train.py

import argparse
import datetime
import logging
import inspect
import math
import os
import gc
import copy

from typing import Dict, Optional, Tuple

import wandb
from moviepy.audio.io.AudioFileClip import AudioFileClip
from moviepy.video.io.VideoFileClip import VideoFileClip
from omegaconf import OmegaConf

import torchaudio
import cv2
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import diffusers
import transformers

from tqdm.auto import tqdm

from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import set_seed

from models.unet_3d_condition import UNet3DConditionModel
from diffusers.models import AutoencoderKL
from diffusers import DPMSolverMultistepScheduler, DDPMScheduler
from diffusers.optimization import get_scheduler
from diffusers.utils import check_min_version, export_to_video
from diffusers.utils.import_utils import is_xformers_available
from diffusers.models.attention_processor import AttnProcessor2_0
from diffusers.models.attention import BasicTransformerBlock

from transformers import CLIPTokenizer

from modules.pipelines.pipeline_audio_to_video import TextToVideoSDPipeline
from modules.beats.BEATs import BEATs, BEATsConfig
from modules.temp_tokens.embedder import TempEmbedder
from modules.text_encoder.modeling_clip_tempotokens import CLIPTextModel, CLIPEncoder
from utils.dataset import VideoJsonDataset, SingleVideoDataset, \
    ImageDataset, VideoFolderDataset, CachedDataset
from einops import rearrange

from utils.lora import (
    extract_lora_ups_down,
    inject_trainable_lora,
    inject_trainable_lora_extended,
    save_lora_weight,
    train_patch_pipe,
    monkeypatch_or_replace_lora,
    monkeypatch_or_replace_lora_extended
)

already_printed_trainables = False

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


def combine_video_audio(video_path, audio_path):
    video = VideoFileClip(video_path)
    audio = AudioFileClip(audio_path)

    # Set the audio of the video clip to the provided audio
    video = video.set_audio(audio)

    output_path = video_path[:-4] + '_c.mp4'

    try:
        # Write the final video with audio to the output path
        video.write_videofile(output_path, codec="libx264", audio_codec="aac")
        return output_path

    except:
        return


def cut_wav_file(input_path, output_path, start_time, end_time):
    # Load the WAV file
    audio, sr = torchaudio.load(input_path)

    # Calculate the start and end samples
    start_sample = int(start_time * sr)
    end_sample = int(end_time * sr)

    # Extract the desired portion of the audio
    extracted_audio = audio[:, start_sample:end_sample]

    # Save the extracted audio as a new WAV file
    torchaudio.save(output_path, extracted_audio, sr)


def create_logging(logging, logger, accelerator):
    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)


def accelerate_set_verbose(accelerator):
    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()


def get_train_dataset(dataset_types, train_data, tokenizer):
    train_datasets = []

    # Loop through all available datasets, get the name, then add to list of data to process.
    for DataSet in [VideoJsonDataset, SingleVideoDataset, ImageDataset, VideoFolderDataset]:
        for dataset in dataset_types:
            if dataset == DataSet.__getname__():
                train_datasets.append(DataSet(**train_data, tokenizer=tokenizer))

    if len(train_datasets) > 0:
        return train_datasets
    else:
        raise ValueError("Dataset type not found: 'json', 'single_video', 'folder', 'image'")


def extend_datasets(datasets, dataset_items, extend=False):
    biggest_data_len = max(x.__len__() for x in datasets)
    extended = []
    for dataset in datasets:
        if dataset.__len__() == 0:
            del dataset
            continue
        if dataset.__len__() < biggest_data_len:
            for item in dataset_items:
                if extend and item not in extended and hasattr(dataset, item):
                    print(f"Extending {item}")

                    value = getattr(dataset, item)
                    value *= biggest_data_len
                    value = value[:biggest_data_len]

                    setattr(dataset, item, value)

                    print(f"New {item} dataset length: {dataset.__len__()}")
                    extended.append(item)


def export_to_video(video_frames, output_video_path, fps):
    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
    h, w, _ = video_frames[0].shape
    video_writer = cv2.VideoWriter(output_video_path, fourcc, fps=fps, frameSize=(w, h))
    for i in range(len(video_frames)):
        img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR)
        video_writer.write(img)


def create_output_folders(output_dir, config, run_name='test'):
    now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
    # out_dir = os.path.join(output_dir, f"train_{now}")
    out_dir = os.path.join(output_dir, run_name)

