def train_fada(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
critic = Critic2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
critic = critic.cuda()
criterion = torch.nn.CrossEntropyLoss()
loss_class = torch.nn.CrossEntropyLoss()
loss_domain = torch.nn.CrossEntropyLoss()
res_dir = os.path.join(config['res_dir'],
'snr{}-lr{}'.format(config['snr'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_dann')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(critic)
logging.debug(config)
optimizer = optim.Adam([{
'params': extractor.parameters()
}, {
'params': classifier.parameters()
}, {
'params': critic.parameters()
}],
lr=config['lr'])
# TODO
discriminator = main_models.DCD(input_features=128)
optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=0.001)
# source samples预训练
#--------------pretrain g and h for step 1---------------------------------
for epoch in range(config['n_epoches_1']):
for data, labels in config['source_train_loader']:
data = data.to(device)
labels = labels.to(device)
optimizer.zero_grad()
y_pred = classifier(extractor(data))
loss = loss_class(y_pred, labels)
loss.backward()
optimizer.step()
acc = 0
for data, labels in config['target_test_loader']:
data = data.to(device)
labels = labels.to(device)
y_test_pred = classifier(extractor(data))
acc += (torch.max(y_test_pred,
1)[1] == labels).float().mean().item()
accuracy = round(acc / float(len(config['target_test_loader'])), 3)
print("step1----Epoch %d/%d accuracy: %.3f " %
(epoch + 1, config['n_epoches_1'], accuracy))
#-----------------train DCD for step 2--------------------------------
# X_s,Y_s=dataloader.sample_data()
# X_t,Y_t=dataloader.create_target_samples(config['n_target_samples'])
for epoch in range(config['n_epoches_2']):
# for data,labels in config['source_train_loader']:
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, label_target = iter_target.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target, label_target = data_target.cuda(
), label_target.cuda()
groups, aa = dataloader.sample_groups(data_source,
label_source,
data_target,
label_target,
seed=epoch)
# groups, aa = dataloader.sample_groups(X_s,Y_s,X_t,Y_t,seed=epoch)
n_iters = 4 * len(groups[1])
index_list = torch.randperm(n_iters)
mini_batch_size = 40 #use mini_batch train can be more stable
loss_mean = []
X1 = []
X2 = []
ground_truths = []
for index in range(n_iters):
ground_truth = index_list[index] // len(groups[1])
x1, x2 = groups[ground_truth][index_list[index] -
len(groups[1]) * ground_truth]
X1.append(x1)
X2.append(x2)
ground_truths.append(ground_truth)
#select data for a mini-batch to train
if (index + 1) % mini_batch_size == 0:
X1 = torch.stack(X1)
X2 = torch.stack(X2)
ground_truths = torch.LongTensor(ground_truths)
X1 = X1.to(device)
X2 = X2.to(device)
ground_truths = ground_truths.to(device)
optimizer_D.zero_grad()
X_cat = torch.cat([extractor(X1), extractor(X2)], 1)
y_pred = discriminator(X_cat.detach())
loss = loss_domain(y_pred, ground_truths)
loss.backward()
optimizer_D.step()
loss_mean.append(loss.item())
X1 = []
X2 = []
ground_truths = []
print("step2----Epoch %d/%d loss:%.3f" %
(epoch + 1, config['n_epoches_2'], np.mean(loss_mean)))
def train_pada(config):
if config['network'] == 'inceptionv1':
extractor_s = InceptionV1(num_classes=32)
extractor_t = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor_s = InceptionV1s(num_classes=32)
extractor_t = InceptionV1s(num_classes=32)
else:
extractor_s = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
extractor_t = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier_s = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
classifier_t = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor_s = extractor_s.cuda()
classifier_s = classifier_s.cuda()
extractor_t = extractor_t.cuda()
classifier_t = classifier_t.cuda()
cdan_random = config['random_layer']
res_dir = os.path.join(
config['res_dir'],
'normal{}-{}-cons{}-lr{}'.format(config['normal'], config['network'],
config['pada_cons_w'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_pada')
print(config)
set_log_config(res_dir)
