train_openfold.py

import argparse
import logging
import os

#os.environ["CUDA_VISIBLE_DEVICES"] = "0"
#os.environ["MASTER_ADDR"]="10.119.81.14"
#os.environ["MASTER_PORT"]="42069"
#os.environ["NODE_RANK"]="0"

import random
import time

import numpy as np
import pytorch_lightning as pl
from pytorch_lightning.callbacks.lr_monitor import LearningRateMonitor
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
from pytorch_lightning.loggers import WandbLogger
from pytorch_lightning.plugins.training_type import DeepSpeedPlugin, DDPPlugin
from pytorch_lightning.plugins.environments import SLURMEnvironment
import torch

from openfold.config import model_config
from openfold.data.data_modules import (
    OpenFoldDataModule,
    DummyDataLoader,
)
from openfold.model.model import AlphaFold
from openfold.model.torchscript import script_preset_
from openfold.utils.callbacks import (
    EarlyStoppingVerbose,
)
from openfold.utils.exponential_moving_average import ExponentialMovingAverage
from openfold.utils.argparse import remove_arguments
from openfold.utils.loss import AlphaFoldLoss, lddt_ca
from openfold.utils.seed import seed_everything
from openfold.utils.tensor_utils import tensor_tree_map
from scripts.zero_to_fp32 import (
    get_fp32_state_dict_from_zero_checkpoint
)

from openfold.utils.logger import PerformanceLoggingCallback


class OpenFoldWrapper(pl.LightningModule):
    def __init__(self, config):
        super(OpenFoldWrapper, self).__init__()
        self.config = config
        self.model = AlphaFold(config)
        self.loss = AlphaFoldLoss(config.loss)
        self.ema = ExponentialMovingAverage(
            model=self.model, decay=config.ema.decay
        )
        
        self.cached_weights = None

    def forward(self, batch):
        return self.model(batch)

    def training_step(self, batch, batch_idx):
        if(self.ema.device != batch["aatype"].device):
            self.ema.to(batch["aatype"].device)

        # Run the model
        outputs = self(batch)
        
        # Remove the recycling dimension
        batch = tensor_tree_map(lambda t: t[..., -1], batch)

        # Compute loss
        loss = self.loss(outputs, batch)

        # Log it
        self.log("train/loss", loss, on_step=True, logger=True)

        return loss

    def on_before_zero_grad(self, *args, **kwargs):
        self.ema.update(self.model)

    def validation_step(self, batch, batch_idx):
        # At the start of validation, load the EMA weights
        if(self.cached_weights is None):
            self.cached_weights = self.model.state_dict()
            self.model.load_state_dict(self.ema.state_dict()["params"])
       
        # Calculate validation loss
        outputs = self(batch)
        batch = tensor_tree_map(lambda t: t[..., -1], batch)
        lddt_ca_score = lddt_ca(
            outputs["final_atom_positions"],
            batch["all_atom_positions"],
            batch["all_atom_mask"],
            eps=self.config.globals.eps,
            per_residue=False,
        )
        self.log("val/lddt_ca", lddt_ca_score, logger=True)

        batch["use_clamped_fape"] = 0.
        loss = self.loss(outputs, batch)
        self.log("val/loss", loss, logger=True)

    def validation_epoch_end(self, _):
        # Restore the model weights to normal
        self.model.load_state_dict(self.cached_weights)
        self.cached_weights = None

    def configure_optimizers(self, 
        learning_rate: float = 1e-3,
        eps: float = 1e-5,
    ) -> torch.optim.Adam:
        # Ignored as long as a DeepSpeed optimizer is configured
        return torch.optim.Adam(
            self.model.parameters(), 
            lr=learning_rate, 
            eps=eps
        )

    def on_load_checkpoint(self, checkpoint):
        self.ema.load_state_dict(checkpoint["ema"])

    def on_save_checkpoint(self, checkpoint):
        checkpoint["ema"] = self.ema.state_dict()


def main(args):
    if(args.seed is not None):
        seed_everything(args.seed) 

    config = model_config(
        "initial_training", 
        train=True, 
        low_prec=(args.precision == "16")
    ) 
    
    model_module = OpenFoldWrapper(config)
    if(args.resume_from_ckpt and args.resume_model_weights_only):
        sd = get_fp32_state_dict_from_zero_checkpoint(args.resume_from_ckpt)
        sd = {k[len("module."):]:v for k,v in sd.items()}
        model_module.load_state_dict(sd)
        logging.info("Successfully loaded model weights...")