    os.makedirs(out_dir, exist_ok=True)
    os.makedirs(f"{out_dir}/samples", exist_ok=True)
    # OmegaConf.save(config, os.path.join(out_dir, 'config.yaml'))

    return out_dir


def load_primary_models(pretrained_model_path):
    noise_scheduler = DDPMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
    tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
    text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder")
    vae = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae")
    unet = UNet3DConditionModel.from_pretrained(pretrained_model_path, subfolder="unet")
    checkpoint = torch.load(
        'models/BEATs/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt')
    cfg = BEATsConfig(checkpoint['cfg'])
    beats = BEATs(cfg)
    beats.load_state_dict(checkpoint['model'])
    beats.eval()
    beats.predictor = None

    try:
        at_embedder = TempEmbedder()
    except:
        at_embedder = TempEmbedder()

    return noise_scheduler, tokenizer, text_encoder, vae, unet, beats, at_embedder


def unet_and_text_g_c(unet, text_encoder, unet_enable, text_enable):
    unet._set_gradient_checkpointing(value=unet_enable)
    text_encoder._set_gradient_checkpointing(CLIPEncoder, value=text_enable)


def freeze_models(models_to_freeze):
    for model in models_to_freeze:
        if model is not None: model.requires_grad_(False)


def is_attn(name):
    return ('attn1' or 'attn2' == name.split('.')[-1])


def set_processors(attentions):
    for attn in attentions: attn.set_processor(AttnProcessor2_0())


def set_torch_2_attn(unet):
    optim_count = 0

    for name, module in unet.named_modules():
        if is_attn(name):
            if isinstance(module, torch.nn.ModuleList):
                for m in module:
                    if isinstance(m, BasicTransformerBlock):
                        set_processors([m.attn1, m.attn2])
                        optim_count += 1
    if optim_count > 0:
        print(f"{optim_count} Attention layers using Scaled Dot Product Attention.")


def handle_memory_attention(enable_xformers_memory_efficient_attention, enable_torch_2_attn, unet):
    try:
        is_torch_2 = hasattr(F, 'scaled_dot_product_attention')

        if enable_xformers_memory_efficient_attention and not is_torch_2:
            if is_xformers_available():
                # from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
                # unet.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
                pass
            else:
                raise ValueError("xformers is not available. Make sure it is installed correctly")

        if enable_torch_2_attn and is_torch_2:
            set_torch_2_attn(unet)
    except:
        print("Could not enable memory efficient attention for xformers or Torch 2.0.")


def inject_lora(use_lora, model, replace_modules, is_extended=False, dropout=0.0, lora_path='', r=16):
    injector = (
        inject_trainable_lora if not is_extended
        else
        inject_trainable_lora_extended
    )

    params = None
    negation = None

    if os.path.exists(lora_path):
        try:
            for f in os.listdir(lora_path):
                if f.endswith('.pt'):
                    lora_file = os.path.join(lora_path, f)

                    if 'text_encoder' in f and isinstance(model, CLIPTextModel):
                        monkeypatch_or_replace_lora(
                            model,
                            torch.load(lora_file),
                            target_replace_module=replace_modules,
                            r=r
                        )
                        print("Successfully loaded Text Encoder LoRa.")

                    if 'unet' in f and isinstance(model, UNet3DConditionModel):
                        monkeypatch_or_replace_lora_extended(
                            model,
                            torch.load(lora_file),
                            target_replace_module=replace_modules,
                            r=r
                        )
                        print("Successfully loaded UNET LoRa.")

        except Exception as e:
            print(e)
            print("Could not load LoRAs. Injecting new ones instead...")

    if use_lora:
        REPLACE_MODULES = replace_modules
        injector_args = {
            "model": model,
            "target_replace_module": REPLACE_MODULES,
            "r": r
        }
        if not is_extended: injector_args['dropout_p'] = dropout

        params, negation = injector(**injector_args)
        for _up, _down in extract_lora_ups_down(
                model,
                target_replace_module=REPLACE_MODULES):

            if all(x is not None for x in [_up, _down]):
                print(f"Lora successfully injected into {model.__class__.__name__}.")