logging.debug('train_pada')
# logging.debug(extractor)
# logging.debug(classifier)
logging.debug(config)
if config['models'] == 'PADA':
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
elif cdan_random:
random_layer = RandomLayer([config['n_flattens'], config['n_class']],
config['n_hiddens'])
ad_net = AdversarialNetwork(config['n_hiddens'], config['n_hiddens'])
random_layer.cuda()
else:
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'] * config['n_class'],
config['n_hiddens'])
ad_net = ad_net.cuda()
optimizer_s = torch.optim.Adam([{
'params': extractor_s.parameters(),
'lr': config['lr']
}, {
'params': classifier_s.parameters(),
'lr': config['lr']
}])
optimizer_t = torch.optim.Adam([{
'params': extractor_t.parameters(),
'lr': config['lr']
}, {
'params': classifier_t.parameters(),
'lr': config['lr']
}])
optimizer_ad = torch.optim.Adam(ad_net.parameters(), lr=config['lr'])
def train_stage1(extractor_s, classifier_s, config, epoch):
extractor_s.train()
classifier_s.train()
# STAGE 1:
# 在labeled source上训练extractor_s和classifier_s
# 训练完成后freeze这两个model
iter_source = iter(config['source_train_loader'])
len_source_loader = len(config['source_train_loader'])
for step in range(1, len_source_loader + 1):
data_source, label_source = iter_source.next()
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
optimizer_s.zero_grad()
h_s = extractor_s(data_source)
h_s = h_s.view(h_s.size(0), -1)
source_preds = classifier_s(h_s)
cls_loss = nn.CrossEntropyLoss()(source_preds, label_source)
cls_loss.backward()
optimizer_s.step()
def train(extractor_s, classifier_s, extractor_t, classifier_t, ad_net,
config, epoch):
start_epoch = 0
# extractor_s.train()
# classifier_s.train()
# ad_net.train()
# # STAGE 1:
# # 在labeled source上训练extractor_s和classifier_s
# # 训练完成后freeze这两个model
# iter_source = iter(config['source_train_loader'])
# len_source_loader = len(config['source_train_loader'])
# for step in range(1, len_source_loader + 1):
# data_source, label_source = iter_source.next()
# if torch.cuda.is_available():
# data_source, label_source = data_source.cuda(), label_source.cuda()
# optimizer_s.zero_grad()
# h_s = extractor_s(data_source)
# h_s = h_s.view(h_s.size(0), -1)
# source_preds = classifier_s(h_s)
# cls_loss = nn.CrossEntropyLoss()(source_preds, label_source)
# cls_loss.backward()
# optimizer_s.step()
# for param in extractor_s.parameters():
# param.requires_grad = False
# for param in classifier_s.parameters():
# param.requires_grad = False
# STAGE 2:
# 使用新的extractor和classifier进行DANN训练
# 不同的地方是，每个target 同时使用extractor_s和extractor_t
extractor_t.train()
classifier_t.train()
ad_net.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for step in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, label_target = iter_target.next()
if step % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target, label_target = data_target.cuda(
), label_target.cuda()
optimizer_t.zero_grad()
optimizer_ad.zero_grad()
h_s = extractor_t(data_source)
h_s = h_s.view(h_s.size(0), -1)
h_t = extractor_t(data_target)
h_t = h_t.view(h_t.size(0), -1)
source_preds = classifier_t(h_s)
cls_loss = nn.CrossEntropyLoss()(source_preds, label_source)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
target_preds = classifier_t(h_t)
softmax_output_t = nn.Softmax(dim=1)(target_preds)
if config['target_labeling'] == 1:
cls_loss += nn.CrossEntropyLoss()(target_preds, label_target)
feature = torch.cat((h_s, h_t), 0)
softmax_output = torch.cat((softmax_output_s, softmax_output_t), 0)
if epoch > start_epoch:
gamma = 2 / (1 + math.exp(-10 *
(epoch) / config['n_epochs'])) - 1
if config['models'] == 'CDAN-E':
entropy = loss_func.Entropy(softmax_output)
d_loss = loss_func.CDAN(
[feature, softmax_output], ad_net, gamma, entropy,
loss_func.calc_coeff(num_iter * (epoch - start_epoch) +
step), random_layer)
elif config['models'] == 'CDAN':
d_loss = loss_func.CDAN([feature, softmax_output], ad_net,
gamma, None, None, random_layer)
elif config['models'] == 'PADA':
d_loss = loss_func.DANN(feature, ad_net, gamma)
else:
raise ValueError('Method cannot be recognized.')
