def train(config): ## set pre-process prep_dict = {} prep_config = config["prep"] prep_dict["source"] = prep.image_train(**config["prep"]['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) if prep_config["test_10crop"]: prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params']) else: prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data dsets = {} dset_loaders = {} data_config = config["data"] train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \ transform=prep_dict["source"]) # print(prep_dict["source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \ transform=prep_dict["target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) if prep_config["test_10crop"]: for i in range(10): dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"][i]) for i in range(10)] dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \ shuffle=False, num_workers=4) for dset in dsets['test']] else: dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=4) class_num = config["network"]["params"]["class_num"] ## set base network net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.cuda() # base_network = base_network.cpu() ## add additional network for some methods if config["loss"]["random"]: random_layer = network.RandomLayer( [base_network.output_num(), class_num], config["loss"]["random_dim"]) ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None ad_net = network.AdversarialNetwork( base_network.output_num() * class_num, 1024) if config["loss"]["random"]: random_layer.cuda() ad_net = ad_net.cuda() # ad_net = ad_net.cpu() parameter_list = base_network.get_parameters() + ad_net.get_parameters() ## set optimizer optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) transfer_loss_value = classifier_loss_value = total_loss_value = 0.0 best_acc = 0.0 for i in range(config["num_iterations"]): if i % config["test_interval"] == config["test_interval"] - 1: base_network.train(False) temp_acc = image_classification_test(dset_loaders, \ base_network, test_10crop=prep_config["test_10crop"]) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str) if i % config["snapshot_interval"] == 0: torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \ "iter_{:05d}_model.pth.tar".format(i))) print("it_train: {:05d} / {:05d} start".format( i, config["num_iterations"])) loss_params = config["loss"] ## train one iter base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, labels_target = iter_target.next() inputs_source, inputs_target, labels_source = inputs_source.cuda( ), inputs_target.cuda(), labels_source.cuda() # inputs_source, inputs_target, labels_source = inputs_source.cpu(), inputs_target.cpu(), labels_source.cpu() features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) if config['method'] == 'CDAN+E': entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) elif config['method'] == 'CDAN': transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer) elif config['method'] == 'DANN': transfer_loss = loss.DANN(features, ad_net) else: raise ValueError('Method cannot be recognized.') classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss total_loss.backward() optimizer.step() print("it_train: {:05d} / {:05d} over".format( i, config["num_iterations"])) torch.save(best_model, osp.join(config["output_path"], "best_model.pth.tar")) return best_acc
def train(args, model, ad_net, random_layer, train_loader, train_loader1, optimizer, optimizer_ad, epoch, start_epoch, method, D_s, D_t, G_s2t, G_t2s, criterion_Sem, criterion_GAN, criterion_cycle, criterion_identity, optimizer_G, optimizer_D_t, optimizer_D_s, classifier1, classifier1_optim, fake_S_buffer, fake_T_buffer): model.train() len_source = len(train_loader) len_target = len(train_loader1) if len_source < len_target: num_iter = len_source else: num_iter = len_target for batch_idx in range(num_iter - 1): if batch_idx % len_source == 0: iter_source = iter(train_loader) if batch_idx % len_target == 0: iter_target = iter(train_loader1) data_source, label_source = iter_source.next() # data_source, label_source = data_source.cuda(), label_source.cuda() data_target, label_target = iter_target.next() # data_target = data_target.cuda() optimizer.zero_grad() optimizer_ad.zero_grad() features_source, outputs_source = model(data_source) features_target, outputs_target = model(data_target) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) #feature, output = model(torch.cat((data_source, data_target), 0)) loss = nn.CrossEntropyLoss()(outputs.narrow(0, 0, data_source.size(0)), label_source) softmax_output = nn.Softmax(dim=1)(outputs) output1 = classifier1(features) softmax_output1 = nn.Softmax(dim=1)(output1) softmax_output = ( 1 - args.cla_plus_weight ) * softmax_output + args.cla_plus_weight * softmax_output1 if epoch > start_epoch: if method == 'CDAN-E': entropy = loss_func.Entropy(softmax_output) loss += loss_func.CDAN( [features, softmax_output], ad_net, entropy, network.calc_coeff(num_iter * (epoch - start_epoch) + batch_idx), random_layer) elif method == 'CDAN': loss += loss_func.CDAN([features, softmax_output], ad_net, None, None, random_layer) elif method == 'DANN': loss += loss_func.DANN(features, ad_net) else: raise ValueError('Method cannot be recognized.') # Cycle num_feature = features.size(0) # =================train discriminator T real_label = Variable(torch.ones(num_feature)) # real_label = Variable(torch.ones(num_feature)).cuda() fake_label = Variable(torch.zeros(num_feature)) # fake_label = Variable(torch.zeros(num_feature)).cuda() # 训练生成器 optimizer_G.zero_grad() # Identity loss same_t = G_s2t(features_target) loss_identity_t = criterion_identity(same_t, features_target) same_s = G_t2s(features_source) loss_identity_s = criterion_identity(same_s, features_source) # Gan loss fake_t = G_s2t(features_source) pred_fake = D_t(fake_t) loss_G_s2t = criterion_GAN(pred_fake, label_source.float()) fake_s = G_t2s(features_target) pred_fake = D_s(fake_s) loss_G_t2s = criterion_GAN(pred_fake, label_source.float()) # cycle loss recovered_s = G_t2s(fake_t) loss_cycle_sts = criterion_cycle(recovered_s, features_source) recovered_t = G_s2t(fake_s) loss_cycle_tst = criterion_cycle(recovered_t, features_target) # sem loss pred_recovered_s = model.classifier(recovered_s) pred_fake_t = model.classifier(fake_t) loss_sem_t2s = criterion_Sem(pred_recovered_s, pred_fake_t) pred_recovered_t = model.classifier(recovered_t) pred_fake_s = model.classifier(fake_s) loss_sem_s2t = criterion_Sem(pred_recovered_t, pred_fake_s) loss_cycle = loss_cycle_tst + loss_cycle_sts weight_in_loss_g = args.weight_in_loss_g.split(',') loss_G = float(weight_in_loss_g[0]) * (loss_identity_s + loss_identity_t) + \ float(weight_in_loss_g[1]) * (loss_G_s2t + loss_G_t2s) + \ float(weight_in_loss_g[2])* loss_cycle + \ float(weight_in_loss_g[3]) * (loss_sem_s2t + loss_sem_t2s) # 训练softmax分类器 outputs_fake = classifier1(fake_t.detach()) # 分类器优化 classifier_loss1 = nn.CrossEntropyLoss()(outputs_fake, label_source) classifier1_optim.zero_grad() classifier_loss1.backward() classifier1_optim.step() total_loss = loss + args.cyc_loss_weight * loss_G total_loss.backward() optimizer.step() optimizer_G.step() ###### Discriminator S ###### optimizer_D_s.zero_grad() # Real loss pred_real = D_s(features_source.detach()) loss_D_real = criterion_GAN(pred_real, real_label) # Fake loss fake_s = fake_S_buffer.push_and_pop(fake_s) pred_fake = D_s(fake_s.detach()) loss_D_fake = criterion_GAN(pred_fake, fake_label) # Total loss loss_D_s = loss_D_real + loss_D_fake loss_D_s.backward() optimizer_D_s.step() ################################### ###### Discriminator t ###### optimizer_D_t.zero_grad() # Real loss pred_real = D_t(features_target.detach()) loss_D_real = criterion_GAN(pred_real, real_label) # Fake loss fake_t = fake_T_buffer.push_and_pop(fake_t) pred_fake = D_t(fake_t.detach()) loss_D_fake = criterion_GAN(pred_fake, fake_label) # Total loss loss_D_t = loss_D_real + loss_D_fake loss_D_t.backward() optimizer_D_t.step() if epoch > start_epoch: optimizer_ad.step() if (batch_idx + epoch * num_iter) % args.log_interval == 0: print( 'Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tLoss+G: {:.6f}' .format(epoch, batch_idx * args.batch_size, num_iter * args.batch_size, 100. * batch_idx / num_iter, loss.item(), total_loss.item()))