            break

    return params, negation


def save_lora(model, name, condition, replace_modules, step, save_path):
    if condition and replace_modules is not None:
        save_path = f"{save_path}/{step}_{name}.pt"
        save_lora_weight(model, save_path, replace_modules)


def handle_lora_save(
        use_unet_lora,
        use_text_lora,
        model,
        save_path,
        checkpoint_step,
        unet_target_modules,
        text_encoder_target_modules
):
    os.makedirs(f"{save_path}/lora", exist_ok=True)
    save_path = f"{save_path}/lora/lora_{str(checkpoint_step)}"
    os.makedirs(save_path, exist_ok=True)

    save_lora(
        model.unet,
        'unet',
        use_unet_lora,
        unet_target_modules,
        checkpoint_step,
        save_path,
    )
    save_lora(
        model.text_encoder,
        'text_encoder',
        use_text_lora,
        text_encoder_target_modules,
        checkpoint_step,
        save_path
    )

    train_patch_pipe(model, use_unet_lora, use_text_lora)


def param_optim(model, condition, extra_params=None, is_lora=False, negation=None):
    return {
        "model": model,
        "condition": condition,
        'extra_params': extra_params,
        'is_lora': is_lora,
        "negation": negation
    }


def create_optim_params(name='param', params=None, lr=5e-6, extra_params=None):
    params = {
        "name": name,
        "params": params,
        "lr": lr
    }

    if extra_params is not None:
        for k, v in extra_params.items():
            params[k] = v

    return params


def negate_params(name, negation):
    # We have to do this if we are co-training with LoRA.
    # This ensures that parameter groups aren't duplicated.
    if negation is None: return False
    for n in negation:
        if n in name and 'temp' not in name:
            return True
    return False


def create_optimizer_params(model_list, lr):
    import itertools
    optimizer_params = []

    for optim in model_list:
        model, condition, extra_params, is_lora, negation = optim.values()
        # Check if we are doing LoRA training.
        if is_lora and condition:
            params = create_optim_params(
                params=itertools.chain(*model),
                extra_params=extra_params
            )
            optimizer_params.append(params)
            continue

        # If this is true, we can train it.
        if condition:
            for n, p in model.named_parameters():
                should_negate = 'lora' in n
                if should_negate: continue

                params = create_optim_params(n, p, lr, extra_params)
                optimizer_params.append(params)

    return optimizer_params


def get_optimizer(use_8bit_adam):
    if 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`"
            )

        return bnb.optim.AdamW8bit
    else:
        return torch.optim.AdamW


def is_mixed_precision(accelerator):
    weight_dtype = torch.float32

    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16

    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16

    return weight_dtype


def cast_to_gpu_and_type(model_list, accelerator, weight_dtype):
    for model in model_list:
        if model is not None: model.to(accelerator.device, dtype=weight_dtype)


def handle_cache_latents(
        should_cache,
        output_dir,
        train_dataloader,
        train_batch_size,
        vae,
        beats,
        cached_latent_dir=None
):
    # Cache latents by storing them in VRAM.
    # Speeds up training and saves memory by not encoding during the train loop.
    if not should_cache: return None
    vae.to('cuda', dtype=torch.float16)
    vae.enable_slicing()

    beats.to('cuda')
    beats.eval()

    cached_latent_dir = (
        os.path.abspath(cached_latent_dir) if cached_latent_dir is not None else None
    )

    if cached_latent_dir is None:
        cache_save_dir = f"{output_dir}/cached_latents"
        os.makedirs(cache_save_dir, exist_ok=True)

        for i, batch in enumerate(tqdm(train_dataloader, desc="Caching Latents.")):

            try:
                save_name = f"cached_{i}"
                full_out_path = f"{cache_save_dir}/{save_name}.pt"

                pixel_values = batch['pixel_values'].to('cuda', dtype=torch.float16)
                batch['pixel_values'] = tensor_to_vae_latent(pixel_values, vae)

                audio_values = batch['audio_values'].to('cuda')
                batch['audio_values'] = beats.extract_features(audio_values)[1]

                for k, v in batch.items(): batch[k] = v[0]

                torch.save(batch, full_out_path)
                del pixel_values
                del batch

                # We do this to avoid fragmentation from casting latents between devices.
                torch.cuda.empty_cache()

            except:
                continue
    else:
        cache_save_dir = cached_latent_dir

    return torch.utils.data.DataLoader(
        CachedDataset(cache_dir=cache_save_dir),
        batch_size=train_batch_size,
        shuffle=True,
        num_workers=0
    )


def handle_trainable_modules(model, trainable_modules=None, is_enabled=True, negation=None):
    global already_printed_trainables

