fine_tune.py

# Copyright 2023 Natural Synthetics Inc.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import math
import os
import traceback
from pathlib import Path
import time
import torch
import torch.utils.checkpoint
import torch.multiprocessing as mp
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from diffusers import AutoencoderKL
from diffusers.optimization import get_scheduler
from diffusers import DDPMScheduler
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
import torch.nn.functional as F
import gc
from typing import Callable
from PIL import Image
import numpy as np
from concurrent.futures import ThreadPoolExecutor
from hotshot_xl.models.unet import UNet3DConditionModel
from hotshot_xl.pipelines.hotshot_xl_pipeline import HotshotXLPipeline
from hotshot_xl.utils import get_crop_coordinates, res_to_aspect_map, scale_aspect_fill
from einops import rearrange
from torch.utils.data import Dataset, DataLoader
from datetime import timedelta
from accelerate.utils.dataclasses import InitProcessGroupKwargs
from diffusers.utils import is_wandb_available

if is_wandb_available():
    import wandb

logger = get_logger(__file__)


class HotshotXLDataset(Dataset):

    def __init__(self, directory: str, make_sample_fn: Callable):
        """

        Training data folder needs to look like:
        + training_samples
        --- + sample_001
        ------- + frame_0.jpg
        ------- + frame_1.jpg
        ------- + ...
        ------- + frame_n.jpg
        ------- + prompt.txt
        --- + sample_002
        ------- + frame_0.jpg
        ------- + frame_1.jpg
        ------- + ...
        ------- + frame_n.jpg
        ------- + prompt.txt

        Args:
            directory: base directory of the training samples
            make_sample_fn: a delegate call to load the images and prep the sample for batching
        """
        samples_dir = [os.path.join(directory, p) for p in os.listdir(directory)]
        samples_dir = [p for p in samples_dir if os.path.isdir(p)]
        samples = []

        for d in samples_dir:
            file_paths = [os.path.join(d, p) for p in os.listdir(d)]
            image_fps = [f for f in file_paths if os.path.splitext(f)[1] in {".png", ".jpg"}]
            with open(os.path.join(d, "prompt.txt")) as f:
                prompt = f.read().strip()

            samples.append({
                "image_fps": image_fps,
                "prompt": prompt
            })

        self.samples = samples
        self.length = len(samples)
        self.make_sample_fn = make_sample_fn

    def __len__(self):
        return self.length

    def __getitem__(self, index):
        return self.make_sample_fn(
            self.samples[index]
        )


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default="hotshotco/Hotshot-XL",
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--unet_resume_path",
        type=str,
        default=None,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )

    parser.add_argument(
        "--data_dir",
        type=str,
        required=True,
        help="Path to data to train.",
    )

    parser.add_argument(
        "--report_to",
        type=str,
        default="wandb",
        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("--run_validation_at_start", action="store_true")
    parser.add_argument("--max_vae_encode", type=int, default=None)
    parser.add_argument("--vae_b16", action="store_true")
    parser.add_argument("--disable_optimizer_restore", action="store_true")

    parser.add_argument(
        "--latent_nan_checking",
        action="store_true",
        help="Check if latents contain nans - important if vae is f16",
    )
    parser.add_argument(
        "--test_prompts",
        type=str,
        default=None,
    )
    parser.add_argument(
        "--project_name",
        type=str,
        default="fine-tune-hotshot-xl",
        help="the name of the run",
    )
    parser.add_argument(
        "--run_name",
        type=str,
        default="run-01",
        help="the name of the run",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="output",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument("--noise_offset", type=float, default=0.05, help="The scale of noise offset.")
    parser.add_argument("--seed", type=int, default=111, 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(
        "--aspect_ratio",
        type=str,
        default="1.75",
        choices=list(res_to_aspect_map[512].keys()),
        help="Aspect ratio to train at",
    )

    parser.add_argument("--xformers", action="store_true")

    parser.add_argument(
        "--train_batch_size", type=int, default=8, 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=9999999,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )

    parser.add_argument(
        "--learning_rate",
        type=float,
        default=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(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )

    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")