def train(args, model, ad_net, random_layer, train_loader, train_loader1, optimizer, optimizer_ad, epoch, start_epoch, method, ccp): cl_method = 'ga' #choices=['ga', 'nn', 'free', 'pc', 'forward'] meta_method = 'free' if cl_method == 'ga' else cl_method K = 10 model.train() len_source = len(train_loader) len_target = len(train_loader1) if len_source > len_target: num_iter = len_source else: num_iter = len_target for batch_idx in range(num_iter): if batch_idx % len_source == 0: iter_source = iter(train_loader) if batch_idx % len_target == 0: iter_target = iter(train_loader1) data_source, label_source = iter_source.next() data_source, label_source = data_source.cuda(), label_source.cuda() data_target, label_target = iter_target.next() data_target = data_target.cuda() optimizer.zero_grad() optimizer_ad.zero_grad() feature, output = model(torch.cat((data_source, data_target), 0)) #err_s_label, loss_vector = non_negative_loss (f=output.narrow(0, 0, data_source.size(0)), K=10, labels=label_source, ccp=ccp,beta=0) loss, loss_vector = chosen_loss_c(f=output.narrow( 0, 0, data_source.size(0)), K=K, labels=label_source, ccp=ccp, meta_method=meta_method) #loss = nn.CrossEntropyLoss()(output.narrow(0, 0, data_source.size(0)), label_source) softmax_output = nn.Softmax(dim=1)(output) if cl_method == 'ga': if torch.min(loss_vector).item() < 0: loss_vector_with_zeros = torch.cat( (loss_vector.view(-1, 1), torch.zeros( K, requires_grad=True).view(-1, 1).to(device)), 1) min_loss_vector, _ = torch.min(loss_vector_with_zeros, dim=1) loss = torch.sum(min_loss_vector) loss.backward(retain_graph=True) for group in optimizer.param_groups: for p in group['params']: p.grad = -1 * p.grad else: loss.backward(retain_graph=True) else: loss.backward(retain_graph=True) optimizer.step() optimizer.zero_grad() if epoch > start_epoch: if method == 'CDAN-E': softmax_output = Tsharpen(softmax_output) entropy = loss_func.Entropy(softmax_output) loss2 = loss_func.CDAN( [feature, softmax_output], ad_net, entropy, network.calc_coeff(num_iter * (epoch - start_epoch) + batch_idx), random_layer) elif method == 'CDAN': loss2 = loss_func.CDAN([feature, softmax_output], ad_net, None, None, random_layer) elif method == 'DANN': loss2 = loss_func.DANN(feature, ad_net) else: raise ValueError('Method cannot be recognized.') if epoch > start_epoch: loss2.backward() optimizer.step() optimizer_ad.step() if (batch_idx + epoch * num_iter) % args.log_interval == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss1: {:.6f}'.format( epoch, batch_idx * args.batch_size, num_iter * args.batch_size, 100. * batch_idx / num_iter, loss.item()))
def train(config): ## set pre-process prep_dict = {} prep_config = config["prep"] prep_dict["source"] = prep.image_train(**config["prep"]['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) if prep_config["test_10crop"]: prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params']) else: prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data dsets = {} dset_loaders = {} data_config = config["data"] train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \ transform=prep_dict["source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=0, drop_last=True) dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \ transform=prep_dict["target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=0, drop_last=True) if prep_config["test_10crop"]: for i in range(10): dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"][i]) for i in range(10)] dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \ shuffle=False, num_workers=0) for dset in dsets['test']] else: dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=0) class_num = config["network"]["params"]["class_num"] ## set base network net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.cuda() ## add additional network for some methods if config["loss"]["random"]: random_layer = network.RandomLayer( [base_network.output_num(), class_num], config["loss"]["random_dim"]) ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None ad_net = network.AdversarialNetwork( base_network.output_num() * class_num, 1024) if config["loss"]["random"]: random_layer.cuda() ad_net = ad_net.cuda() parameter_list = base_network.get_parameters() + ad_net.get_parameters() ## set optimizer optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) best_acc = 0.0 best_model = nn.Sequential(base_network) each_log = "" for i in range(config["num_iterations"]): if i % config["test_interval"] == config["test_interval"] - 1: base_network.train(False) temp_acc = image_classification_test(dset_loaders, \ base_network, test_10crop=prep_config["test_10crop"]) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model log_str = "iter: {:05d}, precision: {:.5f}, transfer_loss:{:.4f}, classifier_loss:{:.4f}, total_loss:{:.4f}" \ .format(i, temp_acc, transfer_loss.item(), classifier_loss.item(), total_loss.item()) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str) config["out_file"].write(each_log) config["out_file"].flush() each_log = "" loss_params = config["loss"] ## train one iter base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, labels_target = iter_target.next() inputs_source, inputs_target, labels_source = inputs_source.cuda( ), inputs_target.cuda(), labels_source.cuda() features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) labels_target_fake = torch.max(nn.Softmax(dim=1)(outputs_target), 1)[1] labels = torch.cat((labels_source, labels_target_fake)) entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) mdd_loss = loss.mdd_loss(features=features, labels=labels, left_weight=args.left_weight, right_weight=args.right_weight) max_entropy_loss = loss.EntropicConfusion(features) total_loss = loss_params["trade_off"] * transfer_loss \ + args.cls_weight * classifier_loss \ + args.mdd_weight * mdd_loss \ + args.entropic_weight * max_entropy_loss total_loss.backward() optimizer.step() log_str = "iter: {:05d},transfer_loss:{:.4f}, classifier_loss:{:.4f}, mdd_loss:{:4f}," \ "max_entropy_loss:{:.4f},total_loss:{:.4f}" \ .format(i, transfer_loss.item(), classifier_loss.item(), mdd_loss.item(), max_entropy_loss.item(), total_loss.item()) each_log += log_str + "\n" torch.save( best_model, config['model_output_path'] + "{}_{}_p-{}_e-{}".format( config['log_name'], str(best_acc), str(config["mdd_weight"]), str(config["entropic_weight"]))) return best_acc
def train(args, model, ad_net, source_samples, source_labels, target_samples, target_labels, optimizer, optimizer_ad, epoch, start_epoch, method, source_label_distribution, out_wei_file, cov_mat, pseudo_target_label, class_weights, true_weights): model.train() cov_mat[:] = 0.0 pseudo_target_label[:] = 0.0 len_source = source_labels.shape[0] len_target = target_labels.shape[0] size = max(len_source, len_target) num_iter = int(size / args.batch_size) for batch_idx in range(num_iter): t = time.time() source_idx = np.random.choice(len_source, args.batch_size) target_idx = np.random.choice(len_target, args.batch_size) data_source, label_source = source_samples[source_idx], source_labels[ source_idx] data_target, _ = target_samples[target_idx], target_labels[target_idx] optimizer.zero_grad() optimizer_ad.zero_grad() feature, output = model(torch.cat((data_source, data_target), 0)) if 'IW' in method: ys_onehot = torch.zeros(args.batch_size, 10).to(args.device) ys_onehot.scatter_(1, label_source.view(-1, 1), 1) # Compute weights on source data. if 'ORACLE' in method: weights = torch.mm(ys_onehot, true_weights) else: weights = torch.mm(ys_onehot, model.im_weights) source_preds, target_preds = output[:args.batch_size], output[ args.batch_size:] # Compute the aggregated distribution of pseudo-label on the target domain. pseudo_target_label += torch.sum(F.softmax(target_preds, dim=1), dim=0).view(-1, 1).detach() # Update the covariance matrix on the source domain as well. cov_mat += torch.mm( F.softmax(source_preds, dim=1).transpose(1, 0), ys_onehot).detach() loss = torch.mean( nn.CrossEntropyLoss(weight=class_weights, reduction='none') (output.narrow(0, 0, data_source.size(0)), label_source) * weights) / 10.0 else: loss = nn.CrossEntropyLoss()(output.narrow(0, 0, data_source.size(0)), label_source) if epoch > start_epoch: if method == 'CDAN-E': softmax_output = nn.Softmax(dim=1)(output) entropy = loss_func.Entropy(softmax_output) loss += loss_func.CDAN( [feature, softmax_output], ad_net, entropy, network.calc_coeff(num_iter * (epoch - start_epoch) + batch_idx), None, device=args.device) elif 'IWCDAN-E' in method: softmax_output = nn.Softmax(dim=1)(output) entropy = loss_func.Entropy(softmax_output) loss += loss_func.CDAN( [feature, softmax_output], ad_net, entropy, network.calc_coeff(num_iter * (epoch - start_epoch) + batch_idx), None, weights=weights, device=args.device) elif method == 'CDAN': softmax_output = nn.Softmax(dim=1)(output) loss += loss_func.CDAN([feature, softmax_output], ad_net, None, None, None, device=args.device) elif 'IWCDAN' in method: softmax_output = nn.Softmax(dim=1)(output) loss += loss_func.CDAN([feature, softmax_output], ad_net, None, None, None, weights=weights, device=args.device) elif method == 'DANN': loss += loss_func.DANN(feature, ad_net, args.device) elif 'IWDAN' in method: dloss = loss_func.IWDAN(feature, ad_net, weights) loss += args.mu * dloss elif method == 'NANN': pass else: raise ValueError('Method cannot be recognized.') loss.backward() optimizer.step() if epoch > start_epoch and method != 'NANN': optimizer_ad.step() if 'IW' in method and epoch > start_epoch: pseudo_target_label /= args.batch_size * num_iter cov_mat /= args.batch_size * num_iter # Recompute the importance weight by solving a QP. model.im_weights_update(source_label_distribution, pseudo_target_label.cpu().detach().numpy(), cov_mat.cpu().detach().numpy(), args.device) current_weights = [ round(x, 4) for x in model.im_weights.data.cpu().numpy().flatten() ] write_list(out_wei_file, [ np.linalg.norm(current_weights - true_weights.cpu().numpy().flatten()) ] + current_weights) print( np.linalg.norm(current_weights - true_weights.cpu().numpy().flatten()), current_weights)