    # This can most definitely be refactored :-)
    unfrozen_params = 0
    if trainable_modules is not None:
        for name, module in model.named_modules():
            for tm in tuple(trainable_modules):
                if tm == 'all':
                    model.requires_grad_(is_enabled)
                    unfrozen_params = len(list(model.parameters()))
                    break

                if tm in name and 'lora' not in name:
                    for m in module.parameters():
                        m.requires_grad_(is_enabled)
                        if is_enabled: unfrozen_params += 1

    if unfrozen_params > 0 and not already_printed_trainables:
        already_printed_trainables = True
        print(f"{unfrozen_params} params have been unfrozen for training.")


def tensor_to_vae_latent(t, vae):
    video_length = t.shape[1]

    t = rearrange(t, "b f c h w -> (b f) c h w")
    latents = vae.encode(t).latent_dist.sample()
    latents = rearrange(latents, "(b f) c h w -> b c f h w", f=video_length)
    latents = latents * 0.18215

    return latents


def sample_noise(latents, noise_strength, use_offset_noise):
    b, c, f, *_ = latents.shape
    noise_latents = torch.randn_like(latents, device=latents.device)
    offset_noise = None

    if use_offset_noise:
        offset_noise = torch.randn(b, c, f, 1, 1, device=latents.device)
        noise_latents = noise_latents + noise_strength * offset_noise

    return noise_latents


def should_sample(global_step, validation_steps, validation_data):
    return (global_step % validation_steps == 0) \
           and validation_data.sample_preview


def save_pipe(
        path,
        global_step,
        accelerator,
        unet,
        text_encoder,
        vae,
        output_dir,
        use_unet_lora,
        use_text_lora,
        unet_target_replace_module=None,
        text_target_replace_module=None,
        is_checkpoint=False,
):
    if is_checkpoint:
        save_path = os.path.join(output_dir, f"checkpoint-{global_step}")
        os.makedirs(save_path, exist_ok=True)
    else:
        save_path = output_dir

    # Save the dtypes so we can continue training at the same precision.
    u_dtype, t_dtype, v_dtype = unet.dtype, text_encoder.dtype, vae.dtype

    # Copy the model without creating a reference to it. This allows keeping the state of our lora training if enabled.
    unet_out = copy.deepcopy(accelerator.unwrap_model(unet, keep_fp32_wrapper=False))
    text_encoder_out = copy.deepcopy(accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False))

    pipeline = TextToVideoSDPipeline.from_pretrained(
        path,
        unet=unet_out,
        text_encoder=text_encoder_out,
        vae=vae,
    ).to(torch_dtype=torch.float16)

    handle_lora_save(
        use_unet_lora,
        use_text_lora,
        pipeline,
        output_dir,
        global_step,
        unet_target_replace_module,
        text_target_replace_module
    )

    pipeline.save_pretrained(save_path)

    if is_checkpoint:
        unet, text_encoder = accelerator.prepare(unet, text_encoder)
        models_to_cast_back = [(unet, u_dtype), (text_encoder, t_dtype), (vae, v_dtype)]
        [x[0].to(accelerator.device, dtype=x[1]) for x in models_to_cast_back]

    logger.info(f"Saved model at {save_path} on step {global_step}")

    del pipeline
    del unet_out
    del text_encoder_out
    torch.cuda.empty_cache()
    gc.collect()


def replace_prompt(prompt, token, wlist):
    for w in wlist:
        if w in prompt: return prompt.replace(w, token)
    return prompt


def split_audio_waveform(waveform, num_splits, overlap):
    # Calculate the length of each split tensor
    split_length = (waveform.shape[1] - overlap * (num_splits - 1)) // num_splits

    # Initialize a list to store the split tensors
    split_tensors = []

    # Split the audio waveform into 24 tensors
    for i in range(num_splits):
        start_idx = i * (split_length - overlap)
        end_idx = start_idx + split_length
        split_tensors.append(torch.unsqueeze(waveform[:, start_idx:end_idx], dim=1))

    return torch.cat(split_tensors)


def split_into_frames_with_overlap(embedding_tensor, num_frames, overlap_ratio):
    """
    Split a tensor into multiple frames with overlap.

    Args:
        embedding_tensor (torch.Tensor): Input tensor with shape [1, T, 128].
        num_frames (int): Number of frames to split the tensor into.
        overlap_ratio (float): Overlap ratio between adjacent frames (0 to 1).

