osbp.py

"""
@author: Junguang Jiang
@contact: JiangJunguang1123@outlook.com
"""
import random
import time
import warnings
import sys
import argparse
import shutil
import os.path as osp

import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from torch.optim import SGD
from torch.optim.lr_scheduler import LambdaLR
from torch.utils.data import DataLoader
import torch.nn.functional as F

import utils
from tllib.alignment.osbp import ImageClassifier as Classifier, UnknownClassBinaryCrossEntropy
from tllib.utils.data import ForeverDataIterator
from tllib.utils.metric import accuracy, ConfusionMatrix
from tllib.utils.meter import AverageMeter, ProgressMeter
from tllib.utils.logger import CompleteLogger
from tllib.utils.analysis import collect_feature, tsne, a_distance

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


def main(args: argparse.Namespace):
    logger = CompleteLogger(args.log, args.phase)
    print(args)

    if args.seed is not None:
        random.seed(args.seed)
        torch.manual_seed(args.seed)
        cudnn.deterministic = True
        warnings.warn('You have chosen to seed training. '
                      'This will turn on the CUDNN deterministic setting, '
                      'which can slow down your training considerably! '
                      'You may see unexpected behavior when restarting '
                      'from checkpoints.')

    cudnn.benchmark = True

    # Data loading code
    train_transform = utils.get_train_transform(args.train_resizing, random_horizontal_flip=True,
                                                random_color_jitter=False)
    val_transform = utils.get_val_transform(args.val_resizing)
    print("train_transform: ", train_transform)
    print("val_transform: ", val_transform)

    train_source_dataset, train_target_dataset, val_dataset, test_dataset, num_classes, args.class_names = \
        utils.get_dataset(args.data, args.root, args.source, args.target, train_transform, val_transform)
    train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size,
                                     shuffle=True, num_workers=args.workers, drop_last=True)
    train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size,
                                     shuffle=True, num_workers=args.workers, drop_last=True)
    val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers)
    test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers)

    train_source_iter = ForeverDataIterator(train_source_loader)
    train_target_iter = ForeverDataIterator(train_target_loader)

    # create model
    print("=> using pre-trained model '{}'".format(args.arch))
    backbone = utils.get_model(args.arch)
    pool_layer = nn.Identity() if args.no_pool else None
    classifier = Classifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim, pool_layer=pool_layer).to(device)
    print(classifier)
    unknown_bce = UnknownClassBinaryCrossEntropy(t=0.5)

    # define optimizer and lr scheduler
    optimizer = SGD(classifier.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True)
    lr_scheduler = LambdaLR(optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x)) ** (-args.lr_decay))

    # resume from the best checkpoint
    if args.phase != 'train':
        checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu')
        classifier.load_state_dict(checkpoint)

    # analysis the model
    if args.phase == 'analysis':
        # extract features from both domains
        feature_extractor = nn.Sequential(classifier.backbone, classifier.pool_layer, classifier.bottleneck).to(device)
        source_feature = collect_feature(train_source_loader, feature_extractor, device)
        target_feature = collect_feature(train_target_loader, feature_extractor, device)
        # plot t-SNE
        tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png')
        tsne.visualize(source_feature, target_feature, tSNE_filename)
        print("Saving t-SNE to", tSNE_filename)
        # calculate A-distance, which is a measure for distribution discrepancy
        A_distance = a_distance.calculate(source_feature, target_feature, device)
        print("A-distance =", A_distance)
        return

    if args.phase == 'test':
        acc1 = validate(test_loader, classifier, args)
        print(acc1)
        return

    # start training
    best_h_score = 0.
    for epoch in range(args.epochs):
        # train for one epoch
        train(train_source_iter, train_target_iter, classifier, unknown_bce, optimizer,
              lr_scheduler, epoch, args)

        # evaluate on validation set
        h_score = validate(val_loader, classifier, args)

        # remember best acc@1 and save checkpoint
        torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest'))
        if h_score > best_h_score:
            shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best'))
        best_h_score = max(h_score, best_h_score)

    print("best_h_score = {:3.1f}".format(best_h_score))

    # evaluate on test set
    classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best')))
    h_score = validate(test_loader, classifier, args)
    print("test_h_score = {:3.1f}".format(h_score))

    logger.close()


def train(train_source_iter: ForeverDataIterator, train_target_iter: ForeverDataIterator, model: Classifier,
          unknown_bce: UnknownClassBinaryCrossEntropy, optimizer: SGD,
          lr_scheduler: LambdaLR, epoch: int, args: argparse.Namespace):
    batch_time = AverageMeter('Time', ':4.2f')
    data_time = AverageMeter('Data', ':3.1f')
    losses = AverageMeter('Loss', ':3.2f')
    cls_accs = AverageMeter('Cls Acc', ':3.1f')
    tgt_accs = AverageMeter('Tgt Acc', ':3.1f')
    trans_losses = AverageMeter('Trans Loss', ':3.2f')

    progress = ProgressMeter(
        args.iters_per_epoch,
        [batch_time, data_time, losses, trans_losses, cls_accs, tgt_accs],
        prefix="Epoch: [{}]".format(epoch))

