trainT2I.py

import argparse
import logging
import math
import os
import random
import shutil
from pathlib import Path

import accelerate
import datasets
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.state import AcceleratorState
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset, Dataset
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 CLIPTextModel, CLIPTokenizer
from transformers.utils import ContextManagers

import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.utils import check_min_version, deprecate, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available

import gc

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("0.19.0.dev0")

logger = get_logger(__name__, log_level="INFO")

DATASET_NAME_MAPPING = {
    "lambdalabs/pokemon-blip-captions": ("image", "text"),
}


def make_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, args, accelerator, weight_dtype, epoch):
    logger.info("Running validation... ")

    pipeline = StableDiffusionPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        vae=accelerator.unwrap_model(vae),
        text_encoder=accelerator.unwrap_model(text_encoder),
        tokenizer=tokenizer,
        unet=accelerator.unwrap_model(unet),
        safety_checker=None,
        revision=args.revision,
        torch_dtype=weight_dtype
    )
    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)

    images = []
    for i in range(len(args.validation_prompts)):
        with torch.autocast("cuda"):
            image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]

        images.append(image)

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
            tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
        elif tracker.name == "wandb":
            tracker.log(
                {
                    "validation": [
                        wandb.Image(image, caption=f"{i}: {args.validation_prompts[i]}")
                        for i, image in enumerate(images)
                    ]
                }
            )
        else:
            logger.warn(f"image logging not implemented for {tracker.name}")

    del pipeline
    torch.cuda.empty_cache()

    return images


from dataclasses import dataclass

@dataclass
class TrainingConfig:
    pretrained_model_name_or_path = "SG161222/Realistic_Vision_V4.0"
    image_size = 512  # the generated image resolution
    train_batch_size = 32
    max_train_steps = 10000
    gradient_accumulation_steps = 4
    learning_rate = 1e-4
    feature_weight=0.5
    output_weight=0.5
    lr_warmup_steps = 0
    eval_batch_size = 16  # how many images to sample during evaluation
    num_train_epochs = 1
    mixed_precision = "fp16"  # `no` for float32, `fp16` for automatic mixed precision
    output_dir = "Checkpoints"  
    train_data_dir = "data"
    save_dir = "Saved"
    logging_dir = "logs"
    overwrite_output_dir = True  # overwrite the old model when re-running the notebook
    seed = 647
    report_to = "wandb"
    revision = None
    distill_level = "sd_small" # One of "sd_small" or "sd_tiny"
    resume_from_checkpoint = None
    use_ema = False
    enable_xformers_memory_efficient_attention = True
    gradient_checkpointing = True
    allow_tf32 = True #True for Ampere Backend
    scale_lr = False
    use_8bit_adam = False
    adam_beta1 = 0.9
    adam_beta2 = 0.999
    adam_weight_decay = 1e-2
    adam_epsilon = 1e-8
    max_grad_norm = 1
    dataset_name = None
    cache_dir = None
    image_column = "image"
    caption_column = "caption"
    max_train_samples = None
    validation_prompts = None
    resolution = 512
    center_crop = True
    random_flip = True
    dataloader_num_workers = 0
    lr_scheduler = "cosine"
    tracker_project_name = "text2image-fine-tune"
    noise_offset = 0
    input_perturbation = 0
    prediction_type = None
    checkpointing_steps = 1000
    checkpoints_total_limit = 1
    non_ema_revision = None
    validation_prompts = None
    saving_steps = 5000

def main():
    config = TrainingConfig()
    logging_dir = os.path.join(config.output_dir, config.logging_dir)

    accelerator_project_config = ProjectConfiguration(project_dir=config.output_dir, logging_dir=logging_dir)

    accelerator = Accelerator(
        gradient_accumulation_steps=config.gradient_accumulation_steps,
        mixed_precision=config.mixed_precision,
        log_with=config.report_to,
        project_config=accelerator_project_config,
    )

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()
        diffusers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if config.seed is not None:
        set_seed(config.seed)

    # Handle the repository creation
    if accelerator.is_main_process:
        if config.output_dir is not None:
            os.makedirs(config.output_dir, exist_ok=True)

        # if args.push_to_hub:
        #     repo_id = create_repo(
        #         repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
        #     ).repo_id