    Returns:
        torch.Tensor: Tensor containing the split frames with shape [num_frames, frame_length, 128].
    """
    batch_size, total_length, embedding_dim = embedding_tensor.shape

    # Calculate frame length and overlap amount
    frame_length = total_length // num_frames
    overlap_length = int(frame_length * overlap_ratio)

    # Initialize the output tensor to store the split frames
    output_frames = torch.zeros(num_frames, frame_length, embedding_dim)

    # Split the tensor into frames with overlap
    for i in range(num_frames):
        start_idx = i * (frame_length - overlap_length)
        end_idx = start_idx + frame_length
        output_frames[i] = embedding_tensor[:, start_idx:end_idx, :]

    return output_frames


def main(
        pretrained_model_path: str,
        output_dir: str,
        train_data: Dict,
        validation_data: Dict,
        dataset_types: Tuple[str] = ('json'),
        validation_steps: int = 100,
        trainable_modules: Tuple[str] = ("attn1", "attn2"),
        trainable_text_modules: Tuple[str] = ("all"),
        extra_unet_params=None,
        extra_text_encoder_params=None,
        train_batch_size: int = 1,
        max_train_steps: int = 500,
        learning_rate: float = 5e-5,
        lora_learning_rate: float = 5e-5,
        embedder_learning_rate: float = 5e-5,
        scale_lr: bool = False,
        lr_scheduler: str = "constant",
        lr_warmup_steps: int = 0,
        adam_beta1: float = 0.9,
        adam_beta2: float = 0.999,
        adam_weight_decay: float = 1e-2,
        adam_epsilon: float = 1e-08,
        max_grad_norm: float = 1.0,
        gradient_accumulation_steps: int = 1,
        gradient_checkpointing: bool = False,
        text_encoder_gradient_checkpointing: bool = False,
        checkpointing_steps: int = 500,
        resume_from_checkpoint: Optional[str] = None,
        mixed_precision: Optional[str] = "fp16",
        use_8bit_adam: bool = False,
        enable_xformers_memory_efficient_attention: bool = True,
        enable_torch_2_attn: bool = False,
        seed: Optional[int] = None,
        train_text_encoder: bool = False,
        use_offset_noise: bool = False,
        offset_noise_strength: float = 0.1,
        extend_dataset: bool = False,
        cache_latents: bool = False,
        cached_latent_dir=None,
        use_unet_lora: bool = False,
        use_text_lora: bool = False,
        unet_lora_modules: Tuple[str] = ["ResnetBlock2D"],
        text_encoder_lora_modules: Tuple[str] = ["CLIPEncoderLayer"],
        lora_rank: int = 16,
        lora_path: str = '',
        run_name='test',
        **kwargs
):
    *_, config = inspect.getargvalues(inspect.currentframe())

    if run_name == 'test':
        log_with = None
    else:
        log_with = 'wandb'

    accelerator = Accelerator(
        gradient_accumulation_steps=gradient_accumulation_steps,
        mixed_precision=mixed_precision,
        log_with=log_with,
        project_dir=output_dir
    )

    # Make one log on every process with the configuration for debugging.
    create_logging(logging, logger, accelerator)

    # Initialize accelerate, transformers, and diffusers warnings
    accelerate_set_verbose(accelerator)

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

    # Handle the output folder creation
    if accelerator.is_main_process:
        output_dir = create_output_folders(output_dir, config, run_name)

    # Load scheduler, tokenizer and models.
    noise_scheduler, tokenizer, text_encoder, vae, unet, beats, at_embedder = load_primary_models(pretrained_model_path)

    num_added_tokens = tokenizer.add_tokens("<temp>")
    num_added_tokens += tokenizer.add_tokens("<local1>")
    num_added_tokens += tokenizer.add_tokens("<local2>")
    num_added_tokens += tokenizer.add_tokens("<local3>")
    num_added_tokens += tokenizer.add_tokens("<local4>")
    num_added_tokens += tokenizer.add_tokens("<class>")
    if num_added_tokens < 6:
        raise ValueError(
            f"The tokenizer already contains the token <*>. Please pass a different"
            " `placeholder_token` that is not already in the tokenizer."
        )

    # Resize the token embeddings as we are adding new special tokens to the tokenizer
    text_encoder.resize_token_embeddings(len(tokenizer))

    # Freeze any necessary models
    freeze_models([vae, text_encoder, unet, beats])

    # Enable xformers if available
    handle_memory_attention(enable_xformers_memory_efficient_attention, enable_torch_2_attn, unet)

