train_tar.py

import argparse
import os, sys
import os.path as osp
import torchvision
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import transforms
import network, loss
from torch.utils.data import DataLoader
from data_list import ImageList, ImageList_idx
import random, pdb, math, copy
from sklearn.metrics import confusion_matrix
import torch.nn.functional as F


def Entropy(input_):
    bs = input_.size(0)
    epsilon = 1e-5
    entropy = -input_ * torch.log(input_ + epsilon)
    entropy = torch.sum(entropy, dim=1)
    return entropy


def op_copy(optimizer):
    for param_group in optimizer.param_groups:
        param_group['lr0'] = param_group['lr']
    return optimizer


def lr_scheduler(optimizer, iter_num, max_iter, gamma=10, power=0.75):
    decay = (1 + gamma * iter_num / max_iter)**(-power)
    for param_group in optimizer.param_groups:
        param_group['lr'] = param_group['lr0'] * decay
        param_group['weight_decay'] = 1e-3
        param_group['momentum'] = 0.9
        param_group['nesterov'] = True
    return optimizer


def image_train(resize_size=256, crop_size=224, alexnet=False):
    if not alexnet:
        normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                         std=[0.229, 0.224, 0.225])
    else:
        normalize = Normalize(meanfile='./ilsvrc_2012_mean.npy')
    return transforms.Compose([
        transforms.Resize((resize_size, resize_size)),
        transforms.RandomCrop(crop_size),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(), normalize
    ])


def image_test(resize_size=256, crop_size=224, alexnet=False):
    if not alexnet:
        normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                         std=[0.229, 0.224, 0.225])
    else:
        normalize = Normalize(meanfile='./ilsvrc_2012_mean.npy')
    return transforms.Compose([
        transforms.Resize((resize_size, resize_size)),
        transforms.CenterCrop(crop_size),
        transforms.ToTensor(), normalize
    ])


def data_load(args):
    ## prepare data
    dsets = {}
    dset_loaders = {}
    train_bs = args.batch_size
    txt_src = open(args.s_dset_path).readlines()
    txt_tar = open(args.t_dset_path).readlines()
    txt_test = open(args.test_dset_path).readlines()

    dsize = len(txt_src)
    tr_size = int(0.9*dsize)
    # print(dsize, tr_size, dsize - tr_size)
    _, te_txt = torch.utils.data.random_split(txt_src, [tr_size, dsize - tr_size])
    tr_txt = txt_src

    dsets["source_tr"] = ImageList(tr_txt, transform=image_train())
    dset_loaders["source_tr"] = DataLoader(dsets["source_tr"],
                                           batch_size=train_bs,
                                           shuffle=True,
                                           num_workers=args.worker,
                                           drop_last=False)
    dsets["source_te"] = ImageList(te_txt, transform=image_test())
    dset_loaders["source_te"] = DataLoader(dsets["source_te"],
                                           batch_size=train_bs,
                                           shuffle=True,
                                           num_workers=args.worker,
                                           drop_last=False)
    dsets["target"] = ImageList_idx(txt_tar, transform=image_train())
    dset_loaders["target"] = DataLoader(dsets["target"],
                                        batch_size=train_bs,
                                        shuffle=True,
                                        num_workers=args.worker,
                                        drop_last=False)
    dsets["test"] = ImageList_idx(txt_test, transform=image_test())
    dset_loaders["test"] = DataLoader(dsets["test"],
                                      batch_size=train_bs * 3,
                                      shuffle=False,
                                      num_workers=args.worker,
                                      drop_last=False)