def train(config): ## set pre-process prep_dict = {} prep_config = config["prep"] prep_dict["source"] = prep.image_train(**config["prep"]['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) if prep_config["test_10crop"]: prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params']) else: prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data dsets = {} dset_loaders = {} data_config = config["data"] train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \ transform=prep_dict["source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \ transform=prep_dict["target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) if prep_config["test_10crop"]: for i in range(10): dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"][i]) for i in range(10)] dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \ shuffle=False, num_workers=4) for dset in dsets['test']] else: dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=4) class_num = config["network"]["params"]["class_num"] ## set base network net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.cuda() with torch.no_grad(): cluster_data_loader = {} cluster_data_loader["source"] = DataLoader(dsets["source"], batch_size=100, \ shuffle=True, num_workers=0, drop_last=True) cluster_data_loader["target"] = DataLoader(dsets["source"], batch_size=100, \ shuffle=True, num_workers=0, drop_last=True) ## add additional network for some methods if config["loss"]["random"]: random_layer = network.RandomLayer([base_network.output_num(), class_num], config["loss"]["random_dim"]) ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None ad_net = network.AdversarialNetwork(base_network.output_num() * class_num, 1024) if config["loss"]["random"]: random_layer.cuda() ad_net = ad_net.cuda() parameter_list = base_network.get_parameters() + ad_net.get_parameters() ## set optimizer optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) # dset_loaders["ps_target"]=[] ## train len_train_source = len(dset_loaders["source"]) # len_train_target = len(dset_loaders["ps_target"]) transfer_loss_value = classifier_loss_value = total_loss_value = 0.0 best_acc = 0.0 for i in range(config["num_iterations"]): lamb = adaptation_factor((i+1)/10000) cls_lamb = adaptation_factor(5*(i+1)/10000) epoch = int(i / len_train_source) if i% len_train_source ==0: testing = True pl_update=True print_loss =True # print("epoch: {} ".format(int(i / len_train_source))) if epoch % 5 ==0 and pl_update: pl_update= False # del dset_loaders["ps_target"] pseudo_labeled_targets,target_g_ctr, source_g_ctr = pseudo_labeling(base_network, cluster_data_loader, class_num) global_source_ctr = source_g_ctr.detach_() global_target_ctr = target_g_ctr.detach_() if len(pseudo_labeled_targets["label_list"]) !=0: print("new pl at epoch {}".format(epoch)) pseudo_dataset = PS_ImageList(pseudo_labeled_targets, transform=prep_dict["target"]) dset_loaders["ps_target"] = DataLoader(pseudo_dataset, batch_size=train_bs, \ shuffle=False, num_workers=0, drop_last=True) len_train_target = len(dset_loaders["ps_target"]) else: print("no pl at epoch {}".format(epoch)) # print("pseudo labeling done") # print(i) # if i % config["test_interval"] == config["test_interval"] - 1: if epoch % 5 ==0 and testing and i>0: base_network.train(False) temp_acc,v_loss = image_classification_test(dset_loaders, \ base_network, test_10crop=prep_config["test_10crop"]) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str) testing=False now = datetime.now() current_time = now.strftime("%H:%M:%S") print("epoch: {} ".format(int(i / len_train_source))) print("time: {} ".format(current_time)) print("best acc: {} ".format(best_acc)) print("loss: {} ".format(v_loss)) print("adaptation rate : {}".format(lamb)) print("learning rare : {} {} {} {}".format(optimizer.param_groups[0]["lr"],optimizer.param_groups[1]["lr"],optimizer.param_groups[2]["lr"],optimizer.param_groups[3]["lr"])) print("------------") if i % config["snapshot_interval"] == 0: torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \ "iter_{:05d}_model.pth.tar".format(i))) loss_params = config["loss"] ## train one iter base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() ### if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: # print(i,len_train_target) iter_target = iter(dset_loaders["ps_target"]) try: inputs_source, labels_source, _ = iter_source.next() inputs_target, labels_target = iter_target.next() except StopIteration: iter_target = iter(dset_loaders["ps_target"]) inputs_target, labels_target = iter_target.next() inputs_source, inputs_target, labels_source, labels_target = inputs_source.cuda(), inputs_target.cuda(), labels_source.cuda(), labels_target.cuda() features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) ##class_aware batch_source_centroids = utils.get_batch_centers(features_source, labels_source, class_num) batch_target_centroids = utils.get_batch_centers(features_target,labels_target, class_num) # if i==0: # global_source_ctr = batch_source_centroids # global_target_ctr = batch_target_centroids # if i>0: batch_source_centroids = ctr_adapt_factor* global_source_ctr + (1- ctr_adapt_factor) * batch_source_centroids batch_target_centroids = ctr_adapt_factor * global_target_ctr + (1 - ctr_adapt_factor) * batch_target_centroids global_source_ctr = batch_source_centroids.clone().detach_() global_target_ctr = batch_target_centroids.clone().detach_() # # global_source_ctr = global_source_ctr.cpu().data.numpy() # global_target_ctr.detach_() # ctr_alignment_loss = utils.cosine_distance(global_source_ctr,global_target_ctr,cross=False) # source_p2c_Distances = 0 - utils.cosine_distance(features_source, global_source_ctr, cross=True) # # target_p2c_Distances = 0 - utils.cosine_distance(features_target, global_target_ctr, cross=True) # # # # zero_ctrs_s = torch.unique(torch.where(global_source_ctr==0)[0]) # zero_ctrs_t = torch.unique(torch.where(global_target_ctr == 0)[0]) alignment_index = [] identity = np.eye(class_num) ctr_alignment_count =0 pos = [] post = [] neg =[] negt =[] index_s = np.empty([0,1]) index_t = np.empty([0,1]) itt=0 triplets ={} # with torch.no_grad(): labels = labels_source.cpu().data.numpy() labelt = labels_target.cpu().data.numpy() # zero_ctrs_s = zero_ctrs_s.cpu().data.numpy() # zero_ctrs_t = zero_ctrs_t.cpu().data.numpy() #####npair # labels = labels.cpu().data.numpy() n_pairs = [] for label in set(labels): label_mask = (labels == label) label_indices = np.where(label_mask)[0] if len(label_indices) < 1: continue anchor = np.random.choice(label_indices, 1, replace=False) n_pairs.append([anchor, np.array([label])]) n_pairs = np.array(n_pairs) n_negatives = [] for i in range(len(n_pairs)): negative = np.concatenate([n_pairs[:i, 1], n_pairs[i + 1:, 1]]) n_negatives.append(negative) n_negatives = np.array(n_negatives) n_pairs_s = torch.LongTensor(n_pairs) n_neg_s = torch.LongTensor(n_negatives) n_pairs = [] for label in set(labelt): label_mask = (labelt == label) label_indices = np.where(label_mask)[0] if len(label_indices) < 1: continue anchor = np.random.choice(label_indices, 1, replace=False) n_pairs.append([anchor, np.array([label])]) n_pairs = np.array(n_pairs) n_negatives = [] for i in range(len(n_pairs)): negative = np.concatenate([n_pairs[:i, 1], n_pairs[i + 1:, 1]]) n_negatives.append(negative) n_negatives = np.array(n_negatives) n_pairs_t = torch.LongTensor(n_pairs) n_neg_t = torch.LongTensor(n_negatives) # return torch.LongTensor(n_pairs), torch.LongTensor(n_negatives) ##### for it in range(class_num): label_mask = (labels == it) label_maskt = (labelt == it) idx = np.where(label_mask)[0] idxt = np.where(label_maskt)[0] # idx = torch.flatten(torch.nonzero(labels_source== torch.tensor(it).cuda())) if len(idx) !