path: root/training/strategy/dreambooth.py

from typing import Optional
from types import MethodType
from functools import partial
from contextlib import contextmanager, nullcontext
from pathlib import Path
import itertools

import torch
from torch.utils.data import DataLoader

from accelerate import Accelerator
from transformers import CLIPTextModel
from diffusers import AutoencoderKL, UNet2DConditionModel, DPMSolverMultistepScheduler

from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion
from models.clip.tokenizer import MultiCLIPTokenizer
from training.util import EMAModel
from training.functional import TrainingStrategy, TrainingCallbacks, save_samples


def dreambooth_strategy_callbacks(
    accelerator: Accelerator,
    unet: UNet2DConditionModel,
    text_encoder: CLIPTextModel,
    tokenizer: MultiCLIPTokenizer,
    vae: AutoencoderKL,
    sample_scheduler: DPMSolverMultistepScheduler,
    train_dataloader: DataLoader,
    val_dataloader: Optional[DataLoader],
    sample_output_dir: Path,
    checkpoint_output_dir: Path,
    seed: int,
    placeholder_tokens: list[str],
    placeholder_token_ids: list[list[int]],
    train_text_encoder_cycles: int,
    text_encoder_unfreeze_last_n_layers: int = 2,
    use_emb_decay: bool = False,
    emb_decay_target: float = 0.4,
    emb_decay: float = 1e-2,
    max_grad_norm: float = 1.0,
    use_ema: bool = False,
    ema_inv_gamma: float = 1.0,
    ema_power: int = 1,
    ema_max_decay: float = 0.9999,
    sample_batch_size: int = 1,
    sample_num_batches: int = 1,
    sample_num_steps: int = 20,
    sample_guidance_scale: float = 7.5,
    sample_image_size: Optional[int] = None,
):
    sample_output_dir.mkdir(parents=True, exist_ok=True)
    checkpoint_output_dir.mkdir(parents=True, exist_ok=True)

    weight_dtype = torch.float32
    if accelerator.state.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.state.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16

    save_samples_ = partial(
        save_samples,
        accelerator=accelerator,
        tokenizer=tokenizer,
        vae=vae,
        sample_scheduler=sample_scheduler,
        train_dataloader=train_dataloader,
        val_dataloader=val_dataloader,
        output_dir=sample_output_dir,
        seed=seed,
        batch_size=sample_batch_size,
        num_batches=sample_num_batches,
        num_steps=sample_num_steps,
        guidance_scale=sample_guidance_scale,
        image_size=sample_image_size,
    )

    if use_ema:
        ema_unet = EMAModel(
            unet.parameters(),
            inv_gamma=ema_inv_gamma,
            power=ema_power,
            max_value=ema_max_decay,
        )
        ema_unet.to(accelerator.device)
    else:
        ema_unet = None

    def ema_context():
        if ema_unet is not None:
            return ema_unet.apply_temporary(unet.parameters())
        else:
            return nullcontext()

    @contextmanager
    def on_train(cycle: int):
        unet.train()
        tokenizer.train()

        if cycle < train_text_encoder_cycles:
            text_encoder.train()

        yield

    @contextmanager
    def on_eval():
        unet.eval()
        tokenizer.eval()
        text_encoder.eval()

        with ema_context():
            yield

    def on_before_optimize(cycle: int):
        params_to_clip = [unet.parameters()]
        if cycle < train_text_encoder_cycles:
            params_to_clip.append(text_encoder.parameters())
        accelerator.clip_grad_norm_(itertools.chain(*params_to_clip), max_grad_norm)

        if len(placeholder_tokens) != 0 and use_emb_decay:
            params = [
                p
                for p in text_encoder.text_model.embeddings.parameters()
                if p.grad is not None
            ]
            return torch.stack(params) if len(params) != 0 else None

    @torch.no_grad()
    def on_after_optimize(w, lrs: dict[str, float]):
        if ema_unet is not None:
            ema_unet.step(unet.parameters())

        if w is not None and "emb" in lrs:
            lr = lrs["emb"]
            lambda_ = emb_decay * lr

            if lambda_ != 0:
                norm = w[:, :].norm(dim=-1, keepdim=True)
                w[:].add_(
                    (w[:] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm)
                )

    def on_log():
        if ema_unet is not None:
            return {"ema_decay": ema_unet.decay}
        return {}

    @torch.no_grad()
    def on_checkpoint(step, postfix):
        if postfix != "end":
            return

        print("Saving model...")

        unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=False)
        text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False)

        text_encoder_.text_model.embeddings.persist(False)

        with ema_context():
            pipeline = VlpnStableDiffusion(
                text_encoder=text_encoder_,
                vae=vae,
                unet=unet_,
                tokenizer=tokenizer,
                scheduler=sample_scheduler,
            )
            pipeline.save_pretrained(checkpoint_output_dir)

        del unet_, text_encoder_, pipeline

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

    @torch.no_grad()
    def on_sample(cycle, step):
        unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=True)
        text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True)

        orig_unet_dtype = unet_.dtype
        orig_text_encoder_dtype = text_encoder_.dtype

        unet_.to(dtype=weight_dtype)
        text_encoder_.to(dtype=weight_dtype)

        save_samples_(cycle=cycle, step=step, unet=unet_, text_encoder=text_encoder_)

        unet_.to(dtype=orig_unet_dtype)
        text_encoder_.to(dtype=orig_text_encoder_dtype)

        del unet_, text_encoder_

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

    return TrainingCallbacks(
        on_train=on_train,
        on_eval=on_eval,
        on_before_optimize=on_before_optimize,
        on_after_optimize=on_after_optimize,
        on_log=on_log,
        on_checkpoint=on_checkpoint,
        on_sample=on_sample,
    )


def dreambooth_prepare(
    accelerator: Accelerator,
    text_encoder: CLIPTextModel,
    unet: UNet2DConditionModel,
    optimizer: torch.optim.Optimizer,
    train_dataloader: DataLoader,
    val_dataloader: Optional[DataLoader],
    lr_scheduler: torch.optim.lr_scheduler._LRScheduler,
    text_encoder_unfreeze_last_n_layers: int = 2,
    **kwargs
):
    (
        text_encoder,
        unet,
        optimizer,
        train_dataloader,
        val_dataloader,
        lr_scheduler,
    ) = accelerator.prepare(
        text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler
    )

    if text_encoder_unfreeze_last_n_layers == 0:
        text_encoder.text_model.encoder.requires_grad_(False)
    elif text_encoder_unfreeze_last_n_layers > 0:
        for layer in text_encoder.text_model.encoder.layers[
            : (-1 * text_encoder_unfreeze_last_n_layers)
        ]:
            layer.requires_grad_(False)

    text_encoder.text_model.embeddings.position_embedding.requires_grad_(False)
    # text_encoder.text_model.embeddings.requires_grad_(False)

    return text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler


dreambooth_strategy = TrainingStrategy(
    callbacks=dreambooth_strategy_callbacks, prepare=dreambooth_prepare
)