    return dset_loaders


def cal_acc(loader, netF, netB, netC, flag=False):
    start_test = True
    with torch.no_grad():
        iter_test = iter(loader)
        for i in range(len(loader)):
            data = iter_test.next()
            inputs = data[0]
            labels = data[1]
            inputs = inputs.cuda()
            outputs = netC(netB(netF(inputs)))
            if start_test:
                all_output = outputs.float().cpu()
                all_label = labels.float()
                start_test = False
            else:
                all_output = torch.cat((all_output, outputs.float().cpu()), 0)
                all_label = torch.cat((all_label, labels.float()), 0)
    _, predict = torch.max(all_output, 1)
    accuracy = torch.sum(
        torch.squeeze(predict).float() == all_label).item() / float(
            all_label.size()[0])
    mean_ent = torch.mean(Entropy(
        nn.Softmax(dim=1)(all_output))).cpu().data.item()

    if flag:
        matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
        acc = matrix.diagonal() / matrix.sum(axis=1) * 100
        aacc = acc.mean()
        aa = [str(np.round(i, 2)) for i in acc]
        acc = ' '.join(aa)
        return aacc, acc
    else:
        return accuracy * 100, mean_ent



def train_target(args):
    dset_loaders = data_load(args)
    ## set base network
    netF = network.ResBase(res_name=args.net).cuda()

    netB = network.feat_bootleneck(type=args.classifier,
                                   feature_dim=netF.in_features,
                                   bottleneck_dim=args.bottleneck).cuda()
    netC = network.feat_classifier(type=args.layer,
                                   class_num=args.class_num,
                                   bottleneck_dim=args.bottleneck).cuda()

    modelpath = args.output_dir_src + '/source_F.pt'
    netF.load_state_dict(torch.load(modelpath))
    modelpath = args.output_dir_src + '/source_B.pt'
    netB.load_state_dict(torch.load(modelpath))
    modelpath = args.output_dir_src + '/source_C.pt'
    netC.load_state_dict(torch.load(modelpath))

    param_group = []
    param_group_c=[]
    for k, v in netF.named_parameters():
        #if k.find('bn')!=-1:
        if True:
            param_group += [{'params': v, 'lr': args.lr * 0.1}]

    for k, v in netB.named_parameters():
        if True:
            param_group += [{'params': v, 'lr': args.lr * 1}]
    for k, v in netC.named_parameters():
        param_group_c += [{'params': v, 'lr': args.lr * 1}]

    optimizer = optim.SGD(param_group)
    optimizer = op_copy(optimizer)

    optimizer_c = optim.SGD(param_group_c)
    optimizer_c = op_copy(optimizer_c)


    #building feature bank and score bank
    loader = dset_loaders["target"]
    num_sample=len(loader.dataset)
    fea_bank=torch.randn(num_sample,256)
    score_bank = torch.randn(num_sample, 12).cuda()

    netF.eval()
    netB.eval()
    netC.eval()
    with torch.no_grad():
        iter_test = iter(loader)
        for i in range(len(loader)):
            data = iter_test.next()
            inputs = data[0]
            indx=data[-1]
            #labels = data[1]
            inputs = inputs.cuda()
            output = netB(netF(inputs))
            output_norm=F.normalize(output)
            outputs = netC(output)
            outputs=nn.Softmax(-1)(outputs)
            fea_bank[indx] = output_norm.detach().clone().cpu()
            score_bank[indx] = outputs.detach().clone()  #.cpu()

    max_iter = args.max_epoch * len(dset_loaders["target"])
    interval_iter = max_iter // args.interval
    iter_num = 0

    netF.train()
    netB.train()
    netC.train()
    acc_log=0
    while iter_num < max_iter:
        try:
            inputs_test, _, tar_idx = iter_test.next()
        except:
            iter_test = iter(dset_loaders["target"])
            inputs_test, _, tar_idx = iter_test.next()

        if inputs_test.size(0) == 1:
            continue

        inputs_test = inputs_test.cuda()

        iter_num += 1
        lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
        lr_scheduler(optimizer_c, iter_num=iter_num, max_iter=max_iter)

        features_test = netB(netF(inputs_test))
        outputs_test = netC(features_test)
        softmax_out = nn.Softmax(dim=1)(outputs_test)
        #output_re = softmax_out.unsqueeze(1)

        with torch.no_grad():
            output_f_norm = F.normalize(features_test)
            output_f_ = output_f_norm.cpu().detach().clone()