    # Load scheduler, tokenizer and models.
    noise_scheduler = DDPMScheduler.from_pretrained(config.pretrained_model_name_or_path, subfolder="scheduler")
    tokenizer = CLIPTokenizer.from_pretrained(
        config.pretrained_model_name_or_path, subfolder="tokenizer", revision=config.revision
    )

    def deepspeed_zero_init_disabled_context_manager():
        """
        returns either a context list that includes one that will disable zero.Init or an empty context list
        """
        deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
        if deepspeed_plugin is None:
            return []

        return [deepspeed_plugin.zero3_init_context_manager(enable=False)]

    # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3.
    # For this to work properly all models must be run through `accelerate.prepare`. But accelerate
    # will try to assign the same optimizer with the same weights to all models during
    # `deepspeed.initialize`, which of course doesn't work.
    #
    # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2
    # frozen models from being partitioned during `zero.Init` which gets called during
    # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding
    # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded.
    with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
        text_encoder = CLIPTextModel.from_pretrained(
            config.pretrained_model_name_or_path, subfolder="text_encoder", revision=config.revision
        )
        vae = AutoencoderKL.from_pretrained(
            config.pretrained_model_name_or_path, subfolder="vae", revision=config.revision
        )

    def prepare_unet(unet, model_type):
        assert model_type in ["sd_tiny", "sd_small"]
        # Set mid block to None if mode is other than base
        if model_type != "sd_small":
            unet.mid_block = None
            
        # Commence deletion of resnets/attentions inside the U-net
        # Handle Down Blocks
        for i in range(3):
            delattr(unet.down_blocks[i].resnets, "1")
            delattr(unet.down_blocks[i].attentions, "1")
    
        if model_type == "sd_tiny":
            delattr(unet.down_blocks, "3")
            unet.down_blocks[2].downsamplers = None
    
        else:
            delattr(unet.down_blocks[3].resnets, "1")
        # Handle Up blocks
    
        unet.up_blocks[0].resnets[1] = unet.up_blocks[0].resnets[2]
        delattr(unet.up_blocks[0].resnets, "2")
        for i in range(1, 4):
            unet.up_blocks[i].resnets[1] = unet.up_blocks[i].resnets[2]
            unet.up_blocks[i].attentions[1] = unet.up_blocks[i].attentions[2]
            delattr(unet.up_blocks[i].attentions, "2")
            delattr(unet.up_blocks[i].resnets, "2")
        if model_type == "sd_tiny":
            for i in range(3):
                unet.up_blocks[i] = unet.up_blocks[i + 1]
            delattr(unet.up_blocks, "3")
        torch.cuda.empty_cache()
        gc.collect()

    unet = UNet2DConditionModel.from_pretrained(
        config.pretrained_model_name_or_path, subfolder="unet", revision=config.non_ema_revision
    )

    
    KD_teacher_unet=UNet2DConditionModel.from_pretrained(
        config.pretrained_model_name_or_path, subfolder="unet", revision=config.non_ema_revision
    )

    assert config.distill_level=="sd_small" or config.distill_level=="sd_tiny"
    if config.distill_level:
        prepare_unet(unet, config.distill_level)
        
    unet.load_state_dict(torch.load("trained_unet/unet.pt"))
    # Freeze vae and text_encoder
    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)
    KD_teacher_unet.requires_grad_(False)

    # Create EMA for the unet.
    if config.use_ema:
        ema_unet = UNet2DConditionModel.from_pretrained(
            config.pretrained_model_name_or_path, subfolder="unet", revision=config.revision
        )
        ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config)

    if config.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
                logger.warn(
                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
                )
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")

    def compute_snr(timesteps):
        """
        Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
        """
        alphas_cumprod = noise_scheduler.alphas_cumprod
        sqrt_alphas_cumprod = alphas_cumprod**0.5
        sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5

        # Expand the tensors.
        # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
        sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
        while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
            sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
        alpha = sqrt_alphas_cumprod.expand(timesteps.shape)

        sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
        while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
            sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
        sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)

        # Compute SNR.
        snr = (alpha / sigma) ** 2
        return snr

    # `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 config.use_ema:
                ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))

            for i, model in enumerate(models):
                model.save_pretrained(os.path.join(output_dir, "unet"))