    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )

    parser.add_argument(
        "--mixed_precision",
        type=str,
        default="no",
        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."
        ),
    )

    parser.add_argument(
        "--validate_every_steps",
        type=int,
        default=100,
        help="Run inference every",
    )

    parser.add_argument(
        "--save_n_steps",
        type=int,
        default=100,
        help="Save the model every n global_steps",
    )

    parser.add_argument(
        "--save_starting_step",
        type=int,
        default=100,
        help="The step from which it starts saving intermediary checkpoints",
    )

    parser.add_argument(
        "--nccl_timeout",
        type=int,
        help="nccl_timeout",
        default=3600
    )

    parser.add_argument("--snr_gamma", action="store_true")

    args = parser.parse_args()

    return args


def add_time_ids(
        unet_config,
        unet_add_embedding,
        text_encoder_2: CLIPTextModelWithProjection,
        original_size: tuple,
        crops_coords_top_left: tuple,
        target_size: tuple,
        dtype: torch.dtype):
    add_time_ids = list(original_size + crops_coords_top_left + target_size)

    passed_add_embed_dim = (
            unet_config.addition_time_embed_dim * len(add_time_ids) + text_encoder_2.config.projection_dim
    )
    expected_add_embed_dim = unet_add_embedding.linear_1.in_features

    if expected_add_embed_dim != passed_add_embed_dim:
        raise ValueError(
            f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
        )

    add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
    return add_time_ids


def main():
    global_step = 0
    min_steps_before_validation = 0

    args = parse_args()

    next_save_iter = args.save_starting_step

    if args.save_starting_step < 1:
        next_save_iter = None

    if args.report_to == "wandb":
        if not is_wandb_available():
            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        kwargs_handlers=[InitProcessGroupKwargs(timeout=timedelta(args.nccl_timeout))]
    )

    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
        nonlocal global_step

        for model in models:
            if isinstance(model, type(accelerator.unwrap_model(unet))):
                model.save_pretrained(os.path.join(output_dir, 'unet'))
                # make sure to pop weight so that corresponding model is not saved again
                weights.pop()

    accelerator.register_save_state_pre_hook(save_model_hook)

    set_seed(args.seed)

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

    # Load the tokenizer
    tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer")
    tokenizer_2 = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer_2")

    # Load models and create wrapper for stable diffusion
    text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(args.pretrained_model_name_or_path,
                                                                 subfolder="text_encoder_2")

    vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")

    optimizer_resume_path = None

    if args.unet_resume_path:
        optimizer_fp = os.path.join(args.unet_resume_path, "optimizer.bin")

        if os.path.exists(optimizer_fp):
            optimizer_resume_path = optimizer_fp

        unet = UNet3DConditionModel.from_pretrained(args.unet_resume_path,
                                                    subfolder="unet",
                                                    low_cpu_mem_usage=False,
                                                    device_map=None)

    else:
        unet = UNet3DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")

    if args.xformers:
        vae.set_use_memory_efficient_attention_xformers(True, None)
        unet.set_use_memory_efficient_attention_xformers(True, None)

    unet_config = unet.config
    unet_add_embedding = unet.add_embedding

    unet.requires_grad_(False)

    temporal_params = unet.temporal_parameters()

    for p in temporal_params:
        p.requires_grad_(True)

    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)
    text_encoder_2.requires_grad_(False)

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    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
    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

    learning_rate = args.learning_rate

    params_to_optimize = [
        {'params': temporal_params, "lr": learning_rate},
    ]

    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,
    )

    if optimizer_resume_path and not args.disable_optimizer_restore:
        logger.info("Restoring the optimizer.")
        try:

            old_optimizer_state_dict = torch.load(optimizer_resume_path)

            # Extract only the state
            old_state = old_optimizer_state_dict['state']

            # Set the state of the new optimizer
            optimizer.load_state_dict({'state': old_state, 'param_groups': optimizer.param_groups})

            del old_optimizer_state_dict
            del old_state

            torch.cuda.empty_cache()
            torch.cuda.synchronize()
            gc.collect()

            logger.info(f"Restored the optimizer ok")

        except:
            logger.error("Failed to restore the optimizer...", exc_info=True)
            traceback.print_exc()
            raise

    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")

