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(loss.Entropy(nn.Softmax(dim=1)(all_output))).cpu().data.item()
if flag:
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
matrix = matrix[np.unique(all_label).astype(int),:]
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 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(loss.Entropy(
nn.Softmax(dim=1)(all_output))).cpu().data.item()
if flag:
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
pdb.set_trace()
acc = matrix.diagonal() / matrix.sum(axis=1)
return np.mean(acc), acc
else:
return accuracy, mean_ent
def cal_acc(loader, netF, netB, netC):
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)
all_output = nn.Softmax(dim=1)(all_output)
_, 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(loss.Entropy(all_output)).cpu().data.item()
return accuracy * 100, mean_ent
def cal_acc(loader, net, 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 = net(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)
all_output = nn.Softmax(dim=1)(all_output)
ent = torch.sum(-all_output * torch.log(all_output + args.epsilon),
dim=1) / np.log(all_output.size(1))
accuracy = torch.sum(
torch.squeeze(predict).float() == all_label).item() / float(
all_label.size()[0])
mean_ent = torch.mean(loss.Entropy(
nn.Softmax(dim=1)(all_output))).cpu().data.item()
return accuracy, mean_ent
def get_ent(oh_final_f):
oh_final_f = torch.from_numpy(oh_final_f)
all_output = nn.Softmax(dim=1)(oh_final_f)
out_ent = loss.Entropy(all_output)
mean_ent = torch.mean(out_ent)
out_ent_arr = out_ent.cpu().numpy()
return mean_ent, out_ent_arr
def train(args, config, model, ad_net, random_layer, train_loader,
train_loader1, optimizer, optimizer_ad, epoch):
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
total_loss = 0
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_source, output_source = model(data_source)
feature_target, output_target = model(data_target)
feature = torch.cat((feature_source, feature_target), 0)
output = torch.cat((output_source, output_target), 0)
labels_target_fake = torch.max(nn.Softmax(dim=1)(output_target), 1)[1]
labels = torch.cat((label_source, labels_target_fake))
loss = nn.CrossEntropyLoss()(output.narrow(0, 0, data_source.size(0)),
label_source)
softmax_output = nn.Softmax(dim=1)(output)
if epoch > 0:
entropy = loss_func.Entropy(softmax_output)
loss += loss_func.CDAN([feature, softmax_output], ad_net, entropy,
network.calc_coeff(num_iter * (epoch - 0) +
batch_idx), random_layer)
mdd_loss = args.mdd_weight * loss_func.mdd_digit(
feature, labels, args.left_weight, args.right_weight, args.weight)
loss = loss + mdd_loss
total_loss += loss.data
loss.backward()
optimizer.step()
if epoch > 0:
optimizer_ad.step()
if (batch_idx + epoch * num_iter) % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * args.batch_size, num_iter * args.batch_size,
100. * batch_idx / num_iter, loss.item()))
log_str = "total_loss:{}\n".format(total_loss)
config["out_file"].write(log_str)
config["out_file"].flush()
print(log_str)
def cal_acc(loader, netF=None, netB=None, netC=None, per_class_flag=False, visda_flag=False):
"""Calculate model accuracy on validation set or testing set
:param loader: dataloader
:param netF: feature extractor network
:param netB: bottleneck network
:param netC: classifier network
:param per_class_flag: if True: calculatge per-class average accuracy
:param visda_flag: if True: return acc of each class, else: no need to return acc of each class
:return: overall acc, per-class average acc, str: acc of each class, mean entropy
"""
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)
all_output = nn.Softmax(dim=1)(all_output)
_, predict = torch.max(all_output, 1)
accuracy = torch.sum(torch.squeeze(predict).float() == all_label).item() / float(all_label.size()[0])
accuracy *= 100 # overall accuracy
mean_ent = torch.mean(loss.Entropy(all_output)).cpu().data.item() # average entropy of classification results
if per_class_flag:
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
per_cls_acc_vec = matrix.diagonal() / matrix.sum(axis=1) * 100
per_cls_avg_acc = per_cls_acc_vec.mean() # Per-class avg acc
per_cls_acc_list = [str(np.round(i, 2)) for i in per_cls_acc_vec]
acc_each_cls = ' '.join(per_cls_acc_list) # str: acc of each class
if visda_flag:
# For VisDA, return acc of each cls to be printed
# overall acc, acc of each cls: str, per-class avg acc
return accuracy, acc_each_cls, per_cls_avg_acc
elif per_class_flag:
# For Office-Home and DomainNet, no need to return acc of each class
# overall acc, per-class avg acc, average entropy
return accuracy, per_cls_avg_acc, mean_ent
else:
# overall acc, mean-ent
return accuracy, mean_ent
def train_target(args):
dset_loaders = data_load(args)
param_group = []
model_resnet = network.Res50().cuda()
for k, v in model_resnet.named_parameters():
if k.__contains__('fc'):
v.requires_grad = False
else:
param_group += [{'params': v, 'lr': args.lr}]
optimizer = optim.SGD(param_group, momentum=0.9, weight_decay=5e-4, nesterov=True)
for epoch in tqdm(range(args.max_epoch), leave=False):
model_resnet.eval()
mem_label = obtain_label(dset_loaders['test'], model_resnet, args)
mem_label = torch.from_numpy(mem_label).cuda()
model_resnet.train()
iter_test = iter(dset_loaders['target'])
for _, (inputs_test, _, tar_idx) in tqdm(enumerate(iter_test), leave=False):
if inputs_test.size(0) == 1:
continue
inputs_test = inputs_test.cuda()
pred = mem_label[tar_idx]
features_test, outputs_test = model_resnet(inputs_test)
classifier_loss = loss.CrossEntropyLabelSmooth(num_classes=args.class_num, epsilon=0)(outputs_test, pred)
classifier_loss *= args.cls_par