                # make sure to pop weight so that corresponding model is not saved again
                weights.pop()

        def load_model_hook(models, input_dir):
            if config.use_ema:
                load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
                ema_unet.load_state_dict(load_model.state_dict())
                ema_unet.to(accelerator.device)
                del load_model

            for i in range(len(models)):
                # pop models so that they are not loaded again
                model = models.pop()

                # load diffusers style into model
                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
                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)

    if config.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if config.allow_tf32:
        torch.backends.cuda.matmul.allow_tf32 = True

    if config.scale_lr:
        config.learning_rate = (
            config.learning_rate * config.gradient_accumulation_steps * config.train_batch_size * accelerator.num_processes
        )

    # Initialize the optimizer
    if config.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
            )

        optimizer_cls = bnb.optim.AdamW8bit
    else:
        optimizer_cls = torch.optim.AdamW

    optimizer = optimizer_cls(
        unet.parameters(),
        lr=config.learning_rate,
        betas=(config.adam_beta1, config.adam_beta2),
        weight_decay=config.adam_weight_decay,
        eps=config.adam_epsilon,
    )

    # 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 config.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            config.dataset_name,
            config.dataset_config_name,
            cache_dir=config.cache_dir,
        )
    else:
        def get_text(x):
            with open(x, 'rb') as f:
                data = f.readline()
            return data
        data_files = {}
        data_files["image"] = [os.path.join(config.train_data_dir, path) for path in os.listdir(config.train_data_dir) if path.endswith(".jpg")]
        data_files["caption"] = [os.path.join(config.train_data_dir, path) for path in os.listdir(config.train_data_dir) if path.endswith(".txt")]
        dataset = Dataset.from_dict(
            data_files
        ).cast_column("image", datasets.Image())
        dataset = dataset.map(lambda batch: {"image": batch["image"], "caption":get_text(batch['caption'])})
        # data_files = {}
        # if config.train_data_dir is not None:
        #     data_files["train"] = os.path.join(config.train_data_dir, "*.parquet")
        # dataset = load_dataset(
        #     "imagefolder",
        #     data_files=data_files,
        #     cache_dir=config.cache_dir,
        # )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    # column_names = dataset["train"].column_names
    column_names = dataset.column_names
    # 6. Get the column names for input/target.
    dataset_columns = DATASET_NAME_MAPPING.get(config.dataset_name, None)
    if config.image_column is None:
        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
    else:
        image_column = config.image_column
        if image_column not in column_names:
            raise ValueError(
                f"--image_column' value '{config.image_column}' needs to be one of: {', '.join(column_names)}"
            )
    if config.caption_column is None:
        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
    else:
        caption_column = config.caption_column
        if caption_column not in column_names:
            raise ValueError(
                f"--caption_column' value '{config.caption_column}' needs to be one of: {', '.join(column_names)}"
            )

    # Preprocessing the datasets.
    # We need to tokenize input captions and transform the images.
    def tokenize_captions(examples, is_train=True):
        captions = []
        for caption in examples[caption_column]:
            if 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

    # Preprocessing the datasets.
    train_transforms = transforms.Compose(
        [
            transforms.Resize(config.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(config.resolution) if config.center_crop else transforms.RandomCrop(config.resolution),
            transforms.RandomHorizontalFlip() if config.random_flip else transforms.Lambda(lambda x: x),
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    def preprocess_train(examples):
        images = [image.convert("RGB") for image in examples[image_column]]
        examples["pixel_values"] = [train_transforms(image) for image in images]
        examples["input_ids"] = tokenize_captions(examples)
        return examples

    with accelerator.main_process_first():
        if config.max_train_samples is not None:
            # dataset["train"] = dataset["train"].shuffle(seed=config.seed).select(range(config.max_train_samples))
            dataset = dataset.shuffle(seed=config.seed).select(range(config.max_train_samples))
        # Set the training transforms
        # train_dataset = dataset["train"].with_transform(preprocess_train)
        train_dataset = dataset.with_transform(preprocess_train)

    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()
        input_ids = torch.stack([example["input_ids"] for example in examples])
        return {"pixel_values": pixel_values, "input_ids": input_ids}

