "
+            "In the case that the checkpoint is better than the final trained model, the checkpoint can also be used for inference."
+            "Using a checkpoint for inference requires separate loading of the original pipeline and the individual checkpointed model components."
+            "See https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint for step by step"
+            "instructions."
+        ),
+    )
+    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=5e-6,
+        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(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
+    )
+    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
+    parser.add_argument(
+        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
+    )
+    parser.add_argument(
+        "--dataset_num_workers",
+        type=int,
+        default=1,
+        help=(
+            "Number of subprocesses to use for data loading."
+        ),
+    )
+    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(
+        "--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(
+        "--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(
+        "--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(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
+    parser.add_argument(
+        "--set_grads_to_none",
+        action="store_true",
+        help=(
+            "Save more memory by using setting grads to None instead of zero. Be aware, that this changes certain"
+            " behaviors, so disable this argument if it causes any problems. More info:"
+            " https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html"
+        ),
+    )
+    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 the target image."
+    )
+    parser.add_argument(
+        "--conditioning_image_column",
+        type=str,
+        default="conditioning_image",
+        help="The column of the dataset containing the controlnet conditioning 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(
+        "--proportion_empty_prompts",
+        type=float,
+        default=0,
+        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
+    )
+    parser.add_argument(
+        "--validation_prompt",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of prompts evaluated every `--validation_steps` and logged to `--report_to`."
+            " Provide either a matching number of `--validation_image`s, a single `--validation_image`"
+            " to be used with all prompts, or a single prompt that will be used with all `--validation_image`s."
+        ),
+    )
+    parser.add_argument(
+        "--validation_image",
+        type=str,
+        default=None,
+        nargs="+",
+        help=(
+            "A set of paths to the controlnet conditioning image be evaluated every `--validation_steps`"
+            " and logged to `--report_to`. Provide either a matching number of `--validation_prompt`s, a"
+            " a single `--validation_prompt` to be used with all `--validation_image`s, or a single"
+            " `--validation_image` that will be used with all `--validation_prompt`s."
+        ),
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images to be generated for each `--validation_image`, `--validation_prompt` pair",
+    )
+    parser.add_argument(
+        "--validation_steps",
+        type=int,
+        default=100,
+        help=(
+            "Run validation every X steps. Validation consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`"
+            " and logging the images."
+        ),
+    )
+    parser.add_argument(
+        "--tracker_project_name",
+        type=str,
+        default="train_controlnet",
+        help=(
+            "The `project_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+    parser.add_argument(
+        "--tracker_run_name",
+        type=str,
+        default=None,
+        help=(
+            "The `run_name` argument passed to Accelerator.init_trackers for"
+            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
+        ),
+    )
+
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Specify either `--dataset_name` or `--train_data_dir`")
+
+    if args.dataset_name is not None and args.train_data_dir is not None:
+        raise ValueError("Specify only one of `--dataset_name` or `--train_data_dir`")
+
+    if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1:
+        raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].")
+
+    if args.validation_prompt is not None and args.validation_image is None:
+        raise ValueError("`--validation_image` must be set if `--validation_prompt` is set")
+
+    if args.validation_prompt is None and args.validation_image is not None:
+        raise ValueError("`--validation_prompt` must be set if `--validation_image` is set")
+
+    if (
+        args.validation_image is not None
+        and args.validation_prompt is not None
+        and len(args.validation_image) != 1
+        and len(args.validation_prompt) != 1
+        and len(args.validation_image) != len(args.validation_prompt)
+    ):
+        raise ValueError(
+            "Must provide either 1 `--validation_image`, 1 `--validation_prompt`,"
+            " or the same number of `--validation_prompt`s and `--validation_image`s"
+        )
+
+    if args.resolution % 8 != 0:
+        raise ValueError(
+            "`--resolution` must be divisible by 8 for consistently sized encoded images between the VAE and the controlnet encoder."