if args.ent:
softmax_out = nn.Softmax(dim=1)(outputs_test)
entropy_loss = torch.mean(loss.Entropy(softmax_out))
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gentropy_loss = torch.sum(-msoftmax * torch.log(msoftmax + args.epsilon))
entropy_loss -= gentropy_loss
classifier_loss += entropy_loss * args.ent_par
optimizer.zero_grad()
classifier_loss.backward()
optimizer.step()
model_resnet.eval()
acc, ment = cal_acc(dset_loaders['test'], model_resnet)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(args.dset, epoch+1, args.max_epoch, acc*100)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str+'\n')
# torch.save(model_resnet.state_dict(), osp.join(args.output_dir, 'target.pt'))
return model_resnet
def cal_acc(loader, netF, netB, netC, flag=False, threshold=0.1):
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(loss.Entropy(
nn.Softmax(dim=1)(all_output))).cpu().data.item()
if flag:
all_output = nn.Softmax(dim=1)(all_output)
ent = torch.sum(-all_output * torch.log(all_output + args.epsilon),
dim=1) / np.log(args.class_num)
# predict[ent>threshold] = args.class_num
# from sklearn.mixture import GaussianMixture as GMM
# gmm = GMM(n_components=2, random_state=0).fit(ent.reshape(-1,1))
# labels = gmm.predict(ent.reshape(-1,1))
from sklearn.cluster import KMeans
kmeans = KMeans(2, random_state=0).fit(ent.reshape(-1, 1))
labels = kmeans.predict(ent.reshape(-1, 1))
idx = np.where(labels == 1)[0]
iidx = 0
if ent[idx].mean() > ent.mean():
iidx = 1
predict[np.where(labels == iidx)[0]] = args.class_num
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
matrix = matrix[np.unique(all_label).astype(int), :]
acc = matrix.diagonal() / matrix.sum(axis=1) * 100
unknown_acc = acc[-1:].item()
# accuracy = torch.sum(torch.squeeze(predict).float() == all_label).item() / float(all_label.size()[0]) * 100
# print(np.mean(acc[:-1]), np.mean(acc), unknown_acc)
return np.mean(acc[:-1]), np.mean(acc), unknown_acc
else:
return accuracy, mean_ent
def train(dataloader_src, dataloader_tgt, discriminator, classifier,
train_epochs, writer):
discriminator.train()
classifier.train()
loss_clf = nn.CrossEntropyLoss()
# 复习一下：momentum就是上次更新的方向和这次的梯度反向一样，那么这次就加快速度；
# weight_decay就是 L2 regularization
optimizer = optim.SGD(itertools.chain(classifier.parameters(),
discriminator.parameters()),
lr=1e-3,
momentum=0.9,
weight_decay=0.0009)
loss_clf_ = transfer_loss = 0
for epoch in range(train_epochs):
for (imgs_src,
labels_src), (imgs_tgt,
labels_tgt) in zip(dataloader_src, dataloader_tgt):
imgs_src = Variable(imgs_src.type(FloatTensor)).reshape(
imgs_src.shape[0], -1)
labels_src = Variable(labels_src.type(LongTensor))
imgs_tgt = Variable(imgs_tgt.type(FloatTensor)).reshape(
imgs_tgt.shape[0], -1)
labels_tgt = Variable(labels_tgt.type(FloatTensor))
# train source domain
fea_src, pred_src = classifier(imgs_src)
fea_tgt, pred_tgt = classifier(imgs_tgt)
fea = torch.cat((fea_src, fea_tgt), 0)
pred = torch.cat((pred_src, pred_tgt), 0)
# 计算概率
softmax_out = nn.Softmax(dim=1)(pred)
# 计算熵和discriminator loss
entropy = loss.Entropy(softmax_out)
transfer_loss = loss.CDAN([fea, softmax_out], discriminator,
entropy, networks.calc_coeff(epoch))
# classifier loss
loss_clf_ = loss_clf(pred_src, labels_src)
with OptimizerManager([optimizer]):
total_loss = transfer_loss + loss_clf_
total_loss.backward()
if epoch % 5 == 0:
acc_src, acc_tgt = evaluate(classifier, dataloader_src,
dataloader_tgt)
writer.add_scalar('Train/loss_c_src', loss_clf_, epoch)
writer.add_scalar('Train/transfer_loss', transfer_loss, epoch)
writer.add_scalar('Evaluate/Acc_src', acc_src, epoch)
writer.add_scalar('Evaluate/Acc_tgt', acc_tgt, epoch)
def train(args, model, ad_net, random_layer, train_loader, train_loader1,
optimizer, optimizer_ad, epoch, start_epoch, method):
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_source, output_source = model(data_source)
feature_target, output_target = model(data_target)
feature = torch.cat((feature_source, feature_target), 0)
output = torch.cat((output_source, output_target), 0)
labels_target_fake = torch.max(nn.Softmax(dim=1)(output_target), 1)[1]
labels = torch.cat((label_source, labels_target_fake))
loss = nn.CrossEntropyLoss()(output.narrow(0, 0, data_source.size(0)),
label_source)
softmax_output = nn.Softmax(dim=1)(output)
if epoch > start_epoch:
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), random_layer)
loss = loss + args.mdd_weight * loss_func.mdd_digit(
feature, labels
) + args.entropic_weight * loss_func.EntropicConfusion(feature)
loss.backward()
optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (batch_idx + epoch * num_iter) % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.4f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * args.batch_size, num_iter * args.batch_size,
100. * batch_idx / num_iter, loss.item()))
def train(args, model, ad_net, random_layer, train_loader, train_loader1,
optimizer, optimizer_ad, epoch, start_epoch, method):
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()
print('data_source:', data_source.shape, data_target.shape)
optimizer.zero_grad()
optimizer_ad.zero_grad()
feature, output = model(torch.cat((data_source, data_target), 0))
loss = nn.CrossEntropyLoss()(output.narrow(0, 0, data_source.size(0)),
label_source)
softmax_output = nn.Softmax(dim=1)(output)
if epoch > start_epoch:
if method == 'CDAN-E':
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), random_layer)
elif method == 'CDAN':
loss += loss_func.CDAN([feature, softmax_output], ad_net, None,
None, random_layer)
elif method == 'DANN':
loss += loss_func.DANN(feature, ad_net)
else:
raise ValueError('Method cannot be recognized.')