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=collate_fn,
        batch_size=config.train_batch_size,
        num_workers=config.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) / config.gradient_accumulation_steps)
    if config.max_train_steps is None:
        config.max_train_steps = config.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

    lr_scheduler = get_scheduler(
        config.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=config.lr_warmup_steps * config.gradient_accumulation_steps,
        num_training_steps=config.max_train_steps * config.gradient_accumulation_steps,
    )

    # Prepare everything with our `accelerator`.
    unet, optimizer, train_dataloader, lr_scheduler,KD_teacher_unet = accelerator.prepare(
        unet, optimizer, train_dataloader, lr_scheduler,KD_teacher_unet
    )

    if config.use_ema:
        ema_unet.to(accelerator.device)

    # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
        config.mixed_precision = accelerator.mixed_precision
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16
        config.mixed_precision = accelerator.mixed_precision

    # Move text_encode and vae to gpu and cast to weight_dtype
    text_encoder.to(accelerator.device, dtype=weight_dtype)
    vae.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) / config.gradient_accumulation_steps)
    if overrode_max_train_steps:
        config.max_train_steps = config.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    config.num_train_epochs = math.ceil(config.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(config))
        # print(tracker_config)
        # tracker_config.pop("validation_prompts")
        accelerator.init_trackers(config.tracker_project_name, tracker_config)

    # Train!
    total_batch_size = config.train_batch_size * accelerator.num_processes * config.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {config.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {config.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {config.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {config.max_train_steps}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    if config.resume_from_checkpoint:
        if config.resume_from_checkpoint != "latest":
            path = os.path.basename(config.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = os.listdir(config.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 '{config.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            config.resume_from_checkpoint = None
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(config.output_dir, path))
            global_step = int(path.split("-")[1])

            resume_global_step = global_step * config.gradient_accumulation_steps
            first_epoch = global_step // num_update_steps_per_epoch
            resume_step = resume_global_step % (num_update_steps_per_epoch * config.gradient_accumulation_steps)

    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(global_step, config.max_train_steps), disable=not accelerator.is_local_main_process)
    progress_bar.set_description("Steps")

    KD_teacher = {}
    KD_student= {}

    
    def getActivation(activation,name,residuals_present):
        # the hook signature
        if residuals_present:
            def hook(model, input, output):
                activation[name] = output[0]
        else:
            def hook(model, input, output):
                activation[name] = output
        return hook
        
    def cast_hook(unet,dicts,model_type,teacher=False):
        if teacher:
            for i in range(4):
                unet.down_blocks[i].register_forward_hook(getActivation(dicts,'d'+str(i),True))
            unet.mid_block.register_forward_hook(getActivation(dicts,'m',False))
            for i in range(4):
                unet.up_blocks[i].register_forward_hook(getActivation(dicts,'u'+str(i),False))
        else:
            num_blocks= 4 if model_type=="sd_small" else 3
            for i in range(num_blocks):
                unet.down_blocks[i].register_forward_hook(getActivation(dicts,'d'+str(i),True))
            if model_type=="sd_small"
                unet.mid_block.register_forward_hook(getActivation(dicts,'m',False))
            for i in range(num_blocks):
                unet.up_blocks[i].register_forward_hook(getActivation(dicts,'u'+str(i),False))
            
    
    cast_hook(unet,KD_student,config.distill_level,False)
    cast_hook(KD_teacher_unet,KD_teacher,config.distill_level,True)

    

    for epoch in range(first_epoch, config.num_train_epochs):
        unet.train()
        train_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            # Skip steps until we reach the resumed step
            if config.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
                if step % config.gradient_accumulation_steps == 0:
                    progress_bar.update(1)
                continue

            with accelerator.accumulate(unet):
                # Convert images to latent space
                latents = vae.encode(batch["pixel_values"].to(weight_dtype)).latent_dist.sample()
                latents = latents * vae.config.scaling_factor

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(latents)
                if config.noise_offset:
                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
                    noise += config.noise_offset * torch.randn(
                        (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
                    )
                if config.input_perturbation:
                    new_noise = noise + config.input_perturbation * torch.randn_like(noise)
                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)
                if config.input_perturbation:
                    noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps)
                else:
                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

                # Get the text embedding for conditioning
                encoder_hidden_states = text_encoder(batch["input_ids"])[0]

                # Get the target for loss depending on the prediction type
                if config.prediction_type is not None:
                    # set prediction_type of scheduler if defined
                    noise_scheduler.register_to_config(prediction_type=config.prediction_type)