=0: index_s =np.append(index_s,idx) pos += [it for cc in range(len(idx))] mask = 1- identity[it,:] neg_id = np.nonzero(mask.flatten())[0].flatten() # neg_idx = np.where(np.in1d(neg_id,zero_ctrs_s)!=True)[0] neg += [[neg_id] for cc in range(len(idx))] if len(idxt) !=0: index_t = np.append(index_t, idxt) post += [it for cc in range(len(idxt))] maskt = 1- identity[it,:] neg_idt = np.nonzero(maskt.flatten())[0].flatten() # neg_idxt = np.where(np.in1d(neg_idt, zero_ctrs_t))[0] negt += [[neg_idt] for cc in range(len(idxt))] # negt += [[neg_idt] for cc in range(len(idxt))] # alignment_ctr_idx =idx[torch.nonzero(torch.where(idx ==idxt, idx,0))] if len(idx) != 0 and len(idxt) !=0: ctr_alignment_count +=1 alignment_index +=[it] # alignment_loss +=[utils.cosine_distance(batch_source_centroids[it], batch_source_centroids[it], cross=False)] # tempp = torch.cat(source_loss,0) # posetives_s = torch.cat(pos, dim=0) # negatives_s = torch.cat(neg, dim=0) # posetives_t = torch.cat(post, dim=0) # negatives_t = torch.cat(negt, dim=0) # a_i = torch.LongTensor(index_s.flatten()).cuda() # a_p = torch.LongTensor(pos).cuda() # a_n = torch.LongTensor(neg).cuda() ctr_alignment_loss =0 anchors_s = features_source[index_s.flatten(),:] positive_s = global_source_ctr[pos,:] negative_s = global_source_ctr[neg].squeeze(1) # n_pairs_s = n_pairs_s.cuda().squeeze(2) # n_neg_s = n_neg_s.cuda().squeeze(2) # anchors_s = features_source[n_pairs_s[:, 0]] # positive_s = global_source_ctr[n_pairs_s[:, 1]] # negative_s = global_source_ctr[n_neg_s] # # n_pairs_t = n_pairs_t.cuda().squeeze(2) # # n_neg_t = n_neg_t.cuda().squeeze(2) # anchors_t = features_source[n_pairs_t[:, 0]] # positive_t = global_source_ctr[n_pairs_t[:, 1]] # negative_t = global_source_ctr[n_neg_t] # anchors_s.retain_graph=True # positive_s.retain_graph=True # negative_s.retain_graph=True anchors_t = features_target[index_t.flatten(), :] positive_t = global_target_ctr[post, :] negative_t = global_target_ctr[negt].squeeze(1) # FAT_loss = torch.empty([],requires_grad=True) # FAT_loss.requires_grad = True # FAT_loss.retain_grad() # nfat_s = Variable(n_pair_loss(anchors_s,positive_s, negative_s,class_num,train_bs)) # nfat_t = Variable(n_pair_loss(anchors_t,positive_t, negative_t,class_num,train_bs)) # FAT_loss.requires_grad = True # FAT_loss.retain_grad() FAT_loss = n_pair_loss(anchors_s,positive_s, negative_s,class_num,train_bs) + n_pair_loss(anchors_t,positive_t, negative_t,class_num,train_bs)/2 if len(alignment_index) != 0: ctr_alignment_loss = torch.sum(utils.cosine_distance(batch_source_centroids[alignment_index], batch_target_centroids[alignment_index], cross=False))#/ctr_alignment_count # source_batch_FAT_Loss = torch.mean(torch.cat(source_loss,0), 0)/class_num # target_batch_FAT_Loss = torch.mean(torch.cat(target_loss,0),0)/class_num # # FAT_loss = source_batch_FAT_Loss.add(target_batch_FAT_Loss) ## # print("train loss: ", FAT_loss) # ctr_alignment_loss.grad_required =True # ctr_alignment_loss.retain_grad() features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) if config['method'] == 'CDAN+E': entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) elif config['method'] == 'CDAN': transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer) elif config['method'] == 'DANN': transfer_loss = loss.DANN(features, ad_net) else: raise ValueError('Method cannot be recognized.') classifier_loss = nn.CrossEntropyLoss()(outputs_source/(2), labels_source) total_loss = loss_params["trade_off"] * (transfer_loss) + classifier_loss if lamb >.1: cls_lamb = 1.0 else: cls_lamb = 10*lamb # total_loss = lamb * ( FAT_loss + 10*ctr_alignment_loss) + (transfer_loss) + cls_lamb*classifier_loss # total_loss =transfer_loss + lamb * (FAT_loss + ctr_alignment_loss) + classifier_loss # FAT_loss.backward(retain_graph=True) # optimizer.zero_grad() total_loss.backward() optimizer.step() # my_lr_scheduler.step() if epoch % 5 ==0 and print_loss: print("fat loss ", FAT_loss)#.grad_fn, FAT_loss.requires_grad) print("ctr align: ", ctr_alignment_loss) print("tot: ", total_loss) print("clss: ",classifier_loss) print("trs: ", transfer_loss) print("++++++++++++++++++++++++end of epoch++++++++++++++++++++") print_loss =False
def train(config): #################################################### # Data setting #################################################### prep_dict = {} # 데이터 전처리 transforms 부분 prep_dict["source"] = prep.image_train(**config['prep']['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) prep_dict["test"] = prep.image_test(**config['prep']['params']) dsets = {} dsets["source"]= datasets.ImageFolder(config['s_dset_path'], transform=prep_dict["source"]) dsets["target"]= datasets.ImageFolder(config['t_dset_path'], transform=prep_dict['target']) dsets['test']=datasets.ImageFolder(config['t_dset_path'],transform=prep_dict['test']) data_config = config["data"] train_source_bs = data_config["source"]["batch_size"] #원본은 source와 target 모두 source train bs로 설정되었는데 이를 수정함 train_target_bs = data_config['target']['batch_size'] test_bs = data_config["test"]["batch_size"] dset_loaders = {} dset_loaders["source"]=DataLoader(dsets["source"], batch_size=train_source_bs, shuffle=True, num_workers=4, drop_last=True) # 원본은 drop_last=True dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_target_bs, shuffle=True, num_workers=4, drop_last=True) dset_loaders['test'] = DataLoader(dsets['test'], batch_size=test_bs, shuffle=False, num_workers=4, drop_last=False) #################################################### # Network Setting #################################################### class_num = config["network"]['params']['class_num'] net_config = config["network"] """ config['network'] = {'name': network.ResNetFC, 'params': {'resnet_name': args.net, 'use_bottleneck': True, 'bottleneck_dim': 256, 'new_cls': True, 'class_num': args.class_num} } """ base_network = net_config["name"](**net_config["params"]) #network.ResNetFC base_network = base_network.cuda() # ResNetFC(Resnet, True, 256, True, 12) if config["loss"]["random"]: random_layer = network.RandomLayer([base_network.output_num(), class_num], config["loss"]["random_dim"] ) random_layer.cuda() ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None ad_net = network.AdversarialNetwork(base_network.output_num()*class_num, 1024) ad_net = ad_net.cuda() parameter_list = base_network.get_parameters() + ad_net.get_parameters() #################################################### # Env Setting #################################################### #gpus = config['gpu'].split(',') #if len(gpus) > 1 : #ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) #base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) #################################################### # Optimizer Setting #################################################### optimizer_config = config['optimizer'] optimizer = optimizer_config["type"](parameter_list, **(optimizer_config["optim_params"])) # optim.SGD ''' config['optimizer'] = {'type': optim.SGD, 'optim_params': {'lr': args.lr, 'momentum': 0.9, 'weight_decay': 0.0005, 'nestrov': True}, 'lr_type': "inv", 'lr_param': {"lr": args.lr, 'gamma': 0.001, 'power': 0.75 } ''' param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group['lr']) schedule_param = optimizer_config['lr_param'] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] #################################################### # Train #################################################### len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) transfer_loss_value = 0.0 classifier_loss_value = 0.0 total_loss_value = 0.0 best_acc = 0.0 for i in range(config["num_iterations"]): # num_iterations = batch 수 sys.stdout.write("Iteration : {} \r".format(i)) sys.stdout.flush() loss_params = config["loss"] base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, labels_target = iter_target.next() inputs_source, labels_source = inputs_source.cuda(), labels_source.cuda() inputs_target = inputs_target.cuda() features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) features = torch.cat((features_source, features_target), dim = 0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) if config['method'] == 'CDAN+E': entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) elif config['method'] == 'CDAN': transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer) elif config['method'] == 'DANN': pass # 나중에 정리하기 else: raise ValueError('Method cannot be recognized') classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss total_loss.backward() optimizer.step() #################################################### # Test #################################################### if i % config["test_interval"] == config["test_interval"] - 1: # test interval 마다 base_network.train(False) temp_acc = image_classification_test(dset_loaders, base_network) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model ACC = round(best_acc, 2) * 100 torch.save(best_model, os.path.join(config["output_path"], "iter_{}_model.pth.tar".format(ACC))) log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str)