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

    # Initialize the optimizer
    optimizer_cls = get_optimizer(use_8bit_adam)

    # Use LoRA if enabled.
    unet_lora_params, unet_negation = inject_lora(
        use_unet_lora, unet, unet_lora_modules, is_extended=True,
        r=lora_rank, lora_path=lora_path
    )

    text_encoder_lora_params, text_encoder_negation = inject_lora(
        use_text_lora, text_encoder, text_encoder_lora_modules,
        r=lora_rank, lora_path=lora_path
    )

    # Create parameters to optimize over with a condition (if "condition" is true, optimize it)
    optim_params = [
        param_optim(unet, trainable_modules is not None, extra_params=extra_unet_params, negation=unet_negation),
        param_optim(text_encoder, train_text_encoder and not use_text_lora, extra_params=extra_text_encoder_params,
                    negation=text_encoder_negation
                    ),
        param_optim(text_encoder_lora_params, use_text_lora, is_lora=True, extra_params={"lr": lora_learning_rate}),
        param_optim(unet_lora_params, use_unet_lora, is_lora=True, extra_params={"lr": lora_learning_rate}),
        param_optim(at_embedder, True, is_lora=False, extra_params={"lr": embedder_learning_rate}),
    ]

    params = create_optimizer_params(optim_params, learning_rate)

    # Create Optimizer
    optimizer = optimizer_cls(
        params,
        lr=learning_rate,
        betas=(adam_beta1, adam_beta2),
        weight_decay=adam_weight_decay,
        eps=adam_epsilon,
    )

    # Scheduler
    lr_scheduler = get_scheduler(
        lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=lr_warmup_steps * gradient_accumulation_steps,
        num_training_steps=max_train_steps * gradient_accumulation_steps,
    )

    # Get the training dataset based on types (json, single_video, image)
    train_datasets = get_train_dataset(dataset_types, train_data, tokenizer)

    # Extend datasets that are less than the greatest one. This allows for more balanced training.
    attrs = ['video_files']
    extend_datasets(train_datasets, attrs, extend=extend_dataset)

    # Process one dataset
    if len(train_datasets) == 1:
        train_dataset = train_datasets[0]

    # Process many datasets
    else:
        train_dataset = torch.utils.data.ConcatDataset(train_datasets)

        # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=train_batch_size,
        shuffle=True
    )

    # Latents caching
    cached_data_loader = handle_cache_latents(
        cache_latents,
        output_dir,
        train_dataloader,
        train_batch_size,
        vae,
        beats,
        cached_latent_dir=f'outputs/{run_name}/cached_latents/'
    )

    if cached_data_loader is not None:
        train_dataloader = cached_data_loader

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

    # Use Gradient Checkpointing if enabled.
    unet_and_text_g_c(
        unet,
        text_encoder,
        gradient_checkpointing,
        text_encoder_gradient_checkpointing
    )

    # Enable VAE slicing to save memory.
    vae.enable_slicing()

    # 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 = is_mixed_precision(accelerator)

    # Move text encoders, and VAE to GPU
    models_to_cast = [text_encoder, vae, unet, beats]
    cast_to_gpu_and_type(models_to_cast, accelerator, 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) / gradient_accumulation_steps)

    # Afterwards we recalculate our number of training epochs
    num_train_epochs = math.ceil(max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.

    if accelerator.is_main_process:
        accelerator.init_trackers("audio-to-video", config=config)

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

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

    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(global_step, max_train_steps), disable=not accelerator.is_local_main_process)
    progress_bar.set_description("Steps")
    validation_idx = 0

    def finetune_unet(batch, train_encoder=False):

        # Check if we are training the text encoder
        text_trainable = (train_text_encoder or use_text_lora)

        # Unfreeze UNET Layers
        if (global_step == 0 and trainable_modules is not None) or (global_step == 0 and use_unet_lora):
            already_printed_trainables = False
            unet.train()
            handle_trainable_modules(
                unet,
                trainable_modules,
                is_enabled=True,
                negation=unet_negation
            )

        # Convert video to latent space
        pixel_values = batch["pixel_values"].float()

        if not cache_latents:
            latents = tensor_to_vae_latent(pixel_values, vae)
        else:
            latents = pixel_values

        # Get video length
        video_length = latents.shape[2]

        # Sample noise that we'll add to the latents
        noise = sample_noise(latents, offset_noise_strength, use_offset_noise)
        bsz = latents.shape[0]

        # Sample a random timestep for each video
        timesteps = torch.randint(0, noise_scheduler.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)

