donut/train_stream.py

from transformers import VisionEncoderDecoderConfig, DonutProcessor, VisionEncoderDecoderModel
import torch
from torch.utils.data import DataLoader
from pytorch_lightning.loggers import WandbLogger
from pytorch_lightning.callbacks import ModelCheckpoint
import pytorch_lightning as pl
import os
from huggingface_hub import login
import argparse
from sconf import Config
from utils.checkpoint import CustomCheckpointIO
from utils.donut_dataset_stream import DonutDataset
from utils.donut_model_pl import DonutModelPLModule
from utils.callbacks import PushToHubCallback
import warnings
from datasets import load_dataset




def main(config, hug_token):

    config_vision = VisionEncoderDecoderConfig.from_pretrained(
        config.pretrained_model_path)
    config_vision.encoder.image_size = config.image_size
    config_vision.decoder.max_length = config.max_length

    processor = DonutProcessor.from_pretrained(config.start_model_path)
    model = VisionEncoderDecoderModel.from_pretrained(
        config.pretrained_model_path, config=config_vision)

    processor.image_processor.size = config.image_size[::-1]
    processor.image_processor.do_align_long_axis = False

    added_tokens = []

    train_dataset = DonutDataset(
                        config.dataset_path, 
                        processor=processor, 
                        model=model, 
                        max_length=config.max_length,
                        split="train", 
                        task_start_token="<s_cord-v2>", 
                        prompt_end_token="<s_cord-v2>",
                        added_tokens=added_tokens,
                        sort_json_key=False,  # cord dataset is preprocessed, so no need for this
                    )

    val_dataset = DonutDataset(
                        config.dataset_path, 
                        processor=processor, 
                        model=model, 
                        max_length=config.max_length,
                        split="validation", 
                        task_start_token="<s_cord-v2>", 
                        prompt_end_token="<s_cord-v2>",
                        added_tokens=added_tokens,
                        sort_json_key=False,  # cord dataset is preprocessed, so no need for this
                    )

    model.config.pad_token_id = processor.tokenizer.pad_token_id
    model.config.decoder_start_token_id = processor.tokenizer.convert_tokens_to_ids(['<s_cord-v2>'])[0]

    train_dataloader = DataLoader(train_dataset, batch_size=1, shuffle=True, num_workers=4)
    val_dataloader = DataLoader(val_dataset, batch_size=1, shuffle=False, num_workers=4)

    login(hug_token, True)

    model_module = DonutModelPLModule(config.train_config.toDict(), processor, model, max_length=config.max_length, train_dataloader=train_dataloader, val_dataloader=val_dataloader)
    
    wandb_logger = WandbLogger(project="Donut", name=config.wandb_test_name)

    checkpoint_callback = ModelCheckpoint(
        monitor="val_metric",
        dirpath=config.checkpoint_path,
        filename="artifacts",
        save_top_k=1,
        save_last=False,
        mode="min",
    )

    custom_ckpt = CustomCheckpointIO()

    trainer = pl.Trainer(
        accelerator="gpu" if torch.cuda.is_available() else 'cpu', # change to gpu
        devices=1,
        max_epochs=config.train_config.max_epochs,
        val_check_interval=config.train_config.val_check_interval,
        check_val_every_n_epoch=config.train_config.check_val_every_n_epoch,
        gradient_clip_val=config.train_config.gradient_clip_val,
        precision=16, # we'll use mixed precision
        plugins=custom_ckpt,
        num_sanity_val_steps=0,
        logger=wandb_logger,
        callbacks=[PushToHubCallback(output_model_path=config.output_model_path), checkpoint_callback],
    )

    trainer.fit(model_module)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True)
    args, left_argv = parser.parse_known_args()
    config = Config(args.config)
    config.argv_update(left_argv)

    hug_token = os.environ.get("HUG_TOKEN", None)

    if not torch.cuda.is_available():
        warnings.warn("You don't have cuda available, training might be taking long time or impossible")

    if not hug_token:
        raise Exception("You need to set up HUG_TOKEN in enviroments to push output model to hub")
    main(config, hug_token)