examples/research_projects/multi_token_textual_inversion/textual_inversion_flax.py

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

import jax
import jax.numpy as jnp
import numpy as np
import optax
import PIL
import torch
import torch.utils.checkpoint
import transformers
from flax import jax_utils
from flax.training import train_state
from flax.training.common_utils import shard
from huggingface_hub import create_repo, upload_folder

# 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 CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel, set_seed

from diffusers import (
    FlaxAutoencoderKL,
    FlaxDDPMScheduler,
    FlaxPNDMScheduler,
    FlaxStableDiffusionPipeline,
    FlaxUNet2DConditionModel,
)
from diffusers.pipelines.stable_diffusion import FlaxStableDiffusionSafetyChecker
from diffusers.utils import check_min_version


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("0.14.0.dev0")

logger = logging.getLogger(__name__)


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    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(
        "--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=42, 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(
        "--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_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
    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("--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(
        "--use_auth_token",
        action="store_true",
        help=(
            "Will use the token generated when running `huggingface-cli login` (necessary to use this script with"
            " private models)."
        ),
    )
    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("--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 in the style of {}",
    "a rendering in the style of {}",
    "a cropped painting in the style of {}",
    "the painting in the style of {}",
    "a clean painting in the style of {}",
    "a dirty painting in the style of {}",
    "a dark painting in the style of {}",
    "a picture in the style of {}",
    "a cool painting in the style of {}",
    "a close-up painting in the style of {}",
    "a bright painting in the style of {}",
    "a cropped painting in the style of {}",
    "a good painting in the style of {}",
    "a close-up painting in the style of {}",
    "a rendition in the style of {}",
    "a nice painting in the style of {}",
    "a small painting in the style of {}",
    "a weird painting in the style of {}",
    "a large painting in the style of {}",
]


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 not 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 resize_token_embeddings(model, new_num_tokens, initializer_token_id, placeholder_token_id, rng):
    if model.config.vocab_size == new_num_tokens or new_num_tokens is None:
        return
    model.config.vocab_size = new_num_tokens

    params = model.params
    old_embeddings = params["text_model"]["embeddings"]["token_embedding"]["embedding"]
    old_num_tokens, emb_dim = old_embeddings.shape

    initializer = jax.nn.initializers.normal()

    new_embeddings = initializer(rng, (new_num_tokens, emb_dim))
    new_embeddings = new_embeddings.at[:old_num_tokens].set(old_embeddings)
    new_embeddings = new_embeddings.at[placeholder_token_id].set(new_embeddings[initializer_token_id])
    params["text_model"]["embeddings"]["token_embedding"]["embedding"] = new_embeddings

    model.params = params
    return model


def get_params_to_save(params):
    return jax.device_get(jax.tree_util.tree_map(lambda x: x[0], params))


def main():
    args = parse_args()

    if args.seed is not None:
        set_seed(args.seed)

    if jax.process_index() == 0:
        if args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)

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

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    # Setup logging, we only want one process per machine to log things on the screen.
    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
    if jax.process_index() == 0:
        transformers.utils.logging.set_verbosity_info()
    else:
        transformers.utils.logging.set_verbosity_error()

    # 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 = FlaxCLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder")
    vae, vae_params = FlaxAutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
    unet, unet_params = FlaxUNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")

    # Create sampling rng
    rng = jax.random.PRNGKey(args.seed)
    rng, _ = jax.random.split(rng)
    # Resize the token embeddings as we are adding new special tokens to the tokenizer
    text_encoder = resize_token_embeddings(
        text_encoder, len(tokenizer), initializer_token_id, placeholder_token_id, rng
    )
    original_token_embeds = text_encoder.params["text_model"]["embeddings"]["token_embedding"]["embedding"]

    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",
    )

    def collate_fn(examples):
        pixel_values = torch.stack([example["pixel_values"] for example in examples])
        input_ids = torch.stack([example["input_ids"] for example in examples])

        batch = {"pixel_values": pixel_values, "input_ids": input_ids}
        batch = {k: v.numpy() for k, v in batch.items()}

        return batch

    total_train_batch_size = args.train_batch_size * jax.local_device_count()
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset, batch_size=total_train_batch_size, shuffle=True, drop_last=True, collate_fn=collate_fn
    )

    # Optimization
    if args.scale_lr:
        args.learning_rate = args.learning_rate * total_train_batch_size

    constant_scheduler = optax.constant_schedule(args.learning_rate)

    optimizer = optax.adamw(
        learning_rate=constant_scheduler,
        b1=args.adam_beta1,
        b2=args.adam_beta2,
        eps=args.adam_epsilon,
        weight_decay=args.adam_weight_decay,
    )

    def create_mask(params, label_fn):
        def _map(params, mask, label_fn):
            for k in params:
                if label_fn(k):
                    mask[k] = "token_embedding"
                else:
                    if isinstance(params[k], dict):
                        mask[k] = {}
                        _map(params[k], mask[k], label_fn)
                    else:
                        mask[k] = "zero"

        mask = {}
        _map(params, mask, label_fn)
        return mask

    def zero_grads():
        # from https://github.com/deepmind/optax/issues/159#issuecomment-896459491
        def init_fn(_):
            return ()

        def update_fn(updates, state, params=None):
            return jax.tree_util.tree_map(jnp.zeros_like, updates), ()

        return optax.GradientTransformation(init_fn, update_fn)

