textual_inversion.py

import argparse
import itertools
import math
import os
import random
from pathlib import Path
from typing import Optional

import numpy as np
import PIL
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    PNDMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker

# from diffusers.utils import check_min_version
from huggingface_hub import HfFolder, Repository, whoami

# TODO: remove and import from diffusers.utils when the new version of diffusers is released
from packaging import version
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer

if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
    PIL_INTERPOLATION = {
        "linear": PIL.Image.Resampling.BILINEAR,
        "bilinear": PIL.Image.Resampling.BILINEAR,
        "bicubic": PIL.Image.Resampling.BICUBIC,
        "lanczos": PIL.Image.Resampling.LANCZOS,
        "nearest": PIL.Image.Resampling.NEAREST,
    }
else:
    PIL_INTERPOLATION = {
        "linear": PIL.Image.LINEAR,
        "bilinear": PIL.Image.BILINEAR,
        "bicubic": PIL.Image.BICUBIC,
        "lanczos": PIL.Image.LANCZOS,
        "nearest": PIL.Image.NEAREST,
    }
# ------------------------------------------------------------------------------


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
# check_min_version("0.10.0.dev0")


logger = get_logger(__name__)


def save_progress(text_encoder, placeholder_token_id, accelerator, args, save_path):
    logger.info("Saving embeddings")
    learned_embeds = (
        accelerator.unwrap_model(text_encoder)
        .get_input_embeddings()
        .weight[placeholder_token_id]
    )
    learned_embeds_dict = {args.placeholder_token: learned_embeds.detach().cpu()}
    torch.save(learned_embeds_dict, save_path)


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--save_steps",
        type=int,
        default=500,
        help="Save learned_embeds.bin every X updates steps.",
    )
    parser.add_argument(
        "--only_save_embeds",
        action="store_true",
        default=False,
        help="Save only the embeddings for the new concept.",
    )
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--tokenizer_name",
        type=str,
        default=None,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default=None,
        required=True,
        help="A folder containing the training data.",
    )
    parser.add_argument(
        "--placeholder_token",
        type=str,
        default=None,
        required=True,
        help="A token to use as a placeholder for the concept.",
    )
    parser.add_argument(
        "--initializer_token",
        type=str,
        default=None,
        required=True,
        help="A token to use as initializer word.",
    )
    parser.add_argument(
        "--learnable_property",
        type=str,
        default="object",
        help="Choose between 'object' and 'style'",
    )
    parser.add_argument(
        "--repeats",
        type=int,
        default=100,
        help="How many times to repeat the training data.",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="text-inversion-model",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--seed", type=int, default=None, help="A seed for reproducible training."
    )
    parser.add_argument(
        "--resolution",
        type=int,
        default=512,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--center_crop",
        action="store_true",
        help="Whether to center crop images before resizing to resolution",
    )
    parser.add_argument(
        "--train_batch_size",
        type=int,
        default=16,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument("--num_train_epochs", type=int, default=100)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=5000,
        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(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=True,
        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(
        "--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(
        "--push_to_hub",
        action="store_true",
        help="Whether or not to push the model to the Hub.",
    )
    parser.add_argument(
        "--hub_token",
        type=str,
        default=None,
        help="The token to use to push to the Model Hub.",
    )
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--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(
        "--local_rank",
        type=int,
        default=-1,
        help="For distributed training: local_rank",
    )

    args = parser.parse_args()
    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    if args.train_data_dir is None:
        raise ValueError("You must specify a train data directory.")

    return args


imagenet_templates_small = [
    "a photo of a {}",
    "a rendering of a {}",
    "a cropped photo of the {}",
    "the photo of a {}",
    "a photo of a clean {}",
    "a photo of a dirty {}",
    "a dark photo of the {}",
    "a photo of my {}",
    "a photo of the cool {}",
    "a close-up photo of a {}",
    "a bright photo of the {}",
    "a cropped photo of a {}",
    "a photo of the {}",
    "a good photo of the {}",
    "a photo of one {}",
    "a close-up photo of the {}",
    "a rendition of the {}",
    "a photo of the clean {}",
    "a rendition of a {}",
    "a photo of a nice {}",
    "a good photo of a {}",
    "a photo of the nice {}",
    "a photo of the small {}",
    "a photo of the weird {}",
    "a photo of the large {}",
    "a photo of a cool {}",
    "a photo of a small {}",
]

imagenet_style_templates_small = [
    "a painting of {}, art by *",
    "a rendering of {}, art by *",
    "a cropped painting of {}, art by *",
    "the painting of {}, art by *",
    "a clean painting of {}, art by *",
    "a dirty painting of {}, art by *",
    "a dark painting of {}, art by *",
    "a picture of {}, art by *",
    "a cool painting of {}, art by *",
    "a close-up painting of {}, art by *",
    "a bright painting of {}, art by *",
    "a cropped painting of {}, art by *",
    "a good painting of {}, art by *",
    "a close-up painting of {}, art by *",
    "a rendition of {}, art by *",
    "a nice painting of {}, art by *",
    "a small painting of {}, art by *",
    "a weird painting of {}, art by *",
    "a large painting of {}, art by *",
]


