train_swag.py

import argparse
from os.path import join as pjoin
import torch
from dca.utils import coro_timer, mkdirp, rm
from dca.trainutils import coro_log, do_epoch, do_trainbatch, \
    do_evalbatch, SummaryWriter, check_cuda, deteministic_run, bn_update
from torch.optim import SGD
from dca.optim import schedule_midway_linear_decay, get_weightdecay
from dca.models32 import savemodel
from dca.dataloaders import SVHNInfo, get_svhn_train_loaders
from utils import SWAGMODELS, loadcheckpoint, savecheckpoint


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('arch', default='preresnet20_swag', choices=SWAGMODELS,
                        help='model architecture: ' + ' | '.join(SWAGMODELS))
    parser.add_argument('-j', '--workers', default=1, type=int, metavar='N',
                        help='number of data loading workers')
    parser.add_argument('-tb', '--tbatch', default=128, type=int,
                        metavar='N', help='train mini-batch size')
    parser.add_argument('-vb', '--vbatch', default=128, type=int,
                        metavar='N', help='eval mini-batch size')
    parser.add_argument('-sp', '--tvsplit', default=0.9, type=float,
                        metavar='RATIO',
                        help='ratio of data used for training')
    parser.add_argument('-e', '--epochs', default=300, type=int, metavar='N',
                        help='number of total epochs to run')
    parser.add_argument('-lr', '--learning_rate', default=0.05, type=float,
                        metavar='LR', help='initial learning rate')
    parser.add_argument('-nwd', '--normalized_weightdecay', default=10.0,
                        type=float)
    parser.add_argument('-m', '--momentum', default=0.9, type=float,
                        metavar='M', help='momentum')
    parser.add_argument('-pf', '--printfreq', default=50, type=int,
                        metavar='N', help='print frequency')
    parser.add_argument('-r', '--resume', default='', type=str,
                        help='resume training from checkpoint')
    parser.add_argument('-d', '--device', default='cpu', type=str,
                        metavar='DEV', help='run on cpu/cuda')
    parser.add_argument('-s', '--seed', type=int,
                        help='if specified, fixes seed for reproducibility')
    parser.add_argument('-sd', '--save_dir',
                        help='The directory used to save the trained models',
                        default='save_temp', type=str)
    parser.add_argument('-dd', '--data_dir',
                        help='The directory to store dataset',
                        default='../data', type=str)
    parser.add_argument('-nb', '--bins', default=20, type=int,
                        help='number of bins for ece & reliability diagram')
    parser.add_argument('-tbd', '--tensorboard_dir', default='', type=str,
                        help='if specified, record data for tensorboard.')
    parser.add_argument('-sse', '--swag_start', type=int, default=160,
                        help='SWAG start epoch number')
    parser.add_argument('-slr', '--swag_lr', type=float, default=0.01,
                        help='SWAG learning rate')
    parser.add_argument('-scf', '--swag_collectfreq', type=int, default=1,
                        help='SWAG model collection frequency')
    parser.add_argument('-svf', '--swag_valfreq', type=int, default=5,
                        help='SWAG model validation frequency')
    parser.add_argument('-sdr', '--swag_devrank', type=int, metavar='K',
                        default=20, help='max rank of SWAG deviation matrix')
    parser.add_argument('-sbu', '--swag_bnupdate', action='store_true',
                        help='update BatchNorm for averaged model')
    return parser.parse_args()


if __name__ == '__main__':
    timer = coro_timer()
    t_init = next(timer)
    print(f'>>> Training initiated at {t_init.isoformat()} <<<\n')

    args = get_args()
    print(args, end='\n\n')

    assert 0 <= args.swag_start <= args.epochs

    # if seed is specified, run deterministically
    if args.seed is not None:
        deteministic_run(seed=args.seed)

    # get device for this experiment
    device = torch.device(args.device)

    if device != torch.device('cpu'):
        check_cuda()

    # build train_dir for this experiment
    mkdirp(args.save_dir)

    # compute weight decay
    weight_decay = get_weightdecay(
        args.normalized_weightdecay,
        int(SVHNInfo.counts['train'] * args.tvsplit))