    # switch to train mode
    model.train()

    end = time.time()
    for i in range(args.iters_per_epoch):
        x_s, labels_s = next(train_source_iter)
        x_t, labels_t = next(train_target_iter)

        x_s = x_s.to(device)
        x_t = x_t.to(device)
        labels_s = labels_s.to(device)
        labels_t = labels_t.to(device)

        # measure data loading time
        data_time.update(time.time() - end)

        # compute output
        y_s, _ = model(x_s, grad_reverse=False)
        y_t, _ = model(x_t, grad_reverse=True)

        cls_loss = F.cross_entropy(y_s, labels_s)
        trans_loss = unknown_bce(y_t)
        loss = cls_loss + trans_loss

        cls_acc = accuracy(y_s, labels_s)[0]
        tgt_acc = accuracy(y_t, labels_t)[0]

        losses.update(loss.item(), x_s.size(0))
        trans_losses.update(trans_loss.item(), x_s.size(0))
        cls_accs.update(cls_acc.item(), x_s.size(0))
        tgt_accs.update(tgt_acc.item(), x_t.size(0))

        # compute gradient and do SGD step
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        lr_scheduler.step()

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % args.print_freq == 0:
            progress.display(i)


def validate(val_loader: DataLoader, model: Classifier, args: argparse.Namespace) -> float:
    batch_time = AverageMeter('Time', ':6.3f')
    classes = val_loader.dataset.classes
    confmat = ConfusionMatrix(len(classes))
    progress = ProgressMeter(
        len(val_loader),
        [batch_time],
        prefix='Test: ')

    # switch to evaluate mode
    model.eval()

    with torch.no_grad():
        end = time.time()
        for i, (images, target) in enumerate(val_loader):
            images = images.to(device)
            target = target.to(device)

            # compute output
            output = model(images)

            # measure accuracy and record loss
            confmat.update(target, output.argmax(1))

            # measure elapsed time
            batch_time.update(time.time() - end)
            end = time.time()

            if i % args.print_freq == 0:
                progress.display(i)

        acc_global, accs, iu = confmat.compute()
        all_acc = torch.mean(accs).item() * 100
        known = torch.mean(accs[:-1]).item() * 100
        unknown = accs[-1].item() * 100
        h_score = 2 * known * unknown / (known + unknown)
        if args.per_class_eval:
            print(confmat.format(classes))
        print(' * All {all:.3f} Known {known:.3f} Unknown {unknown:.3f} H-score {h_score:.3f}'
              .format(all=all_acc, known=known, unknown=unknown, h_score=h_score))

    return h_score


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='OSBP for Openset Domain Adaptation')
    # dataset parameters
    parser.add_argument('root', metavar='DIR',
                        help='root path of dataset')
    parser.add_argument('-d', '--data', metavar='DATA', default='Office31', choices=utils.get_dataset_names(),
                        help='dataset: ' + ' | '.join(utils.get_dataset_names()) +
                             ' (default: Office31)')
    parser.add_argument('-s', '--source', help='source domain')
    parser.add_argument('-t', '--target', help='target domain')
    parser.add_argument('--train-resizing', type=str, default='default')
    parser.add_argument('--val-resizing', type=str, default='default')
    # model parameters
    parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet18',
                        choices=utils.get_model_names(),
                        help='backbone architecture: ' +
                             ' | '.join(utils.get_model_names()) +
                             ' (default: resnet18)')
    parser.add_argument('--no-pool', action='store_true',
                        help='no pool layer after the feature extractor.')
    parser.add_argument('--bottleneck-dim', default=256, type=int,
                        help='Dimension of bottleneck')
    # training parameters
    parser.add_argument('-b', '--batch-size', default=32, type=int,
                        metavar='N',
                        help='mini-batch size (default: 32)')
    parser.add_argument('--lr', '--learning-rate', default=0.001, type=float,
                        metavar='LR', help='initial learning rate', dest='lr')
    parser.add_argument('--lr-gamma', default=0.0003, type=float, help='parameter for lr scheduler')
    parser.add_argument('--lr-decay', default=0.75, type=float, help='parameter for lr scheduler')
    parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
                        help='momentum')
    parser.add_argument('--wd', '--weight-decay', default=0.0005, type=float,
                        metavar='W', help='weight decay (default: 5e-4)')
    parser.add_argument('-j', '--workers', default=2, type=int, metavar='N',
                        help='number of data loading workers (default: 4)')
    parser.add_argument('--epochs', default=20, type=int, metavar='N',
                        help='number of total epochs to run')
    parser.add_argument('-i', '--iters-per-epoch', default=500, type=int,
                        help='Number of iterations per epoch')
    parser.add_argument('-p', '--print-freq', default=100, type=int,
                        metavar='N', help='print frequency (default: 100)')
    parser.add_argument('--seed', default=None, type=int,
                        help='seed for initializing training. ')
    parser.add_argument('--per-class-eval', action='store_true',
                        help='whether output per-class accuracy during evaluation')
    parser.add_argument("--log", type=str, default='osbp',
                        help="Where to save logs, checkpoints and debugging images.")
    parser.add_argument("--phase", type=str, default='train', choices=['train', 'test', 'analysis'],
                        help="When phase is 'test', only test the model."
                             "When phase is 'analysis', only analysis the model.")
    args = parser.parse_args()
    main(args)