train.py

import argparse
import collections
import torch
import numpy as np
import data_loader.data_loaders as module_data
import model.loss as module_loss
import model.metric as module_metric
import model.model as module_arch
from parse_config import ConfigParser
from trainer import Trainer
# For num_experts with same settings, we don't want this to set to True.
# This is strongly discouraged because it's misleading: setting it to true does not make it reproducible acorss machine/pytorch version. In addition, it also makes training slower. Use with caution.
deterministic = False
if deterministic:
    # fix random seeds for reproducibility
    SEED = 123
    torch.manual_seed(SEED)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    np.random.seed(SEED)

def main(config):
    logger = config.get_logger('train')

    # setup data_loader instances
    data_loader = config.init_obj('data_loader', module_data)
    # valid_data_loader = data_loader.split_validation()
    valid_data_loader = getattr(module_data, config['data_loader']['type'])(
        config['data_loader']['args']['data_dir'],
        batch_size=128,
        shuffle=False,
        training=False,
        num_workers=2
    )
    # build model architecture, then print to console
    model = config.init_obj('arch', module_arch)
    logger.info(model)

    # get function handles of loss and metrics
    loss_class = getattr(module_loss, config["loss"]["type"])
    if hasattr(loss_class, "require_num_experts") and loss_class.require_num_experts:
        criterion = config.init_obj('loss', module_loss, cls_num_list=data_loader.cls_num_list, num_experts=config["arch"]["args"]["num_experts"])
    else:
        criterion = config.init_obj('loss', module_loss, cls_num_list=data_loader.cls_num_list)
    metrics = [getattr(module_metric, met) for met in config['metrics']]

    # build optimizer, learning rate scheduler. delete every lines containing lr_scheduler for disabling scheduler
    optimizer = config.init_obj('optimizer', torch.optim, model.parameters())

    if "type" in config._config["lr_scheduler"]:
        if config["lr_scheduler"]["type"] == "CustomLR":
            lr_scheduler_args = config["lr_scheduler"]["args"]
            gamma = lr_scheduler_args["gamma"] if "gamma" in lr_scheduler_args else 0.1
            print("Scheduler step1, step2, warmup_epoch, gamma:", (lr_scheduler_args["step1"], lr_scheduler_args["step2"], lr_scheduler_args["warmup_epoch"], gamma))
            def lr_lambda(epoch):
                if epoch >= lr_scheduler_args["step2"]:
                    lr = gamma * gamma
                elif epoch >= lr_scheduler_args["step1"]:
                    lr = gamma
                else:
                    lr = 1

                """Warmup"""
                warmup_epoch = lr_scheduler_args["warmup_epoch"]
                if epoch < warmup_epoch:
                    lr = lr * float(1 + epoch) / warmup_epoch
                return lr
            lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
        else:
            lr_scheduler = config.init_obj('lr_scheduler', torch.optim.lr_scheduler, optimizer)
    else:
        lr_scheduler = None

    # print(data_loader.prior)
    # print(valid_data_loader.prior)

    trainer = Trainer(model, criterion, metrics, optimizer,
                      config=config,
                      data_loader=data_loader,
                      valid_data_loader=valid_data_loader,
                      lr_scheduler=lr_scheduler)

    if 'eval_ensemble' in config._config and config._config['eval_ensemble']:
        trainer.check_ensemble()
    else:
        trainer.train()


if __name__ == '__main__':
    args = argparse.ArgumentParser(description='PyTorch Template')
    args.add_argument('-c', '--config', default=None, type=str,
                      help='config file path (default: None)')
    args.add_argument('-r', '--resume', default=None, type=str,
                      help='path to latest checkpoint (default: None)')
    args.add_argument('-d', '--device', default=None, type=str,
                      help='indices of GPUs to enable (default: all)')

    # custom cli options to modify configuration from default values given in json file.
    CustomArgs = collections.namedtuple('CustomArgs', 'flags type target')
    options = [
        CustomArgs(['--lr', '--learning_rate'], type=float, target='optimizer;args;lr'),
        CustomArgs(['--bs', '--batch_size'], type=int, target='data_loader;args;batch_size'),
        CustomArgs(['--name'], type=str, target='name'),
        CustomArgs(['--epochs'], type=int, target='trainer;epochs'),
        CustomArgs(['--step1'], type=int, target='lr_scheduler;args;step1'),
        CustomArgs(['--step2'], type=int, target='lr_scheduler;args;step2'),
        CustomArgs(['--warmup'], type=int, target='lr_scheduler;args;warmup_epoch'),
        CustomArgs(['--gamma'], type=float, target='lr_scheduler;args;gamma'),
        CustomArgs(['--save_period'], type=int, target='trainer;save_period'),
        CustomArgs(['--reduce_dimension'], type=int, target='arch;args;reduce_dimension'),
        CustomArgs(['--layer2_dimension'], type=int, target='arch;args;layer2_output_dim'),
        CustomArgs(['--layer3_dimension'], type=int, target='arch;args;layer3_output_dim'),
        CustomArgs(['--layer4_dimension'], type=int, target='arch;args;layer4_output_dim'),
        CustomArgs(['--num_experts'], type=int, target='arch;args;num_experts'),
        CustomArgs(['--distribution_aware_diversity_factor'], type=float, target='loss;args;additional_diversity_factor'),
        CustomArgs(['--pos_weight'], type=float, target='arch;args;pos_weight'),
        CustomArgs(['--collaborative_loss'], type=int, target='loss;args;collaborative_loss'),
        CustomArgs(['--distill_checkpoint'], type=str, target='distill_checkpoint')
    ]
    config = ConfigParser.from_args(args, options)
    main(config)