+        )
+
+    return args
+
+class PrefetchIterableDataset(IterableDataset):
+    def __init__(self, image_files, resizer, transforms, conditioning_transforms, tokenizer, prefetch_factor=16):
+        ex_iterable = ExamplesIterable(self._generate_examples, kwargs={
+            "image_files": image_files,
+            "resizer": resizer,
+            "transforms": transforms,
+            "conditioning_transforms": conditioning_transforms,
+            "tokenizer": tokenizer
+        })
+
+        # Call the parent's __init__ with the created ex_iterable
+        super(PrefetchIterableDataset, self).__init__(ex_iterable=ex_iterable)
+        self.image_files = image_files
+        self.resizer = resizer
+        self.transforms = transforms
+        self.conditioning_transforms = conditioning_transforms
+        self.tokenizer = tokenizer
+        self.prefetch_factor = prefetch_factor
+        self.queue = Queue(maxsize=prefetch_factor)
+        # 미리 계산된 input_ids
+        self.empty_input_ids = tokenizer(
+            "",
+            max_length=tokenizer.model_max_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt"
+        ).input_ids[0]
+
+    def _generate_examples(self, image_files, resizer, transforms, conditioning_transforms, tokenizer):
+        for image_file in image_files:
+            yield self.preprocess_image(image_file)
+
+    def preprocess_image(self, image_file):
+        image_path = os.path.join(args.train_data_dir, image_file)
+        image = Image.open(image_path).convert("RGB")
+        image = self.resizer(image)
+        return {
+            "pixel_values": self.transforms(image),
+            "conditioning_pixel_values": self.conditioning_transforms(image),
+            "input_ids": self.empty_input_ids,
+        }
+
+    def producer(self):
+        with Pool(self.num_workers) as p:
+            for item in p.imap(self.preprocess_image, cycle(self.image_files)):
+                self.queue.put(item)
+
+    def __iter__(self):
+        worker_info = get_worker_info()
+        if worker_info is None:  # single-process data loading
+            iter_start = 0
+            iter_end = len(self.image_files)
+        else:  # in a worker process
+            per_worker = int(math.ceil(len(self.image_files) / float(worker_info.num_workers)))
+            worker_id = worker_info.id
+            iter_start = worker_id * per_worker
+            iter_end = min(iter_start + per_worker, len(self.image_files))
+
+        # 각 워커에 할당된 이미지 파일
+        worker_image_files = self.image_files[iter_start:iter_end]
+
+        # 프리페치를 위한 큐 생성
+        queue = Queue(maxsize=self.prefetch_factor)
+
+        # 프로듀서 함수 정의
+        def producer():
+            for image_file in cycle(worker_image_files):  # cycle을 사용하여 무한 반복
+                item = self.preprocess_image(image_file)
+                queue.put(item)
+
+        # 프로듀서 스레드 시작
+        thread = Thread(target=producer)
+        thread.daemon = True
+        thread.start()
+
+        # 아이템 yield
+        while True:
+            yield queue.get()
+
+
+def collate_fn(examples):
+    pixel_values = torch.stack([example["pixel_values"] for example in examples])
+    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    conditioning_pixel_values = torch.stack([example["conditioning_pixel_values"] for example in examples])
+    conditioning_pixel_values = conditioning_pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    input_ids = torch.stack([example["input_ids"] for example in examples])
+
+    return {
+        "pixel_values": pixel_values,
+        "conditioning_pixel_values": conditioning_pixel_values,
+        "input_ids": input_ids,
+    }
+
+def main(args):
+    if args.report_to == "wandb" and args.hub_token is not None:
+        raise ValueError(
+            "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
+            " Please use `huggingface-cli login` to authenticate with the Hub."
+        )
+    os.environ["WANDB__SERVICE_WAIT"] = "300"
+
+    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,
+    )
+
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
+    # 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)
+    logger.info(f"Training/evaluation parameters {args}")
+    if accelerator.is_local_main_process:
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        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 tokenizer
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
+    elif args.pretrained_model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="tokenizer",
+            revision=args.revision,
+            use_fast=False,
+        )
+
+    # import correct text encoder class
+    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)
+
+    # Load scheduler and models
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    text_encoder = text_encoder_cls.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    vae = AutoencoderKL.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
+    )
+    unet = UNet2DConditionModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
+    )
+
+    if args.controlnet_model_name_or_path:
+        logger.info("Loading existing controlnet weights")
+        controlnet = ControlNetModel.from_pretrained(args.controlnet_model_name_or_path)
+    else:
+        logger.info("Initializing controlnet weights from unet")
+        controlnet = ControlNetModel.from_unet(unet)
+
+    # Taken from [Sayak Paul's Diffusers PR #6511](https://github.com/huggingface/diffusers/pull/6511/files)
+    def unwrap_model(model):
+        model = accelerator.unwrap_model(model)
+        model = model._orig_mod if is_compiled_module(model) else model
+        return model
+
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
+        # 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:
+                i = len(weights) - 1
+
+                while len(weights) > 0:
+                    weights.pop()
+                    model = models[i]
+
+                    sub_dir = "controlnet"
+                    model.save_pretrained(os.path.join(output_dir, sub_dir))
+
+                    i -= 1
+
+        def load_model_hook(models, input_dir):
+            while len(models) > 0:
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                # load diffusers style into model
+                load_model = ControlNetModel.from_pretrained(input_dir, subfolder="controlnet")
+                model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
+    vae.requires_grad_(False)
+    unet.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+    controlnet.train()
+
+    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.warning(
+                    "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()
+            controlnet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
+
+    if args.gradient_checkpointing:
+        controlnet.enable_gradient_checkpointing()
+
+    # Check that all trainable models are in full precision
+    low_precision_error_string = (
+        " Please make sure to always have all model weights in full float32 precision when starting training - even if"
+        " doing mixed precision training, copy of the weights should still be float32."