def train(config): ## Define start time start_time = time.time() ## set pre-process prep_dict = {} prep_config = config["prep"] prep_dict["source"] = prep.image_train(**config["prep"]['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data print("Preparing data", flush=True) dsets = {} dset_loaders = {} data_config = config["data"] train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] root_folder = data_config["root_folder"] dsets["source"] = ImageList(open(osp.join(root_folder, data_config["source"]["list_path"])).readlines(), \ transform=prep_dict["source"], root_folder=root_folder, ratios=config["ratios_source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["target"] = ImageList(open(osp.join(root_folder, data_config["target"]["list_path"])).readlines(), \ transform=prep_dict["target"], root_folder=root_folder, ratios=config["ratios_target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["test"] = ImageList(open( osp.join(root_folder, data_config["test"]["list_path"])).readlines(), transform=prep_dict["test"], root_folder=root_folder, ratios=config["ratios_test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=4) test_path = os.path.join(root_folder, data_config["test"]["dataset_path"]) if os.path.exists(test_path): print('Found existing dataset for test', flush=True) with open(test_path, 'rb') as f: [test_samples, test_labels] = pickle.load(f) test_labels = torch.LongTensor(test_labels).to(config["device"]) else: print('Missing test dataset', flush=True) print('Building dataset for test and writing to {}'.format(test_path), flush=True) dset_test = ImageList(open( osp.join(root_folder, data_config["test"]["list_path"])).readlines(), transform=prep_dict["test"], root_folder=root_folder, ratios=config['ratios_test']) loaded_dset_test = LoadedImageList(dset_test) test_samples, test_labels = loaded_dset_test.samples.numpy( ), loaded_dset_test.targets.numpy() with open(test_path, 'wb') as f: pickle.dump([test_samples, test_labels], f) class_num = config["network"]["params"]["class_num"] test_samples, test_labels = sample_ratios(test_samples, test_labels, config['ratios_test']) # compute labels distribution on the source and target domain source_label_distribution = np.zeros((class_num)) for img in dsets["source"].imgs: source_label_distribution[img[1]] += 1 print("Total source samples: {}".format(np.sum(source_label_distribution)), flush=True) print("Source samples per class: {}".format(source_label_distribution), flush=True) source_label_distribution /= np.sum(source_label_distribution) print("Source label distribution: {}".format(source_label_distribution), flush=True) target_label_distribution = np.zeros((class_num)) for img in dsets["target"].imgs: target_label_distribution[img[1]] += 1 print("Total target samples: {}".format(np.sum(target_label_distribution)), flush=True) print("Target samples per class: {}".format(target_label_distribution), flush=True) target_label_distribution /= np.sum(target_label_distribution) print("Target label distribution: {}".format(target_label_distribution), flush=True) mixture = (source_label_distribution + target_label_distribution) / 2 jsd = (scipy.stats.entropy(source_label_distribution, qk=mixture) \ + scipy.stats.entropy(target_label_distribution, qk=mixture)) / 2 print("JSD : {}".format(jsd), flush=True) test_label_distribution = np.zeros((class_num)) for img in test_labels: test_label_distribution[int(img.item())] += 1 print("Test samples per class: {}".format(test_label_distribution), flush=True) test_label_distribution /= np.sum(test_label_distribution) print("Test label distribution: {}".format(test_label_distribution), flush=True) write_list(config["out_wei_file"], [round(x, 4) for x in test_label_distribution]) write_list(config["out_wei_file"], [round(x, 4) for x in source_label_distribution]) write_list(config["out_wei_file"], [round(x, 4) for x in target_label_distribution]) true_weights = torch.tensor( target_label_distribution / source_label_distribution, dtype=torch.float, requires_grad=False)[:, None].to(config["device"]) print("True weights : {}".format(true_weights[:, 0].cpu().numpy())) config["out_wei_file"].write(str(jsd) + "\n") ## set base network net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.to(config["device"]) ## add additional network for some methods if config["loss"]["random"]: random_layer = network.RandomLayer( [base_network.output_num(), class_num], config["loss"]["random_dim"]) ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None if 'CDAN' in config['method']: ad_net = network.AdversarialNetwork( base_network.output_num() * class_num, 1024) else: ad_net = network.AdversarialNetwork(base_network.output_num(), 1024) if config["loss"]["random"]: random_layer.to(config["device"]) ad_net = ad_net.to(config["device"]) parameter_list = ad_net.get_parameters() + base_network.get_parameters() parameter_list[-1]["lr_mult"] = config["lr_mult_im"] ## set optimizer optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] # Maintain two quantities for the QP. cov_mat = torch.tensor(np.zeros((class_num, class_num), dtype=np.float32), requires_grad=False).to(config["device"]) pseudo_target_label = torch.tensor(np.zeros((class_num, 1), dtype=np.float32), requires_grad=False).to( config["device"]) # Maintain one weight vector for BER. class_weights = torch.tensor(1.0 / source_label_distribution, dtype=torch.float, requires_grad=False).to(config["device"]) gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) transfer_loss_value = classifier_loss_value = total_loss_value = 0.0 best_acc = 0.0 print("Preparations done in {:.0f} seconds".format(time.time() - start_time), flush=True) print("Starting training for {} iterations using method {}".format( config["num_iterations"], config['method']), flush=True) start_time_test = start_time = time.time() for i in range(config["num_iterations"]): if i % config["test_interval"] == config["test_interval"] - 1: base_network.train(False) temp_acc = image_classification_test_loaded( test_samples, test_labels, base_network) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc log_str = " iter: {:05d}, sec: {:.0f}, class: {:.5f}, da: {:.5f}, precision: {:.5f}".format( i, time.time() - start_time_test, classifier_loss_value, transfer_loss_value, temp_acc) config["out_log_file"].write(log_str + "\n") config["out_log_file"].flush() print(log_str, flush=True) if 'IW' in config['method']: current_weights = [ round(x, 4) for x in base_network.im_weights.data.cpu().numpy().flatten() ] # write_list(config["out_wei_file"], current_weights) print(current_weights, flush=True) start_time_test = time.time() if i % 500 == -1: print("{} iterations in {} seconds".format( i, time.time() - start_time), flush=True) loss_params = config["loss"] ## train one iter base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() t = time.time() if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, label_source = iter_source.next() inputs_target, _ = iter_target.next() inputs_source, inputs_target, label_source = inputs_source.to( config["device"]), inputs_target.to( config["device"]), label_source.to(config["device"]) features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) if 'IW' in config['method']: ys_onehot = torch.zeros(train_bs, class_num).to(config["device"]) ys_onehot.scatter_(1, label_source.view(-1, 1), 1) # Compute weights on source data. if 'ORACLE' in config['method']: weights = torch.mm(ys_onehot, true_weights) else: weights = torch.mm(ys_onehot, base_network.im_weights) source_preds, target_preds = outputs[:train_bs], outputs[train_bs:] # Compute the aggregated distribution of pseudo-label on the target domain. pseudo_target_label += torch.sum(F.softmax(target_preds, dim=1), dim=0).view(-1, 1).detach() # Update the covariance matrix on the source domain as well. cov_mat += torch.mm( F.softmax(source_preds, dim=1).transpose(1, 0), ys_onehot).detach() if config['method'] == 'CDAN-E': classifier_loss = nn.CrossEntropyLoss()(outputs_source, label_source) entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) total_loss = loss_params["trade_off"] * \ transfer_loss + classifier_loss elif 'IWCDAN-E' in config['method']: classifier_loss = torch.mean( nn.CrossEntropyLoss(weight=class_weights, reduction='none') (outputs_source, label_source) * weights) / class_num entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer, weights=weights, device=config["device"]) total_loss = loss_params["trade_off"] * \ transfer_loss + classifier_loss elif config['method'] == 'CDAN': classifier_loss = nn.CrossEntropyLoss()(outputs_source, label_source) transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer) total_loss = loss_params[ "trade_off"] * transfer_loss + classifier_loss elif 'IWCDAN' in config['method']: classifier_loss = torch.mean( nn.CrossEntropyLoss(weight=class_weights, reduction='none') (outputs_source, label_source) * weights) / class_num transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer, weights=weights) total_loss = loss_params["trade_off"] * \ transfer_loss + classifier_loss elif config['method'] == 'DANN': classifier_loss = nn.CrossEntropyLoss()(outputs_source, label_source) transfer_loss = loss.DANN(features, ad_net, config["device"]) total_loss = loss_params["trade_off"] * \ transfer_loss + classifier_loss elif 'IWDAN' in config['method']: classifier_loss = torch.mean( nn.CrossEntropyLoss(weight=class_weights, reduction='none') (outputs_source, label_source) * weights) / class_num transfer_loss = loss.IWDAN(features, ad_net, weights) total_loss = loss_params["trade_off"] * \ transfer_loss + classifier_loss elif config['method'] == 'NANN': classifier_loss = nn.CrossEntropyLoss()(outputs_source, label_source) total_loss = classifier_loss else: raise ValueError('Method cannot be recognized.') total_loss.backward() optimizer.step() transfer_loss_value = 0 if config[ 'method'] == 'NANN' else transfer_loss.item() classifier_loss_value = classifier_loss.item() total_loss_value = transfer_loss_value + classifier_loss_value if ('IW' in config['method'] ) and i % (config["dataset_mult_iw"] * len_train_source ) == config["dataset_mult_iw"] * len_train_source - 1: pseudo_target_label /= train_bs * \ len_train_source * config["dataset_mult_iw"] cov_mat /= train_bs * len_train_source * config["dataset_mult_iw"] print(i, np.sum(cov_mat.cpu().detach().numpy()), train_bs * len_train_source) # Recompute the importance weight by solving a QP. base_network.im_weights_update( source_label_distribution, pseudo_target_label.cpu().detach().numpy(), cov_mat.cpu().detach().numpy(), config["device"]) current_weights = [ round(x, 4) for x in base_network.im_weights.data.cpu().numpy().flatten() ] write_list(config["out_wei_file"], [ np.linalg.norm(current_weights - true_weights.cpu().numpy().flatten()) ] + current_weights) print( np.linalg.norm(current_weights - true_weights.cpu().numpy().flatten()), current_weights) cov_mat[:] = 0.0 pseudo_target_label[:] = 0.0 return best_acc
def train(args): # prepare data dsets = {} dset_loaders = {} dsets["source"] = ImageList(open(args.source_list).readlines(), \ transform=image_train()) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=args.batch_size, \ shuffle=True, num_workers=4, drop_last=True) dsets["target"] = ImageList(open('data/{}/pseudo_list/{}_{}_list.txt' ''.format(args.dataset,args.source,args.target)).readlines(), transform=image_train(),pseudo=True) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=args.batch_size, \ shuffle=True, num_workers=4, drop_last=True) dsets["test"] = ImageList(open(args.target_list).readlines(), \ transform=image_test()) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=2 * args.batch_size, \ shuffle=False, num_workers=4) #model model = network.ResNet(class_num=args.num_class).cuda() adv_net = network.AdversarialNetwork(in_feature=model.output_num(),hidden_size=1024,max_iter=2000).cuda() parameter_classifier = [model.get_parameters()[2]] parameter_feature = model.get_parameters()[0:2] + adv_net.get_parameters() optimizer_classifier = torch.optim.SGD(parameter_classifier,lr=args.lr,momentum=0.9,weight_decay=0.005) optimizer_feature = torch.optim.SGD(parameter_feature,lr=args.lr,momentum=0.9,weight_decay=0) gpus = args.gpu_id.split(',') if len(gpus) > 1: adv_net = nn.DataParallel(adv_net, device_ids=[int(i) for i in gpus]) model = nn.DataParallel(model, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) best_acc = 0.0 best_model = copy.deepcopy(model) Cs_memory = torch.zeros(args.num_class, 256).cuda() Ct_memory = torch.zeros(args.num_class, 256).cuda() for i in range(args.max_iter): if i % args.test_interval == args.test_interval - 1: model.train(False) temp_acc = image_classification_test(dset_loaders, model) if temp_acc > best_acc: best_acc = temp_acc best_model = copy.deepcopy(model) log_str = "\n iter: {:05d}, \t precision: {:.4f},\t best_acc:{:.4f}".format(i, temp_acc, best_acc) args.log_file.write(log_str) args.log_file.flush() print(log_str) if i % args.snapshot_interval == args.snapshot_interval -1: if not os.path.exists('snapshot'): os.mkdir('snapshot') if not os.path.exists('snapshot/save'): os.mkdir('snapshot/save') torch.save(best_model,'snapshot/save/best_model.pk') model.train(True) adv_net.train(True) optimizer_classifier = lr_schedule.inv_lr_scheduler(optimizer_classifier,i) optimizer_feature = lr_schedule.inv_lr_scheduler(optimizer_feature, i) if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, pseudo_labels_target, weights = iter_target.next() inputs_source, labels_source = inputs_source.cuda(), labels_source.cuda() inputs_target, pseudo_labels_target = inputs_target.cuda(), pseudo_labels_target.cuda() weights = weights.type(torch.Tensor).cuda() features_source, outputs_source = model(inputs_source) features_target, outputs_target = model(inputs_target) features = torch.cat((features_source, features_target), dim=0) source_class_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) adv_loss = utils.loss_adv(features,adv_net) H = torch.mean(utils.Entropy(F.softmax(outputs_target, dim=1))) target_robust_loss = utils.robust_pseudo_loss(outputs_target,pseudo_labels_target,weights) classifier_loss = source_class_loss + target_robust_loss optimizer_classifier.zero_grad() classifier_loss.backward(retain_graph=True) optimizer_classifier.step() if args.baseline == 'MSTN': lam = network.calc_coeff(i,max_iter=2000) elif args.baseline =='DANN': lam = 0.0 pseu_labels_target = torch.argmax(outputs_target, dim=1) loss_sm, Cs_memory, Ct_memory = utils.SM(features_source, features_target, labels_source, pseu_labels_target, Cs_memory, Ct_memory) feature_loss = classifier_loss + adv_loss + lam*loss_sm + lam*H optimizer_feature.zero_grad() feature_loss.backward() optimizer_feature.step() print('step:{: d},\t,source_class_loss:{:.4f},\t,target_robust_loss:{:.4f}' ''.format(i, source_class_loss.item(),target_robust_loss.item())) Cs_memory.detach_() Ct_memory.detach_() return best_acc, best_model
def train_distill(teacher, args): # prepare data dsets = {} dset_loaders = {} dsets["source"] = ImageList(open(args.source_list).readlines(), \ transform=image_train()) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=args.batch_size, \ shuffle=True, num_workers=2, drop_last=True) dsets["target"] = ImageList(open(args.target_list).readlines(), \ transform=image_train(), params=args) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=args.batch_size, \ shuffle=True, num_workers=2, drop_last=True) dsets["test"] = ImageList(open(args.target_list).readlines(), \ transform=image_test()) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=2 * args.batch_size, \ shuffle=False, num_workers=2) #model model = network.ResNet(class_num=args.num_class).cuda() adv_net = network.AdversarialNetwork(in_feature=model.output_num(),hidden_size=1024, max_iter=args.max_iter).cuda() parameter_list = model.get_parameters() + adv_net.get_parameters() optimizer = torch.optim.SGD(parameter_list,lr=args.lr,momentum=0.9,weight_decay=0.005) # model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0) gpus = args.gpu_id.split(',') if len(gpus) > 1: adv_net = nn.DataParallel(adv_net, device_ids=[int(i) for i in gpus]) model = nn.DataParallel(model, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) best_acc = 0.0 best_model = copy.deepcopy(model) print_interval = (args.test_interval // 10) nt_cent = utils.NTXentLoss('cuda', args.batch_size, 0.2, True) Cs_memory = torch.zeros(args.num_class, 256).cuda() Ct_memory = torch.zeros(args.num_class, 256).cuda() max_batch = 100 queue_size = args.batch_size * max_batch queue_data = [torch.randn(queue_size, 256).cuda(), torch.randn(queue_size, args.num_class).cuda()] queue_data_w = [torch.randn(queue_size, 256).cuda(), torch.randn(queue_size, args.num_class).cuda()] # queue_data = [torch.randn(queue_size, 256).cuda(), torch.randn(queue_size, 256).cuda()] queue_labels = [torch.ones(queue_size).cuda() * (args.num_class+1), torch.ones(queue_size).cuda() * (args.num_class+1)] queue_ptr = torch.zeros(1, dtype=torch.long) queue_weight = np.power(np.linspace(.0, 1.0, max_batch), 3) queue_weight = np.repeat(queue_weight, args.batch_size) best_ema_acc = 0.0 for i in