            fea_bank[tar_idx] = output_f_.detach().clone().cpu()
            score_bank[tar_idx] = softmax_out.detach().clone()

            distance = output_f_@fea_bank.T
            _, idx_near = torch.topk(distance,
                                    dim=-1,
                                    largest=True,
                                    k=args.K+1)
            idx_near = idx_near[:, 1:]  #batch x K
            score_near = score_bank[idx_near]    #batch x K x C

            fea_near = fea_bank[idx_near]  #batch x K x num_dim
            fea_bank_re = fea_bank.unsqueeze(0).expand(fea_near.shape[0],-1,-1) # batch x n x dim
            distance_ = torch.bmm(fea_near, fea_bank_re.permute(0,2,1))  # batch x K x n
            _,idx_near_near=torch.topk(distance_,dim=-1,largest=True,k=args.KK+1)  # M near neighbors for each of above K ones
            idx_near_near = idx_near_near[:,:,1:] # batch x K x M
            tar_idx_ = tar_idx.unsqueeze(-1).unsqueeze(-1)
            match = (
                idx_near_near == tar_idx_).sum(-1).float()  # batch x K
            weight = torch.where(
                match > 0., match,
                torch.ones_like(match).fill_(0.1))  # batch x K

            weight_kk = weight.unsqueeze(-1).expand(-1, -1,
                                                    args.KK)  # batch x K x M
            weight_kk = weight_kk.fill_(0.1)

            # removing the self in expanded neighbors, or otherwise you can keep it and not use extra self regularization
            #weight_kk[idx_near_near == tar_idx_]=0

            score_near_kk = score_bank[idx_near_near]  # batch x K x M x C
            #print(weight_kk.shape)
            weight_kk = weight_kk.contiguous().view(weight_kk.shape[0],
                                                    -1)  # batch x KM

            score_near_kk = score_near_kk.contiguous().view(score_near_kk.shape[0], -1,
                                               12)  # batch x KM x C

            score_self = score_bank[tar_idx]


        # nn of nn
        output_re = softmax_out.unsqueeze(1).expand(-1, args.K * args.KK,
                                                    -1)  # batch x C x 1
        const = torch.mean(
            (F.kl_div(output_re, score_near_kk, reduction='none').sum(-1) *
             weight_kk.cuda()).sum(1)) # kl_div here equals to dot product since we do not use log for score_near_kk
        loss = torch.mean(const)


        # nn
        softmax_out_un = softmax_out.unsqueeze(1).expand(-1, args.K,
                                                         -1)  # batch x K x C

        loss += torch.mean((
            F.kl_div(softmax_out_un, score_near, reduction='none').sum(-1) *
            weight.cuda()).sum(1))

        # self, if not explicitly removing the self feature in expanded neighbor then no need for this
        #loss += -torch.mean((softmax_out * score_self).sum(-1))


        msoftmax = softmax_out.mean(dim=0)
        gentropy_loss = torch.sum(msoftmax *
                                    torch.log(msoftmax + args.epsilon))
        loss += gentropy_loss

        optimizer.zero_grad()
        optimizer_c.zero_grad()
        loss.backward()
        optimizer.step()
        optimizer_c.step()


        if iter_num % interval_iter == 0 or iter_num == max_iter:
            netF.eval()
            netB.eval()
            netC.eval()
            if args.dset == 'visda-2017':
                acc_s_te, acc_list = cal_acc(dset_loaders['test'], netF, netB,
                                             netC,flag= True)
                log_str = 'Task: {}, Iter:{}/{}; Accuracy on target = {:.2f}%'.format(
                    args.name, iter_num, max_iter, acc_s_te
                ) + '\n' + 'T: ' + acc_list

            args.out_file.write(log_str + '\n')
            args.out_file.flush()
            print(log_str + '\n')
            netF.train()
            netB.train()
            netC.train()