    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

    device = torch.device('cuda')

    image_transforms = transforms.Compose(
        [
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    def image_to_tensor(img):
        with torch.no_grad():

            if img.mode != "RGB":
                img = img.convert("RGB")

            image = image_transforms(img).to(accelerator.device)

            if image.shape[0] == 1:
                image = image.repeat(3, 1, 1)

            if image.shape[0] > 3:
                image = image[:3, :, :]

        return image

    def make_sample(sample):

        nonlocal unet_config
        nonlocal unet_add_embedding

        images = [Image.open(img) for img in sample['image_fps']]

        og_size = images[0].size

        for i, im in enumerate(images):
            if im.mode != "RGB":
                images[i] = im.convert("RGB")

        aspect_ratio_map = res_to_aspect_map[args.resolution]

        required_size = tuple(aspect_ratio_map[args.aspect_ratio])

        if required_size != og_size:

            def resize_image(x):
                img_size = x.size
                if img_size == required_size:
                    return x.resize(required_size, Image.LANCZOS)

                return scale_aspect_fill(x, required_size[0], required_size[1])

            with ThreadPoolExecutor(max_workers=len(images)) as executor:
                images = list(executor.map(resize_image, images))

        frames = torch.stack([image_to_tensor(x) for x in images])

        l, u, *_ = get_crop_coordinates(og_size, images[0].size)
        crop_coords = (l, u)

        additional_time_ids = add_time_ids(
            unet_config,
            unet_add_embedding,
            text_encoder_2,
            og_size,
            crop_coords,
            (required_size[0], required_size[1]),
            dtype=torch.float32
        ).to(device)

        input_ids_0 = tokenizer(
            sample['prompt'],
            padding="do_not_pad",
            truncation=True,
            max_length=tokenizer.model_max_length,
        ).input_ids

        input_ids_1 = tokenizer_2(
            sample['prompt'],
            padding="do_not_pad",
            truncation=True,
            max_length=tokenizer.model_max_length,
        ).input_ids

        return {
            "frames": frames,
            "input_ids_0": input_ids_0,
            "input_ids_1": input_ids_1,
            "additional_time_ids": additional_time_ids,
        }

    def collate_fn(examples: list) -> dict:

        # Two Text encoders
        # First Text Encoder  -> Penultimate Layer
        # Second Text Encoder -> Pooled Layer

        input_ids_0 = [example['input_ids_0'] for example in examples]
        input_ids_0 = tokenizer.pad({"input_ids": input_ids_0}, padding="max_length",
                                    max_length=tokenizer.model_max_length, return_tensors="pt").input_ids

        prompt_embeds_0 = text_encoder(
            input_ids_0.to(device),
            output_hidden_states=True,
        )

        # we take penultimate embeddings from the first text encoder
        prompt_embeds_0 = prompt_embeds_0.hidden_states[-2]

        input_ids_1 = [example['input_ids_1'] for example in examples]
        input_ids_1 = tokenizer_2.pad({"input_ids": input_ids_1}, padding="max_length",
                                      max_length=tokenizer.model_max_length, return_tensors="pt").input_ids

        # We are only ALWAYS interested in the pooled output of the final text encoder
        prompt_embeds = text_encoder_2(
            input_ids_1.to(device),
            output_hidden_states=True
        )

        pooled_prompt_embeds = prompt_embeds[0]
        prompt_embeds_1 = prompt_embeds.hidden_states[-2]

        prompt_embeds = torch.concat([prompt_embeds_0, prompt_embeds_1], dim=-1)

        *_, h, w = examples[0]['frames'].shape

        return {
            "frames": torch.stack([x['frames'] for x in examples]).to(memory_format=torch.contiguous_format).float(),
            "prompt_embeds": prompt_embeds.to(memory_format=torch.contiguous_format).float(),
            "pooled_prompt_embeds": pooled_prompt_embeds,
            "additional_time_ids": torch.stack([x['additional_time_ids'] for x in examples]),
        }

    # Region - Dataloaders
    dataset = HotshotXLDataset(args.data_dir, make_sample)
    dataloader = DataLoader(dataset, args.train_batch_size, shuffle=True, collate_fn=collate_fn)

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    num_update_steps_per_epoch = math.ceil(len(dataloader) / args.gradient_accumulation_steps)