range(args.max_iter): if i % args.test_interval == args.test_interval - 1: model.train(False) temp_acc = image_classification_test(dset_loaders, model) if temp_acc > best_acc: best_acc = temp_acc best_model = copy.deepcopy(model) log_str = "\niter: {:05d}, \t precision: {:.4f},\t best_acc:{:.4f}".format(i, temp_acc, best_acc) args.log_file.write(log_str) args.log_file.flush() print(log_str) temp_acc = image_classification_test(dset_loaders, teacher) if temp_acc > best_ema_acc: best_ema_acc = temp_acc # best_model = copy.deepcopy(model) log_str = "\niter: {:05d}, \t precision: {:.4f},\t best_ema_acc:{:.4f}".format(i, temp_acc, best_ema_acc) args.log_file.write(log_str) args.log_file.flush() print(log_str) # if i % args.snapshot_interval == args.snapshot_interval -1: # if not os.path.exists(args.save_dir): # os.mkdir(args.save_dir) # torch.save(best_model,os.path.join(args.save_dir, 'initial_model.pk')) model.train(True) adv_net.train(True) teacher.train(False) optimizer = lr_schedule.inv_lr_scheduler(optimizer,i) if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, _, inputs_target_mosaic_w, inputs_target_mosaic_s, labels_target = iter_target.next() inputs_source, inputs_target, labels_source = inputs_source.cuda(), inputs_target.cuda(), labels_source.cuda() inputs_target_mosaic_w, inputs_target_mosaic_s = inputs_target_mosaic_w.cuda(), inputs_target_mosaic_s.cuda() features_source, outputs_source = model(inputs_source) features_target, outputs_target = model(inputs_target) features = torch.cat((features_source, features_target), dim=0) with torch.no_grad(): features_target_teacher, outputs_target_teacher = teacher(inputs_target) adv_loss = utils.loss_adv(features,adv_net) H = torch.mean(utils.Entropy(F.softmax(outputs_target, dim=1))) if args.baseline == 'MSTN': lam = network.calc_coeff(i) elif args.baseline =='DANN': lam = 0.0 prob_max, pseu_labels_target = torch.max(F.softmax(outputs_target, dim=1), dim=1) loss_sm, Cs_memory, Ct_memory = utils.SM(features_source, features_target, labels_source, pseu_labels_target, Cs_memory, Ct_memory) # classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) classifier_loss = 4*utils.cross_entropy_with_logits(outputs_target / 4.0, F.softmax(outputs_target_teacher / 4.0, dim=1)) + nn.CrossEntropyLoss()(outputs_source, labels_source) total_loss = classifier_loss + lam * loss_sm + adv_loss + network.calc_coeff((i-100), high=0.1, max_iter=100)*H prob_max, pseu_labels_target = torch.max(F.softmax(outputs_target, dim=1), dim=1) optimizer.zero_grad() total_loss.backward() optimizer.step() optimizer.zero_grad() mosaic_loss_target = torch.zeros(1) if i < args.max_iter // 5 * 2: alpha = 0.0 else: alpha = 0.5 with _disable_tracking_bn_stats(model): mosaic_features_target_w, mosaic_outputs_target_w = model(inputs_target_mosaic_w) mosaic_features_target_s, mosaic_outputs_target_s = model(inputs_target_mosaic_s) with torch.no_grad(): features_list_w = [mosaic_features_target_w, F.softmax(mosaic_outputs_target_w, dim=1)] features_target_, outputs_target_ = model(inputs_target) outputs_target = alpha * outputs_target_ + (1. - alpha) * outputs_target_teacher prob_max, pseu_labels_target = torch.max(F.softmax(outputs_target, dim=1), dim=1) features_list = [features_target_, F.softmax(outputs_target, dim=1)] labels_list = [pseu_labels_target, pseu_labels_target] utils.rightshift(queue_weight, args.batch_size) for j in range(len(features_list)): queue_data[j][queue_ptr:queue_ptr+args.batch_size, :] = features_list[j] queue_data_w[j][queue_ptr:queue_ptr+args.batch_size, :] = features_list_w[j] queue_labels[j][queue_ptr:queue_ptr+args.batch_size] = labels_list[j] pre_ptr = int(queue_ptr) ptr = ((i+1) % max_batch) * args.batch_size queue_ptr[0] = ptr mosaic_loss_target = (nt_cent(queue_data[1].detach(), F.softmax(mosaic_outputs_target_w, dim=1), queue_labels[1], pseu_labels_target.float(), queue_weight, pre_ptr, class_level=False) + 1.*nt_cent(queue_data_w[1].detach(), F.softmax(mosaic_outputs_target_s, dim=1), queue_labels[1], pseu_labels_target.float(), queue_weight, pre_ptr, class_level=False)) * network.calc_coeff(i, high=0.3, max_iter=50) mosaic_loss = mosaic_loss_target * 1.0 # mosaic_loss = utils.cross_entropy_with_logits(mosaic_outputs_target, F.softmax(outputs_target*1.5, dim=1)) * (network.calc_coeff(i, high=0.5, max_iter=2000)) # mosaic_loss += 0.4*(torch.abs(F.softmax(outputs_target, dim=1).detach() - F.softmax(mosaic_outputs_target, dim=1)).sum(1)).mean(0) mosaic_loss.backward() optimizer.step() if i % print_interval == 0: log_str = 'step:{: d},\t,class_loss:{:.4f},\t,adv_loss:{:.4f}\t,mosaic_loss:{:.4f}\t,mean_prob:{:.4f}'.format(i, classifier_loss.item(), adv_loss.item(), mosaic_loss_target.item(),prob_max.mean().item()) print(log_str) args.log_file.write('\n'+log_str) args.log_file.flush() Cs_memory.detach_() Ct_memory.detach_() return best_acc, best_model
def train(config): ## set pre-process prep_dict = {} dsets = {} dset_loaders = {} data_config = config["data"] prep_config = config["prep"] prep_dict["source"] = prep.image_target(**config["prep"]['params']) prep_dict["target"] = prep.image_target(**config["prep"]['params']) prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \ transform=prep_dict["source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \ transform=prep_dict["target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=4, drop_last=True) dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=4) ## set base network class_num = config["network"]["params"]["class_num"] net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.cuda() ## add additional network for some methods ad_net = network.AdversarialNetwork(class_num, 1024) ad_net = ad_net.cuda() ## set optimizer parameter_list = base_network.get_parameters() + ad_net.get_parameters() optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] #multi gpu gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel( ad_net, device_ids=[int(i) for i, k in enumerate(gpus)]) base_network = nn.DataParallel( base_network, device_ids=[int(i) for i, k in enumerate(gpus)]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) transfer_loss_value = classifier_loss_value = total_loss_value = 0.0 best_acc = 0.0 for i in range(config["num_iterations"]): #test if i % config["test_interval"] == config["test_interval"] - 1: base_network.train(False) temp_acc = image_classification_test(dset_loaders, base_network, gvbg=config["GVBG"]) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str) #save model if i % config["snapshot_interval"] == 0: torch.save(base_network.state_dict(), osp.join(config["output_path"], \ "iter_{:05d}_model.pth.tar".format(i))) ## train one iter base_network.train(True) ad_net.train(True) loss_params = config["loss"] optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() #dataloader if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) #network inputs_source, labels_source = iter_source.next() inputs_target, _ = iter_target.next() inputs_source, inputs_target, labels_source = inputs_source.cuda( ), inputs_target.cuda(), labels_source.cuda() features_source, outputs_source, focal_source = base_network( inputs_source, gvbg=config["GVBG"]) features_target, outputs_target, focal_target = base_network( inputs_target, gvbg=config["GVBG"]) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) focals = torch.cat((focal_source, focal_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) #loss calculation transfer_loss, mean_entropy, gvbg, gvbd = loss.GVB( [softmax_out, focals], ad_net, network.calc_coeff(i), GVBD=config['GVBD']) classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) total_loss = loss_params[ "trade_off"] * transfer_loss + classifier_loss + config[ "GVBG"] * gvbg + abs(config['GVBD']) * gvbd if i % config["print_num"] == 0: log_str = "iter: {:05d}, transferloss: {:.5f}, classifier_loss: {:.5f}, mean entropy:{:.5f}, gvbg:{:.5f}, gvbd:{:.5f}".format( i, transfer_loss, classifier_loss, mean_entropy, gvbg, gvbd) config["out_file"].write(log_str + "\n") config["out_file"].flush() #print(log_str) total_loss.backward() optimizer.step() torch.save(best_model, osp.join(config["output_path"], "best_model.pth.tar")) return best_acc