loss.backward()
optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (batch_idx + epoch * num_iter) % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * args.batch_size, num_iter * args.batch_size,
100. * batch_idx / num_iter, loss.item()))
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)
all_output = nn.Softmax(dim=1)(all_output)
_, 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(loss.Entropy(all_output)).cpu().data.item()
if flag:
all_output = nn.Softmax(dim=1)(all_output)
ent = torch.sum(-all_output * torch.log(all_output + args.epsilon),
dim=1) / np.log(args.class_num)
ent = ent.float().cpu()
initc = np.array([[0], [1]])
kmeans = KMeans(n_clusters=2, random_state=0, init=initc,
n_init=1).fit(ent.reshape(-1, 1))
threshold = (kmeans.cluster_centers_).mean()
predict[ent > threshold] = args.class_num
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
matrix = matrix[np.unique(all_label).astype(int), :]
acc = matrix.diagonal() / matrix.sum(axis=1)
unknown_acc = acc[-1:].item()
return np.mean(acc), np.mean(acc[:-1])
else:
return accuracy, mean_ent
def cal_acc(loader, netF, netB, netC):
k = 0
start_test = True
with torch.no_grad():
iter_test = iter(loader)
for i in range(len(loader)):
data = iter_test.next()
input_images = []
inputs = data[0]
inputs_clone = inputs.clone()
for j in range(inputs_clone.size(0)):
x = transforms.Normalize((-1, ), (2, ))(inputs_clone[j])
input_images.append(transforms.ToPILImage()(x))
labels = data[1]
outputs = netC(netB(netF(inputs)))
#
_, predict = torch.max(outputs.float().cpu(), 1)
for j in range(inputs.size(0)):
folder = args.output_dir + '/inspect/label-{}'.format(
labels[j])
if not osp.exists(folder):
os.makedirs(folder)
subfolder = folder + '/pred-{}'.format(predict[j])
if not osp.exists(subfolder):
os.makedirs(subfolder)
input_images[j].save(subfolder + '/{}.jpg'.format(k))
k += 1
#
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(loss.Entropy(
nn.Softmax(dim=1)(all_output))).cpu().data.item()
return accuracy * 100, mean_ent
def cal_acc_multi(loader, netF_list, netB_list, netC_list, netG_list, args):
start_test = True
with torch.no_grad():
iter_test = iter(loader)
for _ in range(len(loader)):
data = iter_test.next()
inputs = data[0]
labels = data[1]
inputs = inputs.cuda()
outputs_all = torch.zeros(len(args.src), inputs.shape[0], args.class_num)
weights_all = torch.ones(inputs.shape[0], len(args.src))
outputs_all_w = torch.zeros(inputs.shape[0], args.class_num)
for i in range(len(args.src)):
features = netB_list[i](netF_list[i](inputs))
outputs = netC_list[i](features)
weights = netG_list[i](features)
outputs_all[i] = outputs
weights_all[:, i] = weights.squeeze()
z = torch.sum(weights_all, dim=1)
z = z + 1e-16
weights_all = torch.transpose(torch.transpose(weights_all,0,1)/z,0,1)
print(weights_all.mean(dim=0))
outputs_all = torch.transpose(outputs_all, 0, 1)
for i in range(inputs.shape[0]):
outputs_all_w[i] = torch.matmul(torch.transpose(outputs_all[i],0,1), weights_all[i])
if start_test:
all_output = outputs_all_w.float().cpu()
all_label = labels.float()
start_test = False
else:
all_output = torch.cat((all_output, outputs_all_w.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(loss.Entropy(nn.Softmax(dim=1)(all_output))).cpu().data.item()
return accuracy*100, mean_ent
def cal_acc(loader, netF, netH, netB, netC, args, 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()
labels = labels.cuda()
if args.layer in ['add_margin', 'arc_margin', 'sphere']:
labels_forward = labels
else:
labels_forward = None
outputs = netC(netB(netF(inputs)), labels_forward)
if start_test:
all_output = outputs.float().cpu()
all_label = labels.float().cpu()
start_test = False
else:
all_output = torch.cat((all_output, outputs.float().cpu()), 0)
all_label = torch.cat((all_label, labels.float().cpu()), 0)
all_output = nn.Softmax(dim=1)(all_output)
_, 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(loss.Entropy(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 maxent_step(inputs, netF, netH, netB, netC, optim, epsilon=1e-8):
netF.train()
netH.train()
netB.train()
optim.zero_grad()
c3 = netC(netB(netF(inputs)), None)
softmax_out = nn.Softmax(dim=1)(c3)
msoftmax = softmax_out.mean(dim=0)
gentropy = -torch.sum(-msoftmax * torch.log(msoftmax + epsilon))
with torch.no_grad():
entropy = torch.mean(loss.Entropy(softmax_out))
gentropy.backward()
optim.step()
return entropy.item(), gentropy.item()
def cal_acc(loader, netF, netB, netC, visda_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])
accuracy *= 100
mean_ent = torch.mean(loss.Entropy(nn.Softmax(dim=1)(all_output))).cpu().data.item()
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
per_cls_acc_vec = matrix.diagonal() / matrix.sum(axis=1) * 100
per_cls_avg_acc = per_cls_acc_vec.mean() # Per-class avg acc
per_cls_acc_list = [str(np.round(i, 2)) for i in per_cls_acc_vec]
acc_each_cls = ' '.join(per_cls_acc_list)
if visda_flag:
# For VisDA, return acc of each cls to be printed
# overall acc, acc of each cls: str, per-class avg acc
return accuracy, acc_each_cls, per_cls_avg_acc
else:
# For other datasets, need not return acc of each cls
# overall acc, acc of each cls: str, mean-ent
return accuracy, per_cls_avg_acc, mean_ent
def train(args):
ent_loss_record = []
gent_loss_record = []
sent_loss_record = []
total_loss_record = []
dset_loaders = digit_load(args)
## set base network
if args.dset == 'u':
netF = network.LeNetBase() #.cuda()
elif args.dset == 'm':
netF = network.LeNetBase() #.cuda()
elif args.dset == 's':
netF = network.DTNBase() #.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()
param_group = []
learning_rate = args.lr
for k, v in netF.named_parameters():
param_group += [{'params': v, 'lr': learning_rate}]
for k, v in netB.named_parameters():
param_group += [{'params': v, 'lr': learning_rate}]
for k, v in netC.named_parameters():
param_group += [{'params': v, 'lr': learning_rate}]
optimizer = optim.SGD(param_group)
optimizer = op_copy(optimizer)
acc_init = 0
max_iter = args.max_epoch * len(dset_loaders["train"])
interval_iter = max_iter // 10
iter_num = 0
netF.train()