        # Enable text encoder training
        if text_trainable:
            text_encoder.train()

            if use_text_lora:
                text_encoder.text_model.embeddings.requires_grad_(True)

            if global_step == 0 and train_text_encoder:
                handle_trainable_modules(
                    text_encoder,
                    trainable_modules=trainable_text_modules,
                    negation=text_encoder_negation
                )
            cast_to_gpu_and_type([text_encoder], accelerator, torch.float32)

        # Fixes gradient checkpointing training.
        # See: https://github.com/prigoyal/pytorch_memonger/blob/master/tutorial/Checkpointing_for_PyTorch_models.ipynb
        if gradient_checkpointing or text_encoder_gradient_checkpointing:
            unet.eval()
            text_encoder.eval()

        # Encode text embeddings
        token_ids = batch['prompt_ids']

        audio_values = batch['audio_values'].to('cuda')
        audio_features = beats.extract_features(audio_values)[1]

        temporal_token, local_window_1, local_window_2, local_window_3, local_window_4, audio_token = at_embedder(audio_features)

        token_ids = token_ids.repeat(video_length, 1)
        audio_token = torch.unsqueeze(audio_token, dim=1).repeat(1, video_length, 1) if audio_token is not None else None

        encoder_hidden_states = text_encoder(input_ids=token_ids,
                                             audio_token=audio_token,
                                             temp_token=temporal_token,
                                             local_windows=[local_window_1, local_window_2, local_window_3, local_window_4])[0]

        # Get the target for loss depending on the prediction type
        if noise_scheduler.prediction_type == "epsilon":
            target = noise

        elif noise_scheduler.prediction_type == "v_prediction":
            target = noise_scheduler.get_velocity(latents, noise, timesteps)

        else:
            raise ValueError(f"Unknown prediction type {noise_scheduler.prediction_type}")

        # Here we do two passes for video and text training.
        # If we are on the second iteration of the loop, get one frame.
        # This allows us to train text information only on the spatial layers.
        losses = []
        should_truncate_video = (video_length > 1 and text_trainable)

        if should_truncate_video:

            # We detach the encoder hidden states for the first pass (video frames > 1)
            # Then we make a clone of the initial state to ensure we can train it in the loop.
            detached_encoder_state = encoder_hidden_states.clone().detach()
            trainable_encoder_state = encoder_hidden_states.clone()

            for i in range(2):

                should_detach = noisy_latents.shape[2] > 1 and i == 0

                if should_truncate_video and i == 1:
                    noisy_latents = noisy_latents[:,:,1,:,:].unsqueeze(2)
                    target = target[:,:,1,:,:].unsqueeze(2)

                encoder_hidden_states = (
                    detached_encoder_state if should_detach else trainable_encoder_state
                )

                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states=encoder_hidden_states).sample
                loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")

                losses.append(loss)

                # This was most likely single frame training or a single image.
                if video_length == 1 and i == 0: break

            loss = losses[0] if len(losses) == 1 else losses[0] + losses[1]

            # regularization
            loss += 0.01 * torch.mean(torch.abs(temporal_token))

            return loss, latents

        else:
            model_pred = unet(noisy_latents, timesteps, encoder_hidden_states=encoder_hidden_states).sample
            loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
            # regularization
            loss += 0.01 * torch.mean(torch.abs(temporal_token))

            return loss, latents

    for epoch in range(first_epoch, num_train_epochs):
        train_loss = 0.0

        for step, batch in enumerate(train_dataloader):
            # Skip steps until we reach the resumed step
            resume_step = 2000  # change it
            if resume_from_checkpoint and epoch == first_epoch and step < resume_step:
                if step % gradient_accumulation_steps == 0:
                    progress_bar.update(1)
                continue

            with accelerator.accumulate(unet), accelerator.accumulate(text_encoder):

                text_prompt = batch['text_prompt'][0]

                with accelerator.autocast():
                    loss, latents = finetune_unet(batch, train_encoder=train_text_encoder)

                # Gather the losses across all processes for logging (if we use distributed training).
                avg_loss = accelerator.gather(loss.repeat(train_batch_size)).mean()
                train_loss += avg_loss.item() / max((gradient_accumulation_steps // 2), 1)

                # Backpropagate
                try:
                    accelerator.backward(loss)
                    if unet_lora_modules is not None:
                        params_to_clip = (
                            unet.parameters() if not train_text_encoder
                            else
                            list(unet.parameters()) + list(text_encoder.parameters())
                        )
                        accelerator.clip_grad_norm_(params_to_clip, max_grad_norm)

                    optimizer.step()
                    lr_scheduler.step()
                    optimizer.zero_grad(set_to_none=True)

                except Exception as e:
                    print(f"An error has occured during backpropogation! {e}")
                    continue

