iwan.py

"""
@author: Baixu Chen
@contact: cbx_99_hasta@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.modules.classifier import Classifier
from tllib.modules.entropy import entropy
from tllib.modules.domain_discriminator import DomainDiscriminator
from tllib.reweight.iwan import ImportanceWeightModule, ImageClassifier
from tllib.alignment.dann import DomainAdversarialLoss
from tllib.utils.data import ForeverDataIterator
from tllib.utils.metric import accuracy
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
    if args.data == 'ImageNetCaltech':
        classifier = Classifier(backbone, num_classes, head=backbone.copy_head(), pool_layer=pool_layer).to(device)
    else:
        classifier = ImageClassifier(backbone, num_classes, args.bottleneck_dim, pool_layer=pool_layer).to(device)

    # define domain classifier D, D_0
    D = DomainDiscriminator(in_feature=classifier.features_dim, hidden_size=1024, batch_norm=False).to(device)
    D_0 = DomainDiscriminator(in_feature=classifier.features_dim, hidden_size=1024, batch_norm=False).to(device)

    # define optimizer and lr scheduler
    optimizer = SGD(classifier.get_parameters() + D.get_parameters() + D_0.get_parameters(),
                    args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True)
    lr_scheduler = LambdaLR(optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x)) ** (-args.lr_decay))

    # define loss function
    domain_adv_D = DomainAdversarialLoss(D).to(device)
    domain_adv_D_0 = DomainAdversarialLoss(D_0).to(device)
    # define importance weight module
    importance_weight_module = ImportanceWeightModule(D, train_target_dataset.partial_classes_idx)

    # 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 = utils.validate(test_loader, classifier, args, device)
        print(acc1)
        return

    # start training
    best_acc1 = 0.
    for epoch in range(args.epochs):
        # train for one epoch
        train(train_source_iter, train_target_iter, classifier, domain_adv_D, domain_adv_D_0,
              importance_weight_module, optimizer, lr_scheduler, epoch, args)

        # evaluate on validation set
        acc1 = utils.validate(val_loader, classifier, args, device)

        # remember best acc@1 and save checkpoint
        torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest'))
        if acc1 > best_acc1:
            shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best'))
        best_acc1 = max(acc1, best_acc1)

    print("best_acc1 = {:3.1f}".format(best_acc1))

    # evaluate on test set
    classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best')))
    acc1 = utils.validate(test_loader, classifier, args, device)
    print("test_acc1 = {:3.1f}".format(acc1))

    logger.close()


def train(train_source_iter: ForeverDataIterator, train_target_iter: ForeverDataIterator, model: ImageClassifier,
          domain_adv_D: DomainAdversarialLoss, domain_adv_D_0: DomainAdversarialLoss,
          importance_weight_module, optimizer: SGD, lr_scheduler: LambdaLR, epoch: int, args: argparse.Namespace):
    batch_time = AverageMeter('Time', ':5.2f')
    data_time = AverageMeter('Data', ':5.2f')
    losses = AverageMeter('Loss', ':6.2f')
    cls_accs = AverageMeter('Cls Acc', ':3.1f')
    tgt_accs = AverageMeter('Tgt Acc', ':3.1f')
    domain_accs_D = AverageMeter('Domain Acc for D', ':3.1f')
    domain_accs_D_0 = AverageMeter('Domain Acc for D_0', ':3.1f')
    partial_classes_weights = AverageMeter('Partial Weight', ':3.2f')
    non_partial_classes_weights = AverageMeter('Non-Partial Weight', ':3.2f')

    progress = ProgressMeter(
        args.iters_per_epoch,
        [batch_time, data_time, losses, cls_accs, tgt_accs,
         domain_accs_D, domain_accs_D_0, partial_classes_weights, non_partial_classes_weights],
        prefix="Epoch: [{}]".format(epoch))

    # switch to train mode
    model.train()
    domain_adv_D.train()
    domain_adv_D_0.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
        x = torch.cat((x_s, x_t), dim=0)
        y, f = model(x)
        y_s, y_t = y.chunk(2, dim=0)
        f_s, f_t = f.chunk(2, dim=0)

        # classification loss
        cls_loss = F.cross_entropy(y_s, labels_s)

        # domain adversarial loss for D
        adv_loss_D = domain_adv_D(f_s.detach(), f_t.detach())

        # get importance weights
        w_s = importance_weight_module.get_importance_weight(f_s)
        # domain adversarial loss for D_0
        adv_loss_D_0 = domain_adv_D_0(f_s, f_t, w_s=w_s)

        # entropy loss
        y_t = F.softmax(y_t, dim=1)
        entropy_loss = entropy(y_t, reduction='mean')

        loss = cls_loss + 1.5 * args.trade_off * adv_loss_D + \
               args.trade_off * adv_loss_D_0 + args.gamma * entropy_loss

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

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

        losses.update(loss.item(), x_s.size(0))
        cls_accs.update(cls_acc.item(), x_s.size(0))
        tgt_accs.update(tgt_acc.item(), x_s.size(0))
        domain_accs_D.update(domain_adv_D.domain_discriminator_accuracy, x_s.size(0))
        domain_accs_D_0.update(domain_adv_D_0.domain_discriminator_accuracy, x_s.size(0))

        # debug: output class weight averaged on the partial classes and non-partial classes respectively
        partial_class_weight, non_partial_classes_weight = \
            importance_weight_module.get_partial_classes_weight(w_s, labels_s)
        partial_classes_weights.update(partial_class_weight.item(), x_s.size(0))
        non_partial_classes_weights.update(non_partial_classes_weight.item(), x_s.size(0))

        batch_time.update(time.time() - end)
        end = time.time()

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


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='IWAN for Partial Domain Adaptation')
    # dataset parameters
    parser.add_argument('root', metavar='DIR',
                        help='root path of source (and target) 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')
    parser.add_argument('--gamma', default=0.1, type=float,
                        help='the trade-off hyper-parameter for entropy loss(default: 0.1)')
    parser.add_argument('--trade-off', default=3, type=float,
                        help='the trade-off hyper-parameter for transfer loss(default: 3))')
    # training parameters
    parser.add_argument('-b', '--batch-size', default=36, type=int,
                        metavar='N',
                        help='mini-batch size (default: 36)')
    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=0.001, 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=1e-3, type=float,
                        metavar='W', help='weight decay (default: 1e-3)',
                        dest='weight_decay')
    parser.add_argument('-j', '--workers', default=2, type=int, metavar='N',
                        help='number of data loading workers (default: 2)')
    parser.add_argument('--epochs', default=10, type=int, metavar='N',
                        help='number of total epochs to run')
    parser.add_argument('-i', '--iters-per-epoch', default=1000, 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='iwan',
                        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)