netB.train()
netC.train()
while iter_num < max_iter:
try:
inputs_source, strong_inputs, target = iter_source.next()
except:
iter_source = iter(dset_loaders["train"])
inputs_source, strong_inputs, target = iter_source.next()
if inputs_source.size(0) == 1:
continue
iter_num += 1
lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
inputs_source = inputs_source #.cuda()
outputs_source = netC(netB(netF(inputs_source)))
total_loss = torch.tensor(0.0) #.cuda()
softmax_out = nn.Softmax(dim=1)(outputs_source)
if args.ent:
ent_loss = torch.mean(loss.Entropy(softmax_out))
total_loss += ent_loss
ent_loss_record.append(ent_loss.detach().cpu())
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gent_loss = -torch.sum(-msoftmax * torch.log(msoftmax + 1e-5))
gent_loss_record.append(gent_loss.detach().cpu())
total_loss += gent_loss
if args.sent:
sent_loss = compute_aug_loss(strong_inputs, target, netC, netB,
netF)
total_loss += sent_loss
sent_loss_record.append(sent_loss.detach().cpu())
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
total_loss_record.append(total_loss.detach().cpu())
if iter_num % interval_iter == 0 or iter_num == max_iter:
print(iter_num, interval_iter, max_iter)
# netF.eval()
# netB.eval()
# netC.eval()
# acc_s_tr, _ = cal_acc(dset_loaders['train'], netF, netB, netC)
# acc_s_te, _ = cal_acc(dset_loaders['test'], netF, netB, netC)
# log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%/ {:.2f}%'.format(args.dset, iter_num, max_iter, acc_s_tr, acc_s_te)
# args.out_file.write(log_str + '\n')
# args.out_file.flush()
# print(log_str+'\n')
# if acc_s_te >= acc_init:
# acc_init = acc_s_te
# best_netF = netF.state_dict()
# best_netB = netB.state_dict()
# best_netC = netC.state_dict()
# netF.train()
# netB.train()
# netC.train()
best_netF = netF.state_dict()
best_netB = netB.state_dict()
best_netC = netC.state_dict()
torch.save(best_netF, osp.join(args.output_dir, "F.pt"))
torch.save(best_netB, osp.join(args.output_dir, "B.pt"))
torch.save(best_netC, osp.join(args.output_dir, "C.pt"))
fig, (ax1, ax2, ax3, ax4) = plt.subplots(nrows=4,
sharex=True,
figsize=(16, 8))
ax1.plot(list(range(len(ent_loss_record))), ent_loss_record, 'r')
ax2.plot(list(range(len(gent_loss_record))), gent_loss_record, 'g')
ax3.plot(list(range(len(sent_loss_record))), sent_loss_record, 'b')
ax4.plot(list(range(len(total_loss_record))), total_loss_record, 'm')
plt.tight_layout()
plt.savefig(args.output_dir + '/loss.png')
return netF, netB, netC
def train(config):
# set pre-process
prep_config = config["prep"]
prep_dict = {}
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"] = datasets.ImageFolder(data_config['source']['list_path'], transform=prep_dict["source"])
dset_loaders['source'] = getdataloader(dsets['source'], batchsize=train_bs, num_workers=4, drop_last=True, weightsampler=True)
dsets["target"] = datasets.ImageFolder(data_config['target']['list_path'], 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"] = [datasets.ImageFolder(data_config['test']['list_path'],
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"] = datasets.ImageFolder(data_config['test']['list_path'],
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()
# set test_ad_net
test_ad_net = network.AdversarialNetwork(base_network.output_num(), 1024, test_ad_net=True)
test_ad_net = test_ad_net.cuda()
# add additional network for some methods
if config['method'] == 'DANN':
random_layer = None
ad_net = network.AdversarialNetwork(base_network.output_num(), 1024)
elif config['method'] == 'MADA':
random_layer = None
ad_net = network.AdversarialNetworkClassGroup(base_network.output_num(), 1024, class_num)
elif config['method'] == 'proposed':
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)
ad_net_group = network.AdversarialNetworkGroup(config['loss']['random_dim'], 256, class_num, config['center_threshold'])
else:
random_layer = None
ad_net = network.AdversarialNetwork(base_network.output_num(), 1024)
ad_net_group = network.AdversarialNetworkGroup(base_network.output_num(), 1024, class_num, config['center_threshold'])
elif config['method'] == 'base':
pass
else:
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"] and config['method'] != 'base' and config['method'] != 'DANN' and config['method'] != 'MADA':
random_layer.cuda()
if config['method'] != 'base':
ad_net = ad_net.cuda()
if config['method'] == 'proposed':
ad_net_group = ad_net_group.cuda()
# set parameters
if config['method'] == 'proposed':
parameter_list = base_network.get_parameters() + test_ad_net.get_parameters() + ad_net.get_parameters() + ad_net_group.get_parameters()
elif config['method'] == 'base':
parameter_list = base_network.get_parameters() + test_ad_net.get_parameters()
elif config['method'] == 'MADA':
parameter_list = base_network.get_parameters() + test_ad_net.get_parameters() + ad_net.get_parameters()
else:
parameter_list = base_network.get_parameters() + test_ad_net.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"]]
# parallel
gpus = config['gpu'].split(',')
if len(gpus) > 1:
base_network = nn.DataParallel(base_network)
test_ad_net = nn.DataParallel(test_ad_net)
if config['method'] == 'DANN':
ad_net = nn.DataParallel(ad_net)
elif config['method'] == 'proposed':
if config['loss']['random']:
random_layer = nn.DataParallel(random_layer)
ad_net = nn.DataParallel(ad_net)
#将ad_net_group设置成并行将会引发error，原因可能是由于ad_net_group的输出不是tensor类型，parallel还不能支持。
#ad_net_group = nn.DataParallel(ad_net_group)
else:
ad_net = nn.DataParallel(ad_net)
#ad_net_group = nn.DataParallel(ad_net_group)
elif config['method'] == 'base':
pass
else:
# CDAN+E
if config["loss"]["random"]:
random_layer = nn.DataParallel(random_layer)
ad_net = nn.DataParallel(ad_net)
# CDAN
else:
ad_net = nn.DataParallel(ad_net)
## 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) # eval() == train(False) is True
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)))
loss_params = config["loss"]
# train one iter
base_network.train(True)
if config['method'] != 'base':
ad_net.train(True)
if config['method'] == 'proposed':
ad_net_group.train(True)