    # Zero out gradients of layers other than the token embedding layer
    tx = optax.multi_transform(
        {"token_embedding": optimizer, "zero": zero_grads()},
        create_mask(text_encoder.params, lambda s: s == "token_embedding"),
    )

    state = train_state.TrainState.create(apply_fn=text_encoder.__call__, params=text_encoder.params, tx=tx)

    noise_scheduler = FlaxDDPMScheduler(
        beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000
    )
    noise_scheduler_state = noise_scheduler.create_state()

    # Initialize our training
    train_rngs = jax.random.split(rng, jax.local_device_count())

    # Define gradient train step fn
    def train_step(state, vae_params, unet_params, batch, train_rng):
        dropout_rng, sample_rng, new_train_rng = jax.random.split(train_rng, 3)

        def compute_loss(params):
            vae_outputs = vae.apply(
                {"params": vae_params}, batch["pixel_values"], deterministic=True, method=vae.encode
            )
            latents = vae_outputs.latent_dist.sample(sample_rng)
            # (NHWC) -> (NCHW)
            latents = jnp.transpose(latents, (0, 3, 1, 2))
            latents = latents * vae.config.scaling_factor

            noise_rng, timestep_rng = jax.random.split(sample_rng)
            noise = jax.random.normal(noise_rng, latents.shape)
            bsz = latents.shape[0]
            timesteps = jax.random.randint(
                timestep_rng,
                (bsz,),
                0,
                noise_scheduler.config.num_train_timesteps,
            )
            noisy_latents = noise_scheduler.add_noise(noise_scheduler_state, latents, noise, timesteps)
            encoder_hidden_states = state.apply_fn(
                batch["input_ids"], params=params, dropout_rng=dropout_rng, train=True
            )[0]
            # Predict the noise residual and compute loss
            model_pred = unet.apply(
                {"params": unet_params}, noisy_latents, timesteps, encoder_hidden_states, train=False
            ).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(noise_scheduler_state, latents, noise, timesteps)
            else:
                raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

            loss = (target - model_pred) ** 2
            loss = loss.mean()

            return loss

        grad_fn = jax.value_and_grad(compute_loss)
        loss, grad = grad_fn(state.params)
        grad = jax.lax.pmean(grad, "batch")
        new_state = state.apply_gradients(grads=grad)

        # Keep the token embeddings fixed except the newly added embeddings for the concept,
        # as we only want to optimize the concept embeddings
        token_embeds = original_token_embeds.at[placeholder_token_id].set(
            new_state.params["text_model"]["embeddings"]["token_embedding"]["embedding"][placeholder_token_id]
        )
        new_state.params["text_model"]["embeddings"]["token_embedding"]["embedding"] = token_embeds

        metrics = {"loss": loss}
        metrics = jax.lax.pmean(metrics, axis_name="batch")
        return new_state, metrics, new_train_rng

    # Create parallel version of the train and eval step
    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))

    # Replicate the train state on each device
    state = jax_utils.replicate(state)
    vae_params = jax_utils.replicate(vae_params)
    unet_params = jax_utils.replicate(unet_params)

    # Train!
    num_update_steps_per_epoch = math.ceil(len(train_dataloader))

    # Scheduler and math around the number of training steps.
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch

    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    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) = {total_train_batch_size}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")

    global_step = 0

    epochs = tqdm(range(args.num_train_epochs), desc=f"Epoch ... (1/{args.num_train_epochs})", position=0)
    for epoch in epochs:
        # ======================== Training ================================

        train_metrics = []

        steps_per_epoch = len(train_dataset) // total_train_batch_size
        train_step_progress_bar = tqdm(total=steps_per_epoch, desc="Training...", position=1, leave=False)
        # train
        for batch in train_dataloader:
            batch = shard(batch)
            state, train_metric, train_rngs = p_train_step(state, vae_params, unet_params, batch, train_rngs)
            train_metrics.append(train_metric)

            train_step_progress_bar.update(1)
            global_step += 1

            if global_step >= args.max_train_steps:
                break

        train_metric = jax_utils.unreplicate(train_metric)

        train_step_progress_bar.close()
        epochs.write(f"Epoch... ({epoch + 1}/{args.num_train_epochs} | Loss: {train_metric['loss']})")

    # Create the pipeline using using the trained modules and save it.
    if jax.process_index() == 0:
        scheduler = FlaxPNDMScheduler(
            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True
        )
        safety_checker = FlaxStableDiffusionSafetyChecker.from_pretrained(
            "CompVis/stable-diffusion-safety-checker", from_pt=True
        )
        pipeline = FlaxStableDiffusionPipeline(
            text_encoder=text_encoder,
            vae=vae,
            unet=unet,
            tokenizer=tokenizer,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32"),
        )

        pipeline.save_pretrained(
            args.output_dir,
            params={
                "text_encoder": get_params_to_save(state.params),
                "vae": get_params_to_save(vae_params),
                "unet": get_params_to_save(unet_params),
                "safety_checker": safety_checker.params,
            },
        )

        # Also save the newly trained embeddings
        learned_embeds = get_params_to_save(state.params)["text_model"]["embeddings"]["token_embedding"]["embedding"][
            placeholder_token_id
        ]
        learned_embeds_dict = {args.placeholder_token: learned_embeds}
        jnp.save(os.path.join(args.output_dir, "learned_embeds.npy"), learned_embeds_dict)

        if args.push_to_hub:
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )


if __name__ == "__main__":
    main()