    # resume or initialize
    swamodel = None
    if args.resume:
        startepoch, swagmodel, optimizer, scheduler, dic = loadcheckpoint(
            args.resume, device)
        modelargs, modelkwargs = dic['modelargs'], dic['modelkwargs']
        model = swagmodel.basemodel
        print(f'resumed from {args.resume}\n')
    else:
        startepoch = 0
        swagmodel = SWAGMODELS[args.arch](
            SVHNInfo.outclass, args.swag_devrank).to(args.device)
        modelargs, modelkwargs = (
            SVHNInfo.outclass, args.swag_devrank), {}
        model = swagmodel.basemodel
        optimizer = SGD(model.parameters(), args.learning_rate,
                        momentum=args.momentum, weight_decay=weight_decay)
        scheduler = schedule_midway_linear_decay(
            optimizer,
            epochs=args.epochs,
            start_decay=0.5*args.swag_start/args.epochs,
            end_decay=0.9*args.swag_start/args.epochs,
            end_scale=args.swag_lr/args.learning_rate)

    # prep tensorboard if specified
    if args.tensorboard_dir:
        mkdirp(args.tensorboard_dir)
        sw = SummaryWriter(args.tensorboard_dir)
    else:
        sw = None

    # load data
    train_loader, val_loader = get_svhn_train_loaders(
        args.data_dir, args.tvsplit, args.workers,
        (device != torch.device('cpu')), args.tbatch, args.vbatch)

    # perform training
    log_ece = coro_log(sw, args.printfreq, args.bins, args.save_dir)

    # standard training
    print('\n\n')
    print(f'>>> Base training starts at {next(timer)[0].isoformat()} <<<\n')

    for e in range(startepoch, args.swag_start):
        # run training part
        log_ece.send((e, 'train', len(train_loader), None))
        model.train()
        do_epoch(train_loader, do_trainbatch, log_ece, device, model=model,
                 optimizer=optimizer)
        log_ece.throw(StopIteration)
        # update lr scheduler and decay
        scheduler.step()
        # save checkpoint
        savecheckpoint(
            pjoin(args.save_dir, 'checkpoint.pt'), args.arch, modelargs,
            modelkwargs, swagmodel, optimizer, scheduler)

        print(f'>>> Time elapsed: {next(timer)[1]} <<<\n')

        # run evaluation part
        log_ece.send((e, 'val', len(val_loader), None))
        with torch.no_grad():
            model.eval()
            do_epoch(val_loader, do_evalbatch, log_ece, device, model=model)
        bins, _, avgvloss = log_ece.throw(StopIteration)[:3]

        print(f'>>> Time elapsed: {next(timer)[1]} <<<\n')

    # SWAG training
    print('\n\n')
    print(f'>>> SWAG training starts at {next(timer)[0].isoformat()} <<<\n')

    # adjust learning rate to swag_lr
    for param_group in optimizer.param_groups:
        param_group['lr'] = args.swag_lr

    for e in range(max(startepoch, args.swag_start), args.epochs):
        # run training part
        log_ece.send((e, 'train', len(train_loader), None))
        model.train()
        do_epoch(train_loader, do_trainbatch, log_ece, device, model=model,
                 optimizer=optimizer)
        log_ece.throw(StopIteration)
        # update lr scheduler
        scheduler.step()

        print(f'>>> Time elapsed: {next(timer)[1]} <<<\n')

        # collect base  model
        if (e - args.swag_start) % args.swag_collectfreq == 0:
            swagmodel.collect_model()

        # save checkpoint
        savecheckpoint(
            pjoin(args.save_dir, 'checkpoint.pt'), args.arch, modelargs,
            modelkwargs, swagmodel, optimizer, scheduler)

        # run evaluation part
        if (e - args.swag_start) % args.swag_valfreq == args.swag_valfreq - 1:
            with torch.no_grad():
                # validate base model
                log_ece.send((e, 'val', len(val_loader), None))
                model.eval()
                do_epoch(val_loader, do_evalbatch, log_ece, device,
                         model=model)
                log_ece.throw(StopIteration)
                # validate swa model
                log_ece.send((e, 'swaval', len(val_loader), None))
                swamodel = swagmodel.averaged_model(swamodel)
                if args.swag_bnupdate:
                    print('updating BatchNorm ...', end='')
                    bn_update(train_loader, swamodel, device=device)
                    print(' Done.')
                swamodel.eval()
                do_epoch(val_loader, do_evalbatch, log_ece, device,
                         model=swamodel)
                log_ece.throw(StopIteration)

        print(f'>>> Time elapsed: {next(timer)[1]} <<<\n')
    log_ece.close()

    # save final SWAG model
    savemodel(pjoin(args.save_dir, 'best_model.pt'), args.arch, modelargs,
              modelkwargs, swagmodel)
    # remove temporary checkpoint files
    rm(pjoin(args.save_dir, 'checkpoint.pt'))

    print(f'>>> Training completed at {next(timer)[0].isoformat()} <<<\n')