# lr_scheduler
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)
if config['tsne']:
# feature visualization by using T-SNE
if i == int(0.98*config['num_iterations']):
features_source_total = features_source.cpu().detach().numpy()
features_target_total = features_target.cpu().detach().numpy()
elif i > int(0.98*config['num_iterations']) and i < int(0.98*config['num_iterations'])+10:
features_source_total = np.concatenate((features_source_total, features_source.cpu().detach().numpy()))
features_target_total = np.concatenate((features_target_total, features_target.cpu().detach().numpy()))
elif i == int(0.98*config['num_iterations'])+10:
for index in range(config['tsne_num']):
features_embeded = TSNE(perplexity=10,n_iter=5000).fit_transform(np.concatenate((features_source_total, features_target_total)))
fig = plt.figure()
plt.scatter(features_embeded[:len(features_embeded)//2, 0], features_embeded[:len(features_embeded)//2, 1], c='r', s=1)
plt.scatter(features_embeded[len(features_embeded)//2:, 0], features_embeded[len(features_embeded)//2:, 1], c='b', s=1)
plt.savefig(osp.join(config["output_path"], config['method']+'-'+str(index)+'.png'))
plt.close()
else:
pass
assert features_source.size(0) == features_target.size(0), 'The batchsize must be same'
assert outputs_source.size(0) == outputs_target.size(0), 'The batchsize must be same'
# source first, target second
features = torch.cat((features_source, features_target), dim=0)
outputs = torch.cat((outputs_source, outputs_target), dim=0)
# output the A_distance
if i % config["test_interval"] == config["test_interval"] - 1:
A_distance = cal_A_distance(test_ad_net, features)
config['A_distance_file'].write(str(A_distance)+'\n')
config['A_distance_file'].flush()
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)
elif config['method'] == 'MADA':
transfer_loss = loss.MADA(features, softmax_out, ad_net)
elif config['method'] == 'proposed':
entropy = loss.Entropy(softmax_out)
transfer_loss = loss.proposed([features, outputs], labels_source, ad_net, ad_net_group, entropy,
network.calc_coeff(i), i, random_layer, config['loss']['trade_off23'])
elif config['method'] == 'base':
pass
else:
raise ValueError('Method cannot be recognized.')
test_domain_loss = loss.DANN(features.clone().detach(), test_ad_net)
classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source)
if config['method'] == 'base':
total_loss = classifier_loss + test_domain_loss
else:
total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss + test_domain_loss
total_loss.backward()
optimizer.step()
# torch.save(best_model, osp.join(config["output_path"], "best_model.pth.tar"))
return best_acc
def train_target(args):
dset_loaders = data_load(args)
if args.net[0:3] == 'res':
netF = network.ResBase(res_name=args.net).cuda()
elif args.net[0:3] == 'vgg':
netF = network.VGGBase(vgg_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))
netC.eval()
for k, v in netC.named_parameters():
v.requires_grad = False
param_group = []
for k, v in netF.named_parameters():
if args.lr_decay1 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay1}]
else:
v.requires_grad = False
for k, v in netB.named_parameters():
if args.lr_decay2 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay2}]
else:
v.requires_grad = False
optimizer = optim.SGD(param_group)
optimizer = op_copy(optimizer)
max_iter = args.max_epoch * len(dset_loaders["target"])
interval_iter = max_iter // args.interval
iter_num = 0
iter_sw = int(max_iter / 2.0)
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
if iter_num % interval_iter == 0 and args.cls_par > 0:
netF.eval()
netB.eval()
mem_label_soft, mtx_infor_nh, feas_FC = obtain_label(
dset_loaders['test'], netF, netB, netC, args, iter_num,
iter_sw)
mem_label_soft = torch.from_numpy(mem_label_soft).cuda()
feas_all = feas_FC[0]
ops_all = feas_FC[1]
netF.train()
netB.train()
inputs_test = inputs_test.cuda()
iter_num += 1
lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
features_F_self = netF(inputs_test)
features_F_nh = get_mtx_sam_wgt_nh(feas_all, mtx_infor_nh, tar_idx)
features_F_nh = features_F_nh.cuda()
features_F_mix = 0.8 * features_F_self + 0.2 * features_F_nh
outputs_test_mix = netC(netB(features_F_mix))
ops_test_self = netC(netB(features_F_self))
outputs_test_nh = netC(netB(features_F_nh))
if args.cls_par > 0:
log_probs = nn.LogSoftmax(dim=1)(outputs_test_mix)
targets = mem_label_soft[tar_idx]
loss_soft = (-targets * log_probs).sum(dim=1)
classifier_loss = loss_soft.mean()
classifier_loss *= args.cls_par
if iter_num < interval_iter and args.dset == "VISDA-C":
classifier_loss *= 0
else:
classifier_loss = torch.tensor(0.0).cuda()
if args.ent:
softmax_out = nn.Softmax(dim=1)(
outputs_test_mix) # outputs_test_mix
entropy_loss = torch.mean(loss.Entropy(softmax_out))
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gentropy_loss = torch.sum(-msoftmax *
torch.log(msoftmax + args.epsilon))
entropy_loss -= gentropy_loss
im_loss = entropy_loss * args.ent_par
classifier_loss += im_loss
optimizer.zero_grad()
classifier_loss.backward()
optimizer.step()
if iter_num % interval_iter == 0 or iter_num == max_iter:
netF.eval()
netB.eval()
if args.dset == 'VISDA-C':
acc_s_te, acc_list = cal_acc(dset_loaders['test'], netF, netB,
netC, True)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(
args.name, iter_num, max_iter, acc_s_te) + '\n' + acc_list
else:
acc_s_te, _ = cal_acc(dset_loaders['test'], netF, netB, netC,
False)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(
args.name, iter_num, max_iter, acc_s_te)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str + '\n')
netF.train()
netB.train()
if args.issave:
torch.save(
netF.state_dict(),
osp.join(args.output_dir, "target_F_" + args.savename + ".pt"))
torch.save(
netB.state_dict(),
osp.join(args.output_dir, "target_B_" + args.savename + ".pt"))
torch.save(
netC.state_dict(),
osp.join(args.output_dir, "target_C_" + args.savename + ".pt"))
return netF, netB, netC
def train_target(args):
dset_loaders = data_load(args)
## set base network
if args.net[0:3] == 'res':
netF = network.ResBase(res_name=args.net).cuda()
elif args.net[0:3] == 'vgg':
