train.py

"""
Donut
Copyright (c) 2022-present NAVER Corp.
MIT License
"""
import argparse
import datetime
import json
import os
import random
from io import BytesIO
from os.path import basename
from pathlib import Path

import numpy as np
import torch
import pytorch_lightning as pl
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint

from pytorch_lightning.loggers import MLFlowLogger
from pytorch_lightning.plugins import CheckpointIO
from pytorch_lightning.utilities import rank_zero_only
from sconf import Config

from donut import DonutDataset
from lightning_module import DonutDataPLModule, DonutModelPLModule

import mlflow
from mlflow import MlflowClient

os.makedirs("./outputs", exist_ok=True)

class CustomCheckpointIO(CheckpointIO):
    def save_checkpoint(self, checkpoint, path, storage_options=None):
        del checkpoint["state_dict"]
        torch.save(checkpoint, path)

    def load_checkpoint(self, path, storage_options=None):
        checkpoint = torch.load(path + "artifacts.ckpt")
        state_dict = torch.load(path + "pytorch_model.bin")
        checkpoint["state_dict"] = {"model." + key: value for key, value in state_dict.items()}
        return checkpoint

    def remove_checkpoint(self, path) -> None:
        return super().remove_checkpoint(path)


@rank_zero_only
def save_config_file(config, path):
    if not Path(path).exists():
        os.makedirs(path)
    save_path = Path(path) / "config.yaml"
    print(config.dumps())
    with open(save_path, "w") as f:
        f.write(config.dumps(modified_color=None, quote_str=True))
        print(f"Config is saved at {save_path}")

@rank_zero_only
def mlfow_log(model, path):
    mlflow.pytorch.log_model(model, path)

def train(config):
 
    pl.utilities.seed.seed_everything(config.get("seed", 42), workers=True)

    model_module = DonutModelPLModule(config)
    data_module = DonutDataPLModule(config)

    # add datasets to data_module
    datasets = {"train": [], "validation": []}
    path_list = [config.dataset_name_or_paths]
    for i, dataset_name_or_path in enumerate(path_list):
        task_name = os.path.basename(dataset_name_or_path)  # e.g., cord-v2, docvqa, rvlcdip, ...
        # add <-1> for None 
        model_module.model.decoder.add_special_tokens(["<-1/>"]) 
            
        for split in ["train", "validation"]:
            datasets[split].append(
                DonutDataset(
                    dataset_name_or_path=dataset_name_or_path,
                    donut_model=model_module.model,
                    max_length=config.max_length,
                    split=split,
                    task_start_token=config.task_start_tokens[i]
                    if config.get("task_start_tokens", None)
                    else f"<s_{task_name}>",
                    prompt_end_token= f"<s_{task_name}>",
                    sort_json_key=config.sort_json_key,
                )
            )
    data_module.train_datasets = datasets["train"]
    data_module.val_datasets = datasets["validation"]

    mlf_logger = MLFlowLogger(experiment_name=config.exp_name)

    lr_callback = LearningRateMonitor(logging_interval="step")

    checkpoint_callback = ModelCheckpoint(
        monitor="val_metric",
        dirpath=Path(config.result_path) / config.exp_name / config.exp_version,
        filename="artifacts",
        save_top_k=3,
        save_last=False,
        mode="min",
    )

    custom_ckpt = CustomCheckpointIO()
    trainer = pl.Trainer(
        resume_from_checkpoint=config.get("resume_from_checkpoint_path", None),
        num_nodes=config.get("num_nodes", 1),
        devices=1,
        #strategy= pl.strategies.DDPStrategy(timeout=datetime.timedelta(seconds=3600)),
        strategy="ddp",
        accelerator="gpu",
        plugins=custom_ckpt,
        max_epochs=config.max_epochs,
        max_steps=config.max_steps,
        val_check_interval=config.val_check_interval,
        check_val_every_n_epoch=config.check_val_every_n_epoch,
        gradient_clip_val=config.gradient_clip_val,
        precision=16,
        num_sanity_val_steps=0,
        logger=mlf_logger,
        callbacks=[lr_callback, checkpoint_callback],
    )
    with mlflow.start_run(run_id=mlf_logger.run_id) as run:
        trainer.fit(model_module, data_module)
    mlfow_log(model_module.model, "mlfmodel")
    
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True)
    parser.add_argument("--exp_version", type=str, required=False)
    parser.add_argument("--exp_name", type=str, required=False)
    parser.add_argument("--tracking_uri", type=str, required=False)
  
    args, left_argv = parser.parse_known_args()

    config = Config(args.config)
    config.argv_update(left_argv)

    config.exp_name = "donut_distributed_1" if not args.exp_name else args.exp_name # basename(args.config).split(".")[0]
    config.exp_version = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") if not args.exp_version else args.exp_version
    config.tracking_uri = "" if not args.tracking_uri else args.tracking_uri
    save_config_file(config, Path(config.result_path) / config.exp_name / config.exp_version / config.tracking_uri)
    train(config)