            if acc_s_te>acc_log:
                acc_log=acc_s_te
                torch.save(
                    netF.state_dict(),
                    osp.join(args.output_dir, "target_F_" + '2021_'+str(args.K) + ".pt"))
                torch.save(
                    netB.state_dict(),
                    osp.join(args.output_dir,
                                "target_B_" + '2021_' + str(args.K) + ".pt"))
                torch.save(
                    netC.state_dict(),
                    osp.join(args.output_dir,
                                "target_C_" + '2021_' + str(args.K) + ".pt"))


    return netF, netB, netC


def print_args(args):
    s = "==========================================\n"
    for arg, content in args.__dict__.items():
        s += "{}:{}\n".format(arg, content)
    return s



if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='Neighbors')
    parser.add_argument('--gpu_id',
                        type=str,
                        nargs='?',
                        default='9',
                        help="device id to run")
    parser.add_argument('--s', type=int, default=0, help="source")
    parser.add_argument('--t', type=int, default=1, help="target")
    parser.add_argument('--max_epoch',
                        type=int,
                        default=15,
                        help="max iterations")
    parser.add_argument('--interval', type=int, default=15)
    parser.add_argument('--batch_size',
                        type=int,
                        default=64,
                        help="batch_size")
    parser.add_argument('--worker',
                        type=int,
                        default=4,
                        help="number of workers")
    parser.add_argument(
        '--dset',
        type=str,
        default='visda-2017')
    parser.add_argument('--lr', type=float, default=1e-3, help="learning rate")
    parser.add_argument('--net',
                        type=str,
                        default='resnet101')
    parser.add_argument('--seed', type=int, default=2021, help="random seed")

    parser.add_argument('--bottleneck', type=int, default=256)
    parser.add_argument('--K', type=int, default=5)
    parser.add_argument('--KK', type=int, default=5)
    parser.add_argument('--epsilon', type=float, default=1e-5)
    parser.add_argument('--layer',
                        type=str,
                        default="wn",
                        choices=["linear", "wn"])
    parser.add_argument('--classifier',
                        type=str,
                        default="bn",
                        choices=["ori", "bn"])
    parser.add_argument('--output', type=str, default='weight/target/')
    parser.add_argument('--output_src', type=str, default='weight/source/')
    parser.add_argument('--tag', type=str, default='selfplus')
    parser.add_argument('--da',
                        type=str,
                        default='uda')
    parser.add_argument('--issave', type=bool, default=True)
    args = parser.parse_args()

    if args.dset == 'office-home':
        names = ['Art', 'Clipart', 'Product', 'RealWorld']
        args.class_num = 65
    if args.dset == 'visda-2017':
        names = ['train', 'validation']
        args.class_num = 12

    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
    SEED = args.seed
    torch.manual_seed(SEED)
    torch.cuda.manual_seed(SEED)
    np.random.seed(SEED)
    random.seed(SEED)
    torch.backends.cudnn.deterministic = True

    for i in range(len(names)):
        if i == args.s:
            continue
        args.t = i

        folder = './data/'
        args.s_dset_path = folder + args.dset + '/' + names[
            args.s] + '_list.txt'
        args.t_dset_path = folder + args.dset + '/' + names[
            args.t] + '_list.txt'
        args.test_dset_path = folder + args.dset + '/' + names[
            args.t] + '_list.txt'

        args.output_dir_src = osp.join(args.output_src, args.da, args.dset,
                                       names[args.s][0].upper())
        args.output_dir = osp.join(
            args.output, args.da, args.dset,
            names[args.s][0].upper() + names[args.t][0].upper())
        args.name = names[args.s][0].upper() + names[args.t][0].upper()

        if not osp.exists(args.output_dir):
            os.system('mkdir -p ' + args.output_dir)
        if not osp.exists(args.output_dir):
            os.mkdir(args.output_dir)

        args.out_file = open(
            osp.join(args.output_dir, 'log_target.txt'), 'w')
        args.out_file.write(print_args(args) + '\n')
        args.out_file.flush()
        train_target(args)