netF = network.VGGBase(vgg_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()
if not args.ssl == 0:
netR = network.feat_classifier(type='linear', class_num=4, bottleneck_dim=2*args.bottleneck).cuda()
netR_dict, acc_rot = train_target_rot(args)
netR.load_state_dict(netR_dict)
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))
netC.eval()
for k, v in netC.named_parameters():
v.requires_grad = False
param_group = []
for k, v in netF.named_parameters():
if args.lr_decay1 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay1}]
else:
v.requires_grad = False
for k, v in netB.named_parameters():
if args.lr_decay2 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay2}]
else:
v.requires_grad = False
if not args.ssl == 0:
for k, v in netR.named_parameters():
param_group += [{'params': v, 'lr': args.lr * args.lr_decay2}]
netR.train()
optimizer = optim.SGD(param_group)
optimizer = op_copy(optimizer)
max_iter = args.max_epoch * len(dset_loaders["target"])
interval_iter = max_iter // args.interval
iter_num = 0
while iter_num < max_iter:
optimizer.zero_grad()
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
if iter_num % interval_iter == 0 and args.cls_par > 0:
netF.eval()
netB.eval()
mem_label = obtain_label(dset_loaders['target_te'], netF, netB, netC, args)
mem_label = torch.from_numpy(mem_label).cuda()
netF.train()
netB.train()
inputs_test = inputs_test.cuda()
iter_num += 1
lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
if args.cls_par > 0:
pred = mem_label[tar_idx]
features_test = netB(netF(inputs_test))
outputs_test = netC(features_test)
if args.cls_par > 0:
classifier_loss = nn.CrossEntropyLoss()(outputs_test, pred)
classifier_loss *= args.cls_par
if iter_num < interval_iter and args.dset == "VISDA-C":
classifier_loss *= 0
else:
classifier_loss = torch.tensor(0.0).cuda()
if args.ent:
softmax_out = nn.Softmax(dim=1)(outputs_test)
entropy_loss = torch.mean(loss.Entropy(softmax_out))
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gentropy_loss = torch.sum(-msoftmax * torch.log(msoftmax + args.epsilon))
entropy_loss -= gentropy_loss
im_loss = entropy_loss * args.ent_par
classifier_loss += im_loss
classifier_loss.backward()
if not args.ssl == 0:
r_labels_target = np.random.randint(0, 4, len(inputs_test))
r_inputs_target = rotation.rotate_batch_with_labels(inputs_test, r_labels_target)
r_labels_target = torch.from_numpy(r_labels_target).cuda()
r_inputs_target = r_inputs_target.cuda()
f_outputs = netB(netF(inputs_test))
f_outputs = f_outputs.detach()
f_r_outputs = netB(netF(r_inputs_target))
r_outputs_target = netR(torch.cat((f_outputs, f_r_outputs), 1))
rotation_loss = args.ssl * nn.CrossEntropyLoss()(r_outputs_target, r_labels_target)
rotation_loss.backward()
optimizer.step()
if iter_num % interval_iter == 0 or iter_num == max_iter:
netF.eval()
netB.eval()
if args.dset=='VISDA-C':
acc_s_te, acc_list = cal_acc(dset_loaders['test'], netF, netB, netC, True)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(args.name, iter_num, max_iter, acc_s_te) + '\n' + acc_list
else:
acc_s_te, _ = cal_acc(dset_loaders['test'], netF, netB, netC, False)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(args.name, iter_num, max_iter, acc_s_te)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str+'\n')
netF.train()
netB.train()
if args.issave:
torch.save(netF.state_dict(), osp.join(args.output_dir, "target_F_" + args.savename + ".pt"))
torch.save(netB.state_dict(), osp.join(args.output_dir, "target_B_" + args.savename + ".pt"))
torch.save(netC.state_dict(), osp.join(args.output_dir, "target_C_" + args.savename + ".pt"))
return netF, netB, netC
def train_target(args):
dset_loaders = data_load(args)
## set base network
if args.net[0:3] == 'res':
netF = network.ResBase(res_name=args.net).cuda()
elif args.net[0:3] == 'vgg':
netF = network.VGGBase(vgg_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()
args.modelpath = args.output_dir_src + '/source_F.pt'
netF.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir_src + '/source_B.pt'
netB.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir_src + '/source_C.pt'
netC.load_state_dict(torch.load(args.modelpath))
netC.eval()
for k, v in netC.named_parameters():
v.requires_grad = False
param_group = []
for k, v in netF.named_parameters():
if args.lr_decay1 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay1}]
else:
v.requires_grad = False
for k, v in netB.named_parameters():
if args.lr_decay2 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay2}]
else:
v.requires_grad = False
optimizer = optim.SGD(param_group)
optimizer = op_copy(optimizer)
tt = 0
iter_num = 0
max_iter = args.max_epoch * len(dset_loaders["target"])
interval_iter = max_iter // args.interval
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
if iter_num % interval_iter == 0:
netF.eval()
netB.eval()
mem_label, ENT_THRESHOLD = obtain_label(dset_loaders['test'], netF,
netB, netC, args)
mem_label = torch.from_numpy(mem_label).cuda()
netF.train()
netB.train()
inputs_test = inputs_test.cuda()
iter_num += 1
lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
pred = mem_label[tar_idx]
features_test = netB(netF(inputs_test))
outputs_test = netC(features_test)
softmax_out = nn.Softmax(dim=1)(outputs_test)
outputs_test_known = outputs_test[pred < args.class_num, :]
pred = pred[pred < args.class_num]
if len(pred) == 0:
print(tt)
del features_test
del outputs_test
tt += 1
continue
if args.cls_par > 0:
classifier_loss = nn.CrossEntropyLoss()(outputs_test_known, pred)
classifier_loss *= args.cls_par
else:
classifier_loss = torch.tensor(0.0).cuda()
if args.ent:
softmax_out_known = nn.Softmax(dim=1)(outputs_test_known)
entropy_loss = torch.mean(loss.Entropy(softmax_out_known))
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gentropy_loss = torch.sum(-msoftmax *
torch.log(msoftmax + args.epsilon))
entropy_loss -= gentropy_loss
classifier_loss += entropy_loss * args.ent_par
optimizer.zero_grad()
classifier_loss.backward()
optimizer.step()
if iter_num % interval_iter == 0 or iter_num == max_iter:
netF.eval()
netB.eval()