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train_loss": train_loss}, step=global_step)
                train_loss = 0.0

                if global_step % checkpointing_steps == 0:
                    save_pipe(
                        pretrained_model_path,
                        global_step,
                        accelerator,
                        unet,
                        text_encoder,
                        vae,
                        output_dir,
                        use_unet_lora,
                        use_text_lora,
                        unet_lora_modules,
                        text_encoder_lora_modules,
                        is_checkpoint=True
                    )
                    torch.save(at_embedder.state_dict(), f"{output_dir}/at_embedder_{global_step}.pth")

                if should_sample(global_step, validation_steps, validation_data):
                    if global_step == 1: print("Performing validation prompt.")
                    if accelerator.is_main_process:

                        audio_paths = ['EYWZz8l8yKg_110000_120000', '31O2j4aAgYU_30000_40000', '3AzacBprTzU_60000_70000', 'DKSrNxPQrbY_90000_100000']
                        audio_path = f'validation_set/{audio_paths[validation_idx % len(audio_paths)]}.wav'

                        start_audio = validation_idx % 9
                        end_audio = start_audio + 2
                        save_filename = f"{global_step}_{prompt}"
                        out_file = f"{output_dir}/samples/{save_filename}.mp4"
                        out_audio = out_file.replace('mp4', 'wav')
                        cut_wav_file(audio_path, out_audio[2:], start_audio, end_audio)

                        validation_idx += 1

                        with accelerator.autocast():
                            unet.eval()
                            beats.eval()
                            text_encoder.eval()
                            unet_and_text_g_c(unet, text_encoder, False, False)

                            waveform, sample_rate = torchaudio.load(out_audio)
                            audio = waveform.to('cuda')

                            with torch.no_grad():
                                audio_features = beats.extract_features(audio)[1]

                            temporal_token, local_window_1, local_window_2, local_window_3,\
                            local_window_4, audio_token = at_embedder(audio_features)

                            pipeline = TextToVideoSDPipeline.from_pretrained(
                                pretrained_model_path,
                                text_encoder=text_encoder,
                                tokenizer=tokenizer,
                                vae=vae,
                                unet=unet
                            )

                            diffusion_scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
                            pipeline.scheduler = diffusion_scheduler

                            with torch.no_grad():
                                video_frames = pipeline(
                                    audio_token=audio_token,
                                    temporal_token=temporal_token,
                                    local_windows=[local_window_1, local_window_2, local_window_3, local_window_4],
                                    prompt=prompt,
                                    width=validation_data.width,
                                    height=validation_data.height,
                                    num_frames=validation_data.num_frames,
                                    num_inference_steps=validation_data.num_inference_steps,
                                    guidance_scale=validation_data.guidance_scale
                                ).frames
                            export_to_video(video_frames, out_file, 12)
                            out_file = combine_video_audio(out_file, out_audio)
                            if out_file is not None:
                                accelerator.log(
                                    {
                                        "validation": [
                                            wandb.Video(out_file, caption=f'step: {global_step}', fps=train_data['fps'])
                                        ]
                                    }
                                )

                            del pipeline
                            torch.cuda.empty_cache()
                            if os.path.exists(out_audio):
                                os.remove(out_audio)

                    logger.info(f"Saved a new sample to {out_file}")

                    unet_and_text_g_c(
                        unet,
                        text_encoder,
                        gradient_checkpointing,
                        text_encoder_gradient_checkpointing
                    )
            if gradient_accumulation_steps == 1:
                logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
                accelerator.log({"training_loss": loss.detach().item()}, step=step)
                progress_bar.set_postfix(**logs)

            if global_step >= max_train_steps:
                break

    # Create the pipeline using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        save_pipe(
            pretrained_model_path,
            global_step,
            accelerator,
            unet,
            text_encoder,
            vae,
            output_dir,
            use_unet_lora,
            use_text_lora,
            unet_lora_modules,
            text_encoder_lora_modules,
            is_checkpoint=False
        )
    accelerator.end_training()


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, default="./configs/v2/vggsound.yaml")
    parser.add_argument("--run_name", type=str, default="test")

    args = parser.parse_args()
    config = OmegaConf.load(args.config)

    config['run_name'] = args.run_name
    prompt = config['train_data']['fallback_prompt']

    main(**config)