    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

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

    unet, optimizer, lr_scheduler, dataloader = accelerator.prepare(
        unet, optimizer, lr_scheduler, dataloader
    )

    def to_images(video_frames: torch.Tensor):
        import torchvision.transforms as transforms
        to_pil = transforms.ToPILImage()
        video_frames = rearrange(video_frames, "b c f w h -> b f c w h")
        bsz = video_frames.shape[0]
        images = []
        for i in range(bsz):
            video = video_frames[i]
            for j in range(video.shape[0]):
                image = to_pil(video[j])
                images.append(image)
        return images

    def to_video_frames(images: list) -> np.ndarray:
        x = np.stack([np.asarray(img) for img in images])
        return np.transpose(x, (0, 3, 1, 2))

    def run_validation(step=0, node_index=0):

        nonlocal global_step
        nonlocal accelerator

        if args.test_prompts:
            prompts = args.test_prompts.split("|")
        else:
            prompts = [
                "a woman is lifting weights in a gym",
                "a group of people are dancing at a party",
                "a teddy bear doing the front crawl"
            ]

        torch.cuda.empty_cache()
        gc.collect()

        logger.info(f"Running inference to test model at {step} steps")
        with torch.no_grad():

            pipe = HotshotXLPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                unet=accelerator.unwrap_model(unet),
                text_encoder=text_encoder,
                text_encoder_2=text_encoder_2,
                vae=vae,
            )

            videos = []

            aspect_ratio_map = res_to_aspect_map[args.resolution]
            w, h = aspect_ratio_map[args.aspect_ratio]

            for prompt in prompts:
                video = pipe(prompt,
                             width=w,
                             height=h,
                             original_size=(1920, 1080),  # todo - pass in as args?
                             target_size=(args.resolution, args.resolution),
                             num_inference_steps=30,
                             video_length=8,
                             output_type="tensor",
                             generator=torch.Generator().manual_seed(111)).videos

                videos.append(to_images(video))

            for tracker in accelerator.trackers:

                if tracker.name == "wandb":
                    tracker.log(
                        {
                            "validation": [wandb.Video(to_video_frames(video), fps=8, format='mp4') for video in
                                           videos],
                        }, step=global_step
                    )

            del pipe

        return

    # Move text_encode and vae to gpu.
    vae.to(accelerator.device, dtype=torch.bfloat16 if args.vae_b16 else torch.float32)
    text_encoder.to(accelerator.device)
    text_encoder_2.to(accelerator.device)

    # 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(dataloader) / args.gradient_accumulation_steps)
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterward 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 initialize automatically on the main process.

    if accelerator.is_main_process:
        accelerator.init_trackers(args.project_name)

    def bar(prg):
        br = '|' + '█' * prg + ' ' * (25 - prg) + '|'
        return br

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

    if accelerator.is_main_process:
        logger.info("***** Running training *****")
        logger.info(f"  Num examples = {len(dataset)}")
        logger.info(f"  Num Epochs = {args.num_train_epochs}")
        logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
        logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
        logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
        logger.info(f"  Total optimization steps = {args.max_train_steps}")

    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)

    latents_scaler = vae.config.scaling_factor

    def save_checkpoint():
        save_dir = Path(args.output_dir)
        save_dir = str(save_dir)
        save_dir = save_dir.replace(" ", "_")
        if not os.path.exists(save_dir):
            os.makedirs(save_dir, exist_ok=True)
        accelerator.save_state(save_dir)

    def save_checkpoint_and_wait():
        if accelerator.is_main_process:
            save_checkpoint()
        accelerator.wait_for_everyone()

    def save_model_and_wait():
        if accelerator.is_main_process:
            HotshotXLPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                unet=accelerator.unwrap_model(unet),
                text_encoder=text_encoder,
                text_encoder_2=text_encoder_2,
                vae=vae,
            ).save_pretrained(args.output_dir, safe_serialization=True)
        accelerator.wait_for_everyone()

    def compute_loss_from_batch(batch: dict):
        frames = batch["frames"]
        bsz, number_of_frames, c, w, h = frames.shape

        # Convert images to latent space
        with torch.no_grad():

            if args.max_vae_encode:
                latents = []

                x = rearrange(frames, "bs nf c h w -> (bs nf) c h w")

                for latent_index in range(0, x.shape[0], args.max_vae_encode):
                    sample = x[latent_index: latent_index + args.max_vae_encode]

                    latent = vae.encode(sample.to(dtype=vae.dtype)).latent_dist.sample().float()
                    if len(latent.shape) == 3:
                        latent = latent.unsqueeze(0)

                    latents.append(latent)
                    torch.cuda.empty_cache()

                latents = torch.cat(latents, dim=0)
            else:

                # convert the latents from 5d -> 4d, so we can run it though the vae encoder
                x = rearrange(frames, "bs nf c h w -> (bs nf) c h w")

                del frames

                torch.cuda.empty_cache()

                latents = vae.encode(x.to(dtype=vae.dtype)).latent_dist.sample().float()

            if args.latent_nan_checking and torch.any(torch.isnan(latents)):
                accelerator.print("NaN found in latents, replacing with zeros")
                latents = torch.where(torch.isnan(latents), torch.zeros_like(latents), latents)

            latents = rearrange(latents, "(b f) c h w -> b c f h w", b=bsz)

            torch.cuda.empty_cache()

            noise = torch.randn_like(latents, device=latents.device)

            if args.noise_offset:
                # https://www.crosslabs.org//blog/diffusion-with-offset-noise
                noise += args.noise_offset * torch.randn(
                    (latents.shape[0], latents.shape[1], 1, 1, 1), device=latents.device
                )

            # Sample a random timestep for each image
            timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
            timesteps = timesteps.long()  # .repeat_interleave(number_of_frames)
            latents = latents * latents_scaler

            # Add noise to the latents according to the noise magnitude at each timestep
            # (this is the forward diffusion process)

            prompt_embeds = batch['prompt_embeds']
            add_text_embeds = batch['pooled_prompt_embeds']

            additional_time_ids = batch['additional_time_ids']  # .repeat_interleave(number_of_frames, dim=0)

            added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": additional_time_ids}

            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

            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}")

        noisy_latents.requires_grad = True

        model_pred = unet(noisy_latents,
                          timesteps,
                          cross_attention_kwargs=None,
                          encoder_hidden_states=prompt_embeds,
                          added_cond_kwargs=added_cond_kwargs,
                          return_dict=False,
                          )[0]

        if args.snr_gamma:
            # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
            # Since we predict the noise instead of x_0, the original formulation is slightly changed.
            # This is discussed in Section 4.2 of the same paper.
            snr = compute_snr(timesteps)
            mse_loss_weights = (
                    torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
            )
            # We first calculate the original loss. Then we mean over the non-batch dimensions and
            # rebalance the sample-wise losses with their respective loss weights.
            # Finally, we take the mean of the rebalanced loss.
            loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")

            loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
            return loss.mean()
        else:
            return F.mse_loss(model_pred.float(), target.float(), reduction='mean')

    def process_batch(batch: dict):
        nonlocal global_step
        nonlocal next_save_iter

        now = time.time()

        with accelerator.accumulate(unet):

            logging_data = {}
            if global_step == 0:
                # print(f"Running initial validation at step")
                if accelerator.is_main_process and args.run_validation_at_start:
                    run_validation(step=global_step, node_index=accelerator.process_index // 8)
                accelerator.wait_for_everyone()

            loss = compute_loss_from_batch(batch)

            accelerator.backward(loss)

            if accelerator.sync_gradients:
                accelerator.clip_grad_norm_(temporal_params, args.max_grad_norm)

            optimizer.step()

            lr_scheduler.step()
            optimizer.zero_grad()

        # Checks if the accelerator has performed an optimization step behind the scenes
        if accelerator.sync_gradients:
            progress_bar.update(1)
            global_step += 1

        fll = round((global_step * 100) / args.max_train_steps)
        fll = round(fll / 4)
        pr = bar(fll)

        logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0], "loss_time": (time.time() - now)}

        if args.validate_every_steps is not None and global_step > min_steps_before_validation and global_step % args.validate_every_steps == 0:
            if accelerator.is_main_process:
                run_validation(step=global_step, node_index=accelerator.process_index // 8)

            accelerator.wait_for_everyone()

        for key, val in logging_data.items():
            logs[key] = val

        progress_bar.set_postfix(**logs)
        progress_bar.set_description_str("Progress:" + pr)
        accelerator.log(logs, step=global_step)

        if accelerator.is_main_process \
                and next_save_iter is not None \
                and global_step < args.max_train_steps \
                and global_step + 1 == next_save_iter:
            save_checkpoint()

            torch.cuda.empty_cache()
            gc.collect()

            next_save_iter += args.save_n_steps

    for epoch in range(args.num_train_epochs):
        unet.train()

        for step, batch in enumerate(dataloader):
            process_batch(batch)

            if global_step >= args.max_train_steps:
                break

        if global_step >= args.max_train_steps:
            logger.info("Max train steps reached. Breaking while loop")
            break

        accelerator.wait_for_everyone()

    save_model_and_wait()

    accelerator.end_training()


if __name__ == "__main__":
    mp.set_start_method('spawn')
    main()