acc_os1, acc_os2, acc_unknown = cal_acc(dset_loaders['test'], netF,
netB, netC, True,
ENT_THRESHOLD)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}% / {:.2f}% / {:.2f}%'.format(
args.name, iter_num, max_iter, acc_os2, acc_os1, acc_unknown)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str + '\n')
netF.train()
netB.train()
if args.issave:
torch.save(
netF.state_dict(),
osp.join(args.output_dir, "target_F_" + args.savename + ".pt"))
torch.save(
netB.state_dict(),
osp.join(args.output_dir, "target_B_" + args.savename + ".pt"))
torch.save(
netC.state_dict(),
osp.join(args.output_dir, "target_C_" + args.savename + ".pt"))
return netF, netB, netC
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):
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)
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(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'])
tensor_writer = SummaryWriter(config["tensorboard_path"])
## 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()
parameter_list = base_network.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:
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
for i in range(config["num_iterations"]):
if i % config["test_interval"] == config["test_interval"] - 1 or i==0:
base_network.train(False)
temp_acc, output, prediction, label, feature = 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)))
loss_params = config["loss"]
## train one iter
base_network.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)
outputs_target_temp = outputs_target / config['temperature']
target_softmax_out_temp = nn.Softmax(dim=1)(outputs_target_temp)
target_entropy_weight = loss.Entropy(target_softmax_out_temp).detach()
target_entropy_weight = 1 + torch.exp(-target_entropy_weight)
target_entropy_weight = train_bs * target_entropy_weight / torch.sum(target_entropy_weight)
cov_matrix_t = target_softmax_out_temp.mul(target_entropy_weight.view(-1,1)).transpose(1,0).mm(target_softmax_out_temp)
cov_matrix_t = cov_matrix_t / torch.sum(cov_matrix_t, dim=1)
mcc_loss = (torch.sum(cov_matrix_t) - torch.trace(cov_matrix_t)) / class_num
classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source)
total_loss = classifier_loss + mcc_loss
total_loss.backward()
optimizer.step()
tensor_writer.add_scalar('total_loss', total_loss, i)
tensor_writer.add_scalar('classifier_loss', classifier_loss, i)
tensor_writer.add_scalar('cov_matrix_penalty', mcc_loss, i)
torch.save(best_model, osp.join(config["output_path"], "best_model.pth.tar"))
return best_acc
def train_target(args, zz=''):
dset_loaders = data_load(args)
## set base network
if args.net[0:3] == 'res':
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()
args.modelpath = args.output_dir_src + '/source_F_' + str(zz) + '.pt'
netF.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir_src + '/source_B_' + str(zz) + '.pt'
netB.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir_src + '/source_C_' + str(zz) + '.pt'
netC.load_state_dict(torch.load(args.modelpath))
netC.eval()
for k, v in netC.named_parameters():
v.requires_grad = False
param_group = []
for k, v in netF.named_parameters():
if args.lr_decay1 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay1}]
else:
v.requires_grad = False
for k, v in netB.named_parameters():
if args.lr_decay2 > 0:
param_group += [{'params': v, 'lr': args.lr * args.lr_decay2}]
else:
v.requires_grad = False
optimizer = optim.SGD(param_group,
momentum=0.9,
weight_decay=5e-4,
nesterov=True)
for epoch in tqdm(range(args.max_epoch), leave=False):
netF.eval()
netB.eval()
mem_label = obtain_label(dset_loaders['test'], netF, netB, netC, args)
mem_label = torch.from_numpy(mem_label).cuda()
netF.train()
netB.train()
iter_test = iter(dset_loaders['target'])
for _, (inputs_test, _, tar_idx) in tqdm(enumerate(iter_test),
leave=False):
if inputs_test.size(0) == 1:
continue
inputs_test = inputs_test.cuda()
pred = mem_label[tar_idx]
features_test = netB(netF(inputs_test))
outputs_test = netC(features_test)
classifier_loss = loss.CrossEntropyLabelSmooth(
num_classes=args.class_num, epsilon=0)(outputs_test, pred)
classifier_loss *= args.cls_par
if args.ent:
softmax_out = nn.Softmax(dim=1)(outputs_test)
entropy_loss = torch.mean(loss.Entropy(softmax_out))
if args.gent:
msoftmax = softmax_out.mean(dim=0)
gentropy_loss = torch.sum(
-msoftmax * torch.log(msoftmax + args.epsilon))
entropy_loss -= gentropy_loss
classifier_loss += entropy_loss * args.ent_par
optimizer.zero_grad()
classifier_loss.backward()
optimizer.step()
netF.eval()
netB.eval()
acc, _ = cal_acc(dset_loaders['test'], netF, netB, netC)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(
args.name, epoch + 1, args.max_epoch, acc * 100)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str + '\n')
if args.issave:
torch.save(
netF.state_dict(),
osp.join(args.output_dir, 'target_F_' + args.savename + '.pt'))
torch.save(
netB.state_dict(),
osp.join(args.output_dir, 'target_B_' + args.savename + '.pt'))
torch.save(
netC.state_dict(),
osp.join(args.output_dir, 'target_C_' + args.savename + '.pt'))
return netF, netB, netC
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()))))
print("it_train: {:05d} / {:05d} start".format(i, config["num_iterations"]))
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))
# 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.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()
print("it_train: {:05d} / {:05d} over".format(i, config["num_iterations"]))
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
def train(config):
####################################################
# Tensorboard setting
####################################################
#tensor_writer = SummaryWriter(config["tensorboard_path"])
####################################################
# 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, 이렇게 해야 마지막까지 source, target에서 동일한 수로 배치 생성가능
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,
'type' : args.type}
}
"""
base_network = net_config["name"](**net_config["params"])
#network.py에 정의된 ResNetFc() 클래스 호출
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) # 왜 class 수 만큼 곱하지?