class TextualInversionDataset(Dataset):
    def __init__(
        self,
        data_root,
        tokenizer,
        learnable_property="object",  # [object, style]
        size=512,
        repeats=100,
        interpolation="bicubic",
        flip_p=0.5,
        set="train",
        placeholder_token="*",
        center_crop=False,
    ):
        self.data_root = data_root
        self.tokenizer = tokenizer
        self.learnable_property = learnable_property
        self.size = size
        self.placeholder_token = placeholder_token
        self.center_crop = center_crop
        self.flip_p = flip_p

        self.image_paths = [
            os.path.join(self.data_root, file_path)
            for file_path in os.listdir(self.data_root)
        ]

        self.num_images = len(self.image_paths)
        self._length = self.num_images

        if set == "train":
            self._length = self.num_images * repeats

        self.interpolation = {
            "linear": PIL_INTERPOLATION["linear"],
            "bilinear": PIL_INTERPOLATION["bilinear"],
            "bicubic": PIL_INTERPOLATION["bicubic"],
            "lanczos": PIL_INTERPOLATION["lanczos"],
        }[interpolation]

        self.templates = (
            imagenet_style_templates_small
            if learnable_property == "style"
            else imagenet_templates_small
        )
        self.flip_transform = transforms.RandomHorizontalFlip(p=self.flip_p)

    def __len__(self):
        return self._length

    def __getitem__(self, i):
        example = {}
        image = Image.open(self.image_paths[i % self.num_images])

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

        placeholder_string = self.placeholder_token
        text = random.choice(self.templates).format(placeholder_string)

        example["input_ids"] = self.tokenizer(
            text,
            padding="max_length",
            truncation=True,
            max_length=self.tokenizer.model_max_length,
            return_tensors="pt",
        ).input_ids[0]

        # default to score-sde preprocessing
        img = np.array(image).astype(np.uint8)

        if self.center_crop:
            crop = min(img.shape[0], img.shape[1])
            h, w, = (
                img.shape[0],
                img.shape[1],
            )
            img = img[
                (h - crop) // 2 : (h + crop) // 2, (w - crop) // 2 : (w + crop) // 2
            ]

        image = Image.fromarray(img)
        image = image.resize((self.size, self.size), resample=self.interpolation)

        image = self.flip_transform(image)
        image = np.array(image).astype(np.uint8)
        image = (image / 127.5 - 1.0).astype(np.float32)

        example["pixel_values"] = torch.from_numpy(image).permute(2, 0, 1)
        return example


def get_full_repo_name(
    model_id: str, organization: Optional[str] = None, token: Optional[str] = None
):
    if token is None:
        token = HfFolder.get_token()
    if organization is None:
        username = whoami(token)["name"]
        return f"{username}/{model_id}"
    else:
        return f"{organization}/{model_id}"


def freeze_params(params):
    for param in params:
        param.requires_grad = False


def main():
    args = parse_args()
    # logging_dir = os.path.join(args.output_dir, args.logging_dir)

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
    )

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

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.push_to_hub:
            if args.hub_model_id is None:
                repo_name = get_full_repo_name(
                    Path(args.output_dir).name, token=args.hub_token
                )
            else:
                repo_name = args.hub_model_id
            repo = Repository(args.output_dir, clone_from=repo_name)

            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
                if "step_*" not in gitignore:
                    gitignore.write("step_*\n")
                if "epoch_*" not in gitignore:
                    gitignore.write("epoch_*\n")
        elif args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)

    # Load the tokenizer and add the placeholder token as a additional special token
    if args.tokenizer_name:
        tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name)
    elif args.pretrained_model_name_or_path:
        tokenizer = CLIPTokenizer.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="tokenizer"
        )

    # Add the placeholder token in tokenizer
    num_added_tokens = tokenizer.add_tokens(args.placeholder_token)
    if num_added_tokens == 0:
        raise ValueError(
            f"The tokenizer already contains the token {args.placeholder_token}. Please pass a different"
            " `placeholder_token` that is not already in the tokenizer."
        )

    # Convert the initializer_token, placeholder_token to ids
    token_ids = tokenizer.encode(args.initializer_token, add_special_tokens=False)
    # Check if initializer_token is a single token or a sequence of tokens
    if len(token_ids) > 1:
        raise ValueError("The initializer token must be a single token.")

    initializer_token_id = token_ids[0]
    placeholder_token_id = tokenizer.convert_tokens_to_ids(args.placeholder_token)

    # Load models and create wrapper for stable diffusion
    text_encoder = CLIPTextModel.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="text_encoder",
        revision=args.revision,
    )
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="vae",
        revision=args.revision,
    )
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="unet",
        revision=args.revision,
    )