    # TorchScript components of the model
    if(args.script_modules):
        script_preset_(model_module)

    #data_module = DummyDataLoader("new_batch.pickle")
    data_module = OpenFoldDataModule(
        config=config.data, 
        batch_seed=args.seed,
        **vars(args)
    )

    data_module.prepare_data()
    data_module.setup()
    
    callbacks = []
    if(args.checkpoint_every_epoch):
        mc = ModelCheckpoint(
            every_n_epochs=1,
        )
        callbacks.append(mc)

    if(args.early_stopping):
        es = EarlyStoppingVerbose(
            monitor="val/lddt_ca",
            min_delta=args.min_delta,
            patience=args.patience,
            verbose=False,
            mode="max",
            check_finite=True,
            strict=True,
        )
        callbacks.append(es)

    if(args.log_performance):
        global_batch_size = args.num_nodes * args.gpus
        perf = PerformanceLoggingCallback(
            log_file=os.path.join(args.output_dir, "performance_log.json"),
            global_batch_size=global_batch_size,
        )
        callbacks.append(perf)

    loggers = []
    if(args.wandb):
        wdb_logger = WandbLogger(
            name=args.experiment_name,
            save_dir=args.output_dir,
            id=args.wandb_id,
            project=args.wandb_project,
            **{"entity": args.wandb_entity}
        )
        loggers.append(wdb_logger)

    if(args.deepspeed_config_path is not None):
        strategy = DeepSpeedPlugin(
            config=args.deepspeed_config_path,
        )
        if(args.wandb):
            wdb_logger.experiment.save(args.deepspeed_config_path)
            wdb_logger.experiment.save("openfold/config.py")
    elif (args.gpus is not None and args.gpus > 1) or args.num_nodes > 1:
        strategy = DDPPlugin(find_unused_parameters=False)
    else:
        strategy = None
    
    trainer = pl.Trainer.from_argparse_args(
        args,
        default_root_dir=args.output_dir,
        strategy=strategy,
        callbacks=callbacks,
        logger=loggers,
    )

    if(args.resume_model_weights_only):
        ckpt_path = None
    else:
        ckpt_path = args.resume_from_ckpt

    trainer.fit(
        model_module, 
        datamodule=data_module,
        ckpt_path=ckpt_path,
    )


def bool_type(bool_str: str):
    bool_str_lower = bool_str.lower()
    if bool_str_lower in ('false', 'f', 'no', 'n', '0'):
        return False
    elif bool_str_lower in ('true', 't', 'yes', 'y', '1'):
        return True
    else:
        raise ValueError(f'Cannot interpret {bool_str} as bool')