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, # 이거 0.01이여야 하지 않나?
#'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"]] # return optimizer
####################################################
# 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
batch_size = config["data"]["source"]["batch_size"]
for i in range(
config["num_iterations"]): # num_iterations수의 batch가 학습에 사용됨
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()
inputs = torch.cat((inputs_source, inputs_target), dim=0)
features, outputs, tau, cur_mean_source, cur_mean_target, output_mean_source, output_mean_target = base_network(
inputs)
softmax_out = nn.Softmax(dim=1)(outputs)
outputs_source = outputs[:batch_size]
outputs_target = outputs[batch_size:]
if config['method'] == 'CDAN+E' or config['method'] == 'CDAN_TransNorm':
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()
#tensor_writer.add_scalar('total_loss', total_loss.i )
#tensor_writer.add_scalar('classifier_loss', classifier_loss, i)
#tensor_writer.add_scalar('transfer_loss', transfer_loss, i)
####################################################
# 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):
## 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()
## 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
# 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)))
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)
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)
if i % 10 == 0:
print('iter: ', i, 'classifier_loss: ', classifier_loss.data,
'transfer_loss: ', transfer_loss.data)
total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss
total_loss.backward()
optimizer.step()
torch.save(best_model, osp.join(config["output_path"],
"best_model.pth.tar"))
return best_acc
def train_target(args):
dset_loaders = digit_load(args)
## set base network
if args.dset == 'u2m':
netF = network.LeNetBase().cuda()
elif args.dset == 'm2u':
netF = network.LeNetBase().cuda()
elif args.dset == 's2m':
netF = network.DTNBase().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()
args.modelpath = args.output_dir + '/source_F.pt'
netF.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir + '/source_B.pt'
netB.load_state_dict(torch.load(args.modelpath))
args.modelpath = args.output_dir + '/source_C.pt'
netC.load_state_dict(torch.load(args.modelpath))
netC.eval()
for k, v in netC.named_parameters():
v.requires_grad = False
param_group = []
for k, v in netF.named_parameters():
param_group += [{'params': v, 'lr': args.lr}]
for k, v in netB.named_parameters():
param_group += [{'params': v, 'lr': args.lr}]
optimizer = optim.SGD(param_group)
optimizer = op_copy(optimizer)
max_iter = args.max_epoch * len(dset_loaders["target"])
interval_iter = len(dset_loaders["target"])
# interval_iter = max_iter // args.interval
iter_num = 0
while iter_num < max_iter:
optimizer.zero_grad()
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
if iter_num % interval_iter == 0 and args.cls_par > 0:
netF.eval()
netB.eval()
mem_label = obtain_label(dset_loaders['target_te'], netF, netB,
netC, args)
mem_label = torch.from_numpy(mem_label).cuda()
netF.train()
netB.train()
iter_num += 1
lr_scheduler(optimizer, iter_num=iter_num, max_iter=max_iter)
inputs_test = inputs_test.cuda()
features_test = netB(netF(inputs_test))
outputs_test = netC(features_test)
if args.cls_par > 0:
pred = mem_label[tar_idx]
classifier_loss = args.cls_par * nn.CrossEntropyLoss()(
outputs_test, pred)
else:
classifier_loss = torch.tensor(0.0).cuda()
if args.ent:
softmax_out = nn.Softmax(dim=1)(outputs_test)
entropy_loss = torch.mean(loss.Entropy(softmax_out))
# if args.gent:
# msoftmax = softmax_out.mean(dim=0)
# entropy_loss -= torch.sum(-msoftmax * torch.log(msoftmax + 1e-5))
im_loss = entropy_loss * args.ent_par
classifier_loss += im_loss
optimizer.zero_grad()
classifier_loss.backward()
optimizer.step()
if iter_num % interval_iter == 0 or iter_num == max_iter:
netF.eval()
netB.eval()
acc, _ = cal_acc(dset_loaders['test'], netF, netB, netC)
log_str = 'Task: {}, Iter:{}/{}; Accuracy = {:.2f}%'.format(
args.dset, iter_num, max_iter, acc)
args.out_file.write(log_str + '\n')
args.out_file.flush()
print(log_str + '\n')
netF.train()
netB.train()
if args.issave:
torch.save(
netF.state_dict(),
osp.join(args.output_dir, "target_F_" + args.savename + ".pt"))
torch.save(
netB.state_dict(),
osp.join(args.output_dir, "target_B_" + args.savename + ".pt"))
torch.save(
netC.state_dict(),
osp.join(args.output_dir, "target_C_" + args.savename + ".pt"))
return netF, netB, netC