+    )
+
+    if unwrap_model(controlnet).dtype != torch.float32:
+        raise ValueError(
+            f"Controlnet loaded as datatype {unwrap_model(controlnet).dtype}. {low_precision_error_string}"
+        )
+
+    # 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
+
+    # Optimizer creation
+    params_to_optimize = controlnet.parameters()
+    optimizer = optimizer_class(
+        params_to_optimize,
+        lr=args.learning_rate,
+        betas=(args.adam_beta1, args.adam_beta2),
+        weight_decay=args.adam_weight_decay,
+        eps=args.adam_epsilon,
+    )
+
+    logger.info("Loading the training dataset")
+    image_files = [f for f in os.listdir(args.train_data_dir) if
+                   f.lower().endswith(('.png', '.jpg', '.jpeg', '.bmp', '.gif'))]
+
+    # Define image transforms
+    image_resizers = transforms.Compose([
+        transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+        transforms.CenterCrop(args.resolution),
+    ])
+    image_transforms = transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize([0.5], [0.5]),
+    ])
+    conditioning_image_transforms = transforms.Compose([
+        transforms.Grayscale(num_output_channels=3),
+        transforms.ToTensor(),
+    ])
+
+    train_dataset = PrefetchIterableDataset(
+        image_files=image_files,
+        resizer=image_resizers,
+        transforms=image_transforms,
+        conditioning_transforms=conditioning_image_transforms,
+        tokenizer=tokenizer,
+        prefetch_factor=128,
+    )
+
+    logger.info("Creating the training dataloader")
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        batch_size=args.train_batch_size,
+        collate_fn=collate_fn,
+        num_workers=32,
+        pin_memory=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,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+    )
+
+    # Prepare everything with our `accelerator`.
+    controlnet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        controlnet, optimizer, train_dataloader, lr_scheduler
+    )
+
+    # For mixed precision training we cast the text_encoder and vae weights to half-precision
+    # as these models 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 vae, unet and text_encoder to device and cast to weight_dtype
+    vae.to(accelerator.device, dtype=weight_dtype)
+    unet.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+
+    # 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:
+        tracker_config = dict(vars(args))
+
+        # tensorboard cannot handle list types for config
+        tracker_config.pop("validation_prompt")
+        tracker_config.pop("validation_image")
+
+        logger.info(f"Init trackers: {args.tracker_project_name}")
+        accelerator.init_trackers(args.tracker_project_name, config=tracker_config)
+        # if args.tracker_run_name is not None:
+        #     accelerator.trackers[-1].run.name = args.tracker_run_name
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    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
+    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,
+    )
+
+    image_logs = None
+    while True:
+        for batch in train_dataloader:
+            try:
+                with accelerator.accumulate(controlnet):
+                    # 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)
+                    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
+                    encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]
+
+                    controlnet_image = batch["conditioning_pixel_values"].to(dtype=weight_dtype)
+
+                    down_block_res_samples, mid_block_res_sample = controlnet(
+                        noisy_latents,
+                        timesteps,
+                        encoder_hidden_states=encoder_hidden_states,
+                        controlnet_cond=controlnet_image,
+                        return_dict=False,
+                    )
+
+                    # Predict the noise residual
+                    model_pred = unet(
+                        noisy_latents,
+                        timesteps,
+                        encoder_hidden_states=encoder_hidden_states,
+                        down_block_additional_residuals=[
+                            sample.to(dtype=weight_dtype) for sample in down_block_res_samples
+                        ],
+                        mid_block_additional_residual=mid_block_res_sample.to(dtype=weight_dtype),
+                        return_dict=False,
+                    )[0]
+
+                    # Get the target for loss depending on the 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}")
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+
+                    accelerator.backward(loss)
+                    if accelerator.sync_gradients:
+                        params_to_clip = controlnet.parameters()
+                        accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+                    optimizer.step()
+                    lr_scheduler.step()
+                    optimizer.zero_grad(set_to_none=args.set_grads_to_none)
+
+                # Checks if the accelerator has performed an optimization step behind the scenes
+                if accelerator.sync_gradients:
+                    progress_bar.update(1)
+                    global_step += 1
+
+                    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}")
+
+                        if args.validation_prompt is not None and global_step % args.validation_steps == 0:
+                            image_logs = log_validation(
+                                vae,
+                                text_encoder,
+                                tokenizer,
+                                unet,
+                                controlnet,
+                                args,
+                                accelerator,
+                                weight_dtype,
+                                global_step,
+                            )
+
+                logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+                progress_bar.set_postfix(**logs)
+                accelerator.log(logs, step=global_step)
+
+                if global_step >= args.max_train_steps:
+                    break
+            except Exception as e:
+                logger.warning(f"Error processing batch: {str(e)}")
+                continue
+
+        if global_step >= args.max_train_steps:
+            break
+
+    # Create the pipeline using using the trained modules and save it.
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        controlnet = unwrap_model(controlnet)
+        controlnet.save_pretrained(args.output_dir)
+
+        # Run a final round of validation.
+        image_logs = None
+        if args.validation_prompt is not None:
+            image_logs = log_validation(
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                unet=unet,
+                controlnet=None,
+                args=args,
+                accelerator=accelerator,
+                weight_dtype=weight_dtype,
+                step=global_step,
+                is_final_validation=True,
+            )
+
+        if args.push_to_hub:
+            save_model_card(
+                repo_id,
+                image_logs=image_logs,
+                base_model=args.pretrained_model_name_or_path,
+                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_*"],
+            )
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)