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

    # Initialise the newly added placeholder token with the embeddings of the initializer token
    token_embeds = text_encoder.get_input_embeddings().weight.data
    token_embeds[placeholder_token_id] = token_embeds[initializer_token_id]

    # Freeze vae and unet
    freeze_params(vae.parameters())
    freeze_params(unet.parameters())
    # Freeze all parameters except for the token embeddings in text encoder
    params_to_freeze = itertools.chain(
        text_encoder.text_model.encoder.parameters(),
        text_encoder.text_model.final_layer_norm.parameters(),
        text_encoder.text_model.embeddings.position_embedding.parameters(),
    )
    freeze_params(params_to_freeze)

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

    # Initialize the optimizer
    optimizer = torch.optim.AdamW(
        text_encoder.get_input_embeddings().parameters(),  # only optimize the embeddings
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

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

    train_dataset = TextualInversionDataset(
        data_root=args.train_data_dir,
        tokenizer=tokenizer,
        size=args.resolution,
        placeholder_token=args.placeholder_token,
        repeats=args.repeats,
        learnable_property=args.learnable_property,
        center_crop=args.center_crop,
        set="train",
    )
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset, batch_size=args.train_batch_size, shuffle=True
    )

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    num_update_steps_per_epoch = math.ceil(
        len(train_dataloader) / args.gradient_accumulation_steps
    )
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

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

    text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        text_encoder, optimizer, train_dataloader, lr_scheduler
    )

    # Move vae and unet to device
    vae.to(accelerator.device)
    unet.to(accelerator.device)

    # Keep vae and unet in eval model as we don't train these
    vae.eval()
    unet.eval()

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(
        len(train_dataloader) / args.gradient_accumulation_steps
    )
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        accelerator.init_trackers("textual_inversion", config=vars(args))

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

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num 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
    )
    progress_bar.set_description("Steps")
    global_step = 0

    for epoch in range(args.num_train_epochs):
        text_encoder.train()
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(text_encoder):
                # Convert images to latent space
                latents = (
                    vae.encode(batch["pixel_values"]).latent_dist.sample().detach()
                )
                latents = latents * 0.18215

                # Sample noise that we'll add to the latents
                noise = torch.randn(latents.shape).to(latents.device)
                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,
                ).long()

                # Add noise to the latents according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

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

                # Predict the noise residual
                model_pred = unet(
                    noisy_latents, timesteps, encoder_hidden_states
                ).sample

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

                loss = (
                    F.mse_loss(model_pred, target, reduction="none")
                    .mean([1, 2, 3])
                    .mean()
                )
                accelerator.backward(loss)

                # Zero out the gradients for all token embeddings except the newly added
                # embeddings for the concept, as we only want to optimize the concept embeddings
                if accelerator.num_processes > 1:
                    grads = text_encoder.module.get_input_embeddings().weight.grad
                else:
                    grads = text_encoder.get_input_embeddings().weight.grad
                # Get the index for tokens that we want to zero the grads for
                index_grads_to_zero = (
                    torch.arange(len(tokenizer)) != placeholder_token_id
                )
                grads.data[index_grads_to_zero, :] = grads.data[
                    index_grads_to_zero, :
                ].fill_(0)

                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
                if global_step % args.save_steps == 0:
                    save_path = os.path.join(
                        args.output_dir, f"learned_embeds-steps-{global_step}.bin"
                    )
                    save_progress(
                        text_encoder, placeholder_token_id, accelerator, args, save_path
                    )

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

            if global_step >= args.max_train_steps:
                break

        accelerator.wait_for_everyone()

    # Create the pipeline using using the trained modules and save it.
    if accelerator.is_main_process:
        if args.push_to_hub and args.only_save_embeds:
            logger.warn(
                "Enabling full model saving because --push_to_hub=True was specified."
            )
            save_full_model = True
        else:
            save_full_model = not args.only_save_embeds
        if save_full_model:
            pipeline = StableDiffusionPipeline(
                text_encoder=accelerator.unwrap_model(text_encoder),
                vae=vae,
                unet=unet,
                tokenizer=tokenizer,
                scheduler=PNDMScheduler.from_pretrained(
                    args.pretrained_model_name_or_path, subfolder="scheduler"
                ),
                safety_checker=StableDiffusionSafetyChecker.from_pretrained(
                    "CompVis/stable-diffusion-safety-checker"
                ),
                feature_extractor=CLIPFeatureExtractor.from_pretrained(
                    "openai/clip-vit-base-patch32"
                ),
            )
            pipeline.save_pretrained(args.output_dir)
        # Save the newly trained embeddings
        save_path = os.path.join(args.output_dir, "learned_embeds.bin")
        save_progress(text_encoder, placeholder_token_id, accelerator, args, save_path)

        if args.push_to_hub:
            repo.push_to_hub(
                commit_message="End of training", blocking=False, auto_lfs_prune=True
            )

    accelerator.end_training()


if __name__ == "__main__":
    main()