                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

                # Predict the noise residual and compute loss
                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
                with torch.no_grad():
                    teacher_pred=KD_teacher_unet(noisy_latents, timesteps, encoder_hidden_states).sample

                loss_features=0
                if configs.distill_level=="sd_small":
                    for i in range(4):
                        loss_features=loss_features+F.mse_loss(KD_teacher['d'+str(i)],KD_student['d'+str(i)])
                    loss_features=loss_features+F.mse_loss(KD_teacher['m'],KD_student['m'])
                    for i in range(4):
                        loss_features=loss_features+F.mse_loss(KD_teacher['u'+str(i)],KD_student['u'+str(i)])
                else:
                    for i in range(2):
                        loss_features=loss_features+F.mse_loss(KD_teacher['d'+str(i)],KD_student['d'+str(i)])
                    loss_features=loss_features+F.mse_loss(KD_teacher['u'+str(0)],KD_student['d'+str(2)])
                    for i in range(3):
                        loss_features=loss_features+F.mse_loss(KD_teacher['u'+str(i+1)],KD_student['u'+str(i)])
                    
                loss_KD=F.mse_loss(model_pred.float(), teacher_pred.float(), reduction="mean")
                loss_task = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                loss=loss_task+config.output_weight*loss_KD+config.feature_weight*loss_features

                # Gather the losses across all processes for logging (if we use distributed training).
                avg_loss = accelerator.gather(loss.repeat(config.train_batch_size)).mean()
                train_loss += avg_loss.item() / config.gradient_accumulation_steps

                # Backpropagate
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(unet.parameters(), config.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()
                torch.cuda.empty_cache()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                if config.use_ema:
                    ema_unet.step(unet.parameters())
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train_loss": train_loss}, step=global_step)
                train_loss = 0.0
                if global_step % config.saving_steps == 0:
                    pipeline = StableDiffusionPipeline.from_pretrained(
                        config.pretrained_model_name_or_path,
                        text_encoder=text_encoder,
                        vae=vae,
                        unet=unet,
                        revision=config.revision,
                    )
                    pipeline.save_pretrained(config.save_dir + f"/{str(global_step)}")
                if global_step % config.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                        if config.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(config.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) >= config.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - config.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(config.output_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)

                        save_path = os.path.join(config.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")

            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
            progress_bar.set_postfix(**logs)

            if global_step >= config.max_train_steps:
                break

        if accelerator.is_main_process:
            if config.validation_prompts is not None and epoch % config.validation_epochs == 0:
                if config.use_ema:
                    # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
                    ema_unet.store(unet.parameters())
                    ema_unet.copy_to(unet.parameters())
                log_validation(
                    vae,
                    text_encoder,
                    tokenizer,
                    unet,
                    config,
                    accelerator,
                    weight_dtype,
                    global_step,
                )
                if config.use_ema:
                    # Switch back to the original UNet parameters.
                    ema_unet.restore(unet.parameters())

    # Create the pipeline using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        if config.use_ema:
            ema_unet.copy_to(unet.parameters())
        torch.save(unet.state_dict(), "trained_unet/unet.pt")
        pipeline = StableDiffusionPipeline.from_pretrained(
            config.pretrained_model_name_or_path,
            text_encoder=text_encoder,
            vae=vae,
            unet=unet,
            revision=config.revision,
        )
        pipeline.save_pretrained(config.output_dir)

        # Run a final round of inference.
        images = []
        if config.validation_prompts is not None:
            logger.info("Running inference for collecting generated images...")
            pipeline = pipeline.to(accelerator.device)
            pipeline.torch_dtype = weight_dtype
            pipeline.set_progress_bar_config(disable=True)

            if config.enable_xformers_memory_efficient_attention:
                pipeline.enable_xformers_memory_efficient_attention()

            if config.seed is None:
                generator = None
            else:
                generator = torch.Generator(device=accelerator.device).manual_seed(config.seed)

            for i in range(len(config.validation_prompts)):
                with torch.autocast("cuda"):
                    image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]
                images.append(image)

        # if config.push_to_hub:
        #     save_model_card(args, repo_id, images, 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__":
    main()