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "train_data_dir", type=str,
        help="Directory containing training mmCIF files"
    )
    parser.add_argument(
        "train_alignment_dir", type=str,
        help="Directory containing precomputed training alignments"
    )
    parser.add_argument(
        "template_mmcif_dir", type=str,
        help="Directory containing mmCIF files to search for templates"
    )
    parser.add_argument(
        "output_dir", type=str,
        help='''Directory in which to output checkpoints, logs, etc. Ignored
                if not on rank 0'''
    )
    parser.add_argument(
        "max_template_date", type=str,
        help='''Cutoff for all templates. In training mode, templates are also 
                filtered by the release date of the target'''
    )
    parser.add_argument(
        "--distillation_data_dir", type=str, default=None,
        help="Directory containing training PDB files"
    )
    parser.add_argument(
        "--distillation_alignment_dir", type=str, default=None,
        help="Directory containing precomputed distillation alignments"
    )
    parser.add_argument(
        "--val_data_dir", type=str, default=None,
        help="Directory containing validation mmCIF files"
    )
    parser.add_argument(
        "--val_alignment_dir", type=str, default=None,
        help="Directory containing precomputed validation alignments"
    )
    parser.add_argument(
        "--kalign_binary_path", type=str, default='/usr/bin/kalign',
        help="Path to the kalign binary"
    )
    parser.add_argument(
        "--train_mapping_path", type=str, default=None,
        help='''Optional path to a .json file containing a mapping from
                consecutive numerical indices to sample names. Used to filter
                the training set'''
    )
    parser.add_argument(
        "--distillation_mapping_path", type=str, default=None,
        help="""See --train_mapping_path"""
    )
    parser.add_argument(
        "--obsolete_pdbs_file_path", type=str, default=None,
        help="""Path to obsolete.dat file containing list of obsolete PDBs and 
             their replacements."""
    )
    parser.add_argument(
        "--template_release_dates_cache_path", type=str, default=None,
        help="""Output of scripts/generate_mmcif_cache.py run on template mmCIF
                files."""
    )
    parser.add_argument(
        "--use_small_bfd", type=bool_type, default=False,
        help="Whether to use a reduced version of the BFD database"
    )
    parser.add_argument(
        "--seed", type=int, default=None,
        help="Random seed"
    )
    parser.add_argument(
        "--deepspeed_config_path", type=str, default=None,
        help="Path to DeepSpeed config. If not provided, DeepSpeed is disabled"
    )
    parser.add_argument(
        "--checkpoint_every_epoch", action="store_true", default=False,
        help="""Whether to checkpoint at the end of every training epoch"""
    )
    parser.add_argument(
        "--early_stopping", type=bool_type, default=False,
        help="Whether to stop training when validation loss fails to decrease"
    )
    parser.add_argument(
        "--min_delta", type=float, default=0,
        help="""The smallest decrease in validation loss that counts as an 
                improvement for the purposes of early stopping"""
    )
    parser.add_argument(
        "--patience", type=int, default=3,
        help="Early stopping patience"
    )
    parser.add_argument(
        "--resume_from_ckpt", type=str, default=None,
        help="Path to a model checkpoint from which to restore training state"
    )
    parser.add_argument(
        "--resume_model_weights_only", type=bool_type, default=False,
        help="Whether to load just model weights as opposed to training state"
    )
    parser.add_argument(
        "--log_performance", type=bool_type, default=False,
        help="Measure performance"
    )
    parser.add_argument(
        "--wandb", action="store_true", default=False,
    )
    parser.add_argument(
        "--experiment_name", type=str, default=None,
    )
    parser.add_argument(
        "--wandb_id", type=str, default=None,
    )
    parser.add_argument(
        "--wandb_project", type=str, default=None,
    )
    parser.add_argument(
        "--wandb_entity", type=str, default=None,
    )
    parser.add_argument(
        "--script_modules", type=bool_type, default=False,
        help="Whether to TorchScript eligible components of them model"
    )
    parser.add_argument(
        "--train_chain_data_cache_path", type=str, default=None,
    )
    parser.add_argument(
        "--distillation_chain_data_cache_path", type=str, default=None,
    )
    parser.add_argument(
        "--train_epoch_len", type=int, default=10000,
    )
    parser = pl.Trainer.add_argparse_args(parser)
   
    # Disable the initial validation pass
    parser.set_defaults(
        num_sanity_val_steps=0,
    )

    # Remove some buggy/redundant arguments introduced by the Trainer
    remove_arguments(
        parser, 
        [
            "--accelerator", 
            "--resume_from_checkpoint",
            "--reload_dataloaders_every_epoch",
            "--reload_dataloaders_every_n_epochs",
        ]
    ) 

    args = parser.parse_args()

    if(args.seed is None and 
        ((args.gpus is not None and args.gpus > 1) or 
         (args.num_nodes is not None and args.num_nodes > 1))):
        raise ValueError("For distributed training, --seed must be specified")

    # This re-applies the training-time filters at the beginning of every epoch
    args.reload_dataloaders_every_n_epochs = 1

    main(args)