def train(config): ## set pre-process prep_dict = {} prep_config = config["prep"] prep_dict["source"] = prep.image_train(**config["prep"]['params']) prep_dict["target"] = prep.image_train(**config["prep"]['params']) if prep_config["test_10crop"]: prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params']) else: prep_dict["test"] = prep.image_test(**config["prep"]['params']) ## prepare data dsets = {} dset_loaders = {} data_config = config["data"] train_bs = data_config["source"]["batch_size"] test_bs = data_config["test"]["batch_size"] dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \ transform=prep_dict["source"]) dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \ shuffle=True, num_workers=0, drop_last=True) dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \ transform=prep_dict["target"]) dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \ shuffle=True, num_workers=0, drop_last=True) if prep_config["test_10crop"]: for i in range(10): dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"][i]) for i in range(10)] dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \ shuffle=False, num_workers=0) for dset in dsets['test']] else: dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \ transform=prep_dict["test"]) dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \ shuffle=False, num_workers=0) class_num = config["network"]["params"]["class_num"] ## set base network net_config = config["network"] base_network = net_config["name"](**net_config["params"]) base_network = base_network.cuda() ## 添加判别器D_s,D_t,生成器G_s2t,G_t2s z_dimension = 256 D_s = network.models["Discriminator"]() D_s = D_s.cuda() G_s2t = network.models["Generator"](z_dimension, 1024) G_s2t = G_s2t.cuda() D_t = network.models["Discriminator"]() D_t = D_t.cuda() G_t2s = network.models["Generator"](z_dimension, 1024) G_t2s = G_t2s.cuda() criterion_GAN = torch.nn.MSELoss() criterion_cycle = torch.nn.L1Loss() criterion_identity = torch.nn.L1Loss() criterion_Sem = torch.nn.L1Loss() optimizer_G = torch.optim.Adam(itertools.chain(G_s2t.parameters(), G_t2s.parameters()), lr=0.0003) optimizer_D_s = torch.optim.Adam(D_s.parameters(), lr=0.0003) optimizer_D_t = torch.optim.Adam(D_t.parameters(), lr=0.0003) fake_S_buffer = ReplayBuffer() fake_T_buffer = ReplayBuffer() classifier_optimizer = torch.optim.Adam(base_network.parameters(), lr=0.0003) ## 添加分类器 classifier1 = net.Net(256, class_num) classifier1 = classifier1.cuda() classifier1_optim = optim.Adam(classifier1.parameters(), lr=0.0003) ## add additional network for some methods if config["loss"]["random"]: random_layer = network.RandomLayer( [base_network.output_num(), class_num], config["loss"]["random_dim"]) ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024) else: random_layer = None ad_net = network.AdversarialNetwork( base_network.output_num() * class_num, 1024) if config["loss"]["random"]: random_layer.cuda() ad_net = ad_net.cuda() parameter_list = base_network.get_parameters() + ad_net.get_parameters() ## set optimizer optimizer_config = config["optimizer"] optimizer = optimizer_config["type"](parameter_list, \ **(optimizer_config["optim_params"])) param_lr = [] for param_group in optimizer.param_groups: param_lr.append(param_group["lr"]) schedule_param = optimizer_config["lr_param"] lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]] gpus = config['gpu'].split(',') if len(gpus) > 1: ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus]) base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus]) ## train len_train_source = len(dset_loaders["source"]) len_train_target = len(dset_loaders["target"]) transfer_loss_value = classifier_loss_value = total_loss_value = 0.0 best_acc = 0.0 for i in range(config["num_iterations"]): if i % config["test_interval"] == config["test_interval"] - 1: base_network.train(False) temp_acc = image_classification_test(dset_loaders, \ base_network, test_10crop=prep_config["test_10crop"]) temp_model = nn.Sequential(base_network) if temp_acc > best_acc: best_acc = temp_acc best_model = temp_model now = datetime.datetime.now() d = str(now.month) + '-' + str(now.day) + ' ' + str( now.hour) + ':' + str(now.minute) + ":" + str(now.second) torch.save( best_model, osp.join( config["output_path"], "{}_to_{}_best_model_acc-{}_{}.pth.tar".format( args.source, args.target, best_acc, d))) log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc) config["out_file"].write(log_str + "\n") config["out_file"].flush() print(log_str) if i % config["snapshot_interval"] == 0: torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \ "{}_to_{}_iter_{:05d}_model_{}.pth.tar".format(args.source, args.target, i, str( datetime.datetime.utcnow())))) loss_params = config["loss"] ## train one iter classifier1.train(True) base_network.train(True) ad_net.train(True) optimizer = lr_scheduler(optimizer, i, **schedule_param) optimizer.zero_grad() if i % len_train_source == 0: iter_source = iter(dset_loaders["source"]) if i % len_train_target == 0: iter_target = iter(dset_loaders["target"]) inputs_source, labels_source = iter_source.next() inputs_target, labels_target = iter_target.next() inputs_source, inputs_target, labels_source = inputs_source.cuda( ), inputs_target.cuda(), labels_source.cuda() # 提取特征 features_source, outputs_source = base_network(inputs_source) features_target, outputs_target = base_network(inputs_target) features = torch.cat((features_source, features_target), dim=0) outputs = torch.cat((outputs_source, outputs_target), dim=0) softmax_out = nn.Softmax(dim=1)(outputs) outputs_source1 = classifier1(features_source.detach()) outputs_target1 = classifier1(features_target.detach()) outputs1 = torch.cat((outputs_source1, outputs_target1), dim=0) softmax_out1 = nn.Softmax(dim=1)(outputs1) softmax_out = (1 - args.cla_plus_weight ) * softmax_out + args.cla_plus_weight * softmax_out1 if config['method'] == 'CDAN+E': entropy = loss.Entropy(softmax_out) transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer) elif config['method'] == 'CDAN': transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer) elif config['method'] == 'DANN': transfer_loss = loss.DANN(features, ad_net) else: raise ValueError('Method cannot be recognized.') classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) # Cycle num_feature = features_source.size(0) # =================train discriminator T real_label = Variable(torch.ones(num_feature)).cuda() fake_label = Variable(torch.zeros(num_feature)).cuda() # 训练生成器 optimizer_G.zero_grad() # Identity loss same_t = G_s2t(features_target.detach()) loss_identity_t = criterion_identity(same_t, features_target) same_s = G_t2s(features_source.detach()) loss_identity_s = criterion_identity(same_s, features_source) # Gan loss fake_t = G_s2t(features_source.detach()) pred_fake = D_t(fake_t) loss_G_s2t = criterion_GAN(pred_fake, labels_source.float()) fake_s = G_t2s(features_target.detach()) pred_fake = D_s(fake_s) loss_G_t2s = criterion_GAN(pred_fake, labels_source.float()) # cycle loss recovered_s = G_t2s(fake_t) loss_cycle_sts = criterion_cycle(recovered_s, features_source) recovered_t = G_s2t(fake_s) loss_cycle_tst = criterion_cycle(recovered_t, features_target) # sem loss pred_recovered_s = base_network.fc(recovered_s) pred_fake_t = base_network.fc(fake_t) loss_sem_t2s = criterion_Sem(pred_recovered_s, pred_fake_t) pred_recovered_t = base_network.fc(recovered_t) pred_fake_s = base_network.fc(fake_s) loss_sem_s2t = criterion_Sem(pred_recovered_t, pred_fake_s) loss_cycle = loss_cycle_tst + loss_cycle_sts weights = args.weight_in_lossG.split(',') loss_G = float(weights[0]) * (loss_identity_s + loss_identity_t) + \ float(weights[1]) * (loss_G_s2t + loss_G_t2s) + \ float(weights[2]) * loss_cycle + \ float(weights[3]) * (loss_sem_s2t + loss_sem_t2s) # 训练softmax分类器 outputs_fake = classifier1(fake_t.detach()) # 分类器优化 classifier_loss1 = nn.CrossEntropyLoss()(outputs_fake, labels_source) classifier1_optim.zero_grad() classifier_loss1.backward() classifier1_optim.step() total_loss = loss_params[ "trade_off"] * transfer_loss + classifier_loss + args.cyc_loss_weight * loss_G total_loss.backward() optimizer.step() optimizer_G.step() ###### Discriminator S ###### optimizer_D_s.zero_grad() # Real loss pred_real = D_s(features_source.detach()) loss_D_real = criterion_GAN(pred_real, real_label) # Fake loss fake_s = fake_S_buffer.push_and_pop(fake_s) pred_fake = D_s(fake_s.detach()) loss_D_fake = criterion_GAN(pred_fake, fake_label) # Total loss loss_D_s = loss_D_real + loss_D_fake loss_D_s.backward() optimizer_D_s.step() ################################### ###### Discriminator t ###### optimizer_D_t.zero_grad() # Real loss pred_real = D_t(features_target.detach()) loss_D_real = criterion_GAN(pred_real, real_label) # Fake loss fake_t = fake_T_buffer.push_and_pop(fake_t) pred_fake = D_t(fake_t.detach()) loss_D_fake = criterion_GAN(pred_fake, fake_label) # Total loss loss_D_t = loss_D_real + loss_D_fake loss_D_t.backward() optimizer_D_t.step() now = datetime.datetime.now() d = str(now.month) + '-' + str(now.day) + ' ' + str(now.hour) + ':' + str( now.minute) + ":" + str(now.second) torch.save( best_model, osp.join( config["output_path"], "{}_to_{}_best_model_acc-{}_{}.pth.tar".format( args.source, args.target, best_acc, d))) return best_acc