else:
d_loss = 0
# constraints loss
h_s_prev = extractor_s(data_source)
cons_loss = nn.L1Loss()(h_s, h_s_prev)
loss = cls_loss + d_loss + config['pada_cons_w'] * cons_loss
loss.backward()
optimizer_t.step()
if epoch > start_epoch:
optimizer_ad.step()
if (step) % 20 == 0:
print(
'Train Epoch {} closs {:.6f}, dloss {:.6f}, cons_loss {:.6f}, Loss {:.6f}'
.format(epoch, cls_loss.item(), d_loss.item(),
cons_loss.item(), loss.item()))
for epoch in range(1, config['n_epochs'] + 1):
train_stage1(extractor_s, classifier_s, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
print('test on source_test_loader')
test(extractor_s, classifier_s, config['source_test_loader'],
epoch)
# print('test on target_test_loader')
# accuracy = test(extractor_s, classifier_s, config['target_test_loader'], epoch)
extractor_t.load_state_dict(extractor_s.state_dict())
classifier_t.load_state_dict(classifier_s.state_dict())
for param in extractor_s.parameters():
param.requires_grad = False
for param in classifier_s.parameters():
param.requires_grad = False
for epoch in range(1, config['n_epochs'] + 1):
train(extractor_s, classifier_s, extractor_t, classifier_t, ad_net,
config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
accuracy = test(extractor_t, classifier_t,
config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
title = config['models']
draw_confusion_matrix(extractor_t, classifier_t,
config['target_test_loader'], res_dir, epoch,
title)
draw_tsne(extractor_t,
classifier_t,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
def train_cnn(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
res_dir = os.path.join(config['res_dir'], 'normal{}-{}-lr{}'.format(config['normal'],
config['network'],
config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
criterion = torch.nn.CrossEntropyLoss()
# optimizer = optim.Adam(
# list(extractor.parameters()) + list(classifier.parameters()),
# lr = config['lr'])
opts = optim.SGD(
list(extractor.parameters()) + list(classifier.parameters()),
lr = config['lr'], nesterov=True, momentum=0.9)
def train(extractor, classifier, config, epoch):
extractor.train()
classifier.train()
optimizer = inv_lr_scheduler(opts, epoch, gamma=0.01, power=0.75, lr=config['lr'], weight_decay=0.0005)
for step, (features, labels) in enumerate(config['source_train_loader']):
if torch.cuda.is_available():
features, labels = features.cuda(), labels.cuda()
optimizer.zero_grad()
# if config['aux_classifier'] == 1:
# x1, x2, x3 = extractor(features)
# preds = classifier(x1, x2, x3)
preds = classifier(extractor(features))
# print('preds {}, labels {}'.format(preds.shape, labels.shape))
# print(preds[0])
# preds_l = F.softmax(preds, dim=1)
# print('preds_l {}'.format(preds_l.shape))
# print(preds_l[0])
# print('------')
loss = criterion(preds, labels)
loss.backward()
optimizer.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
print('test on source_test_loader')
test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
def train_adda(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], bn=config['bn'])
classifier = Classifier(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], n_class=config['n_class'])
criterion = torch.nn.CrossEntropyLoss()
loss_class = torch.nn.CrossEntropyLoss()
loss_domain = torch.nn.CrossEntropyLoss()
lr = config['lr']
res_dir = os.path.join(config['res_dir'], 'lr{}'.format(config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_adda')
logging.debug(config)
def pretrain(model, config, pretrain_epochs):
model.class_classifier.train()
model.feature.train()
optimizer = optim.Adam(model.parameters(), lr=lr)
for epoch in range(pretrain_epochs):
for step, (features, labels) in enumerate(config['source_train_loader']):
if torch.cuda.is_available():
features, labels = features.cuda(), labels.cuda()
optimizer.zero_grad()
preds = model.class_classify(features)
loss = criterion(preds, labels)
loss.backward()
optimizer.step()
def train(model, config, epoch):
model.class_classifier.train()
model.feature.train()
# LEARNING_RATE = lr / math.pow((1 + 10 * (epoch - 1) / config['n_epochs']), 0.75)
# print('epoch {}, learning rate{: .4f}'.format(epoch, LEARNING_RATE) )
# optimizer = torch.optim.SGD([
# {'params': model.feature.parameters()},
# {'params': model.class_classifier.parameters(), 'lr': LEARNING_RATE},
# ], lr= LEARNING_RATE / 10, momentum=momentum, weight_decay=l2_decay)
optimizer = optim.Adam(model.parameters(), lr=lr)
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for i in range(1, num_iter):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(), label_source.cuda()
data_target = data_target.cuda()
optimizer.zero_grad()
class_output_s, domain_output, embeddings_s = model.dann(input_data=data_source, alpha=gamma)
# print('domain_output {}'.format(domain_output.size()))
err_s_label = loss_class(class_output_s, label_source)
domain_label = torch.zeros(data_source.size(0)).long().cuda()
err_s_domain = loss_domain(domain_output, domain_label)
# Training model using target data
domain_label = torch.ones(data_target.size(0)).long().cuda()
class_output_t, domain_output, embeddings_t = model.dann(input_data=data_target, alpha=gamma)
err_t_domain = loss_domain(domain_output, domain_label)
err = err_s_label + err_s_domain + err_t_domain
if i % 50 == 0:
print('err_s_label {}, err_s_domain {}, gamma {}, err_t_domain {}, total err {}'.format(err_s_label.item(), err_s_domain.item(), gamma, err_t_domain.item(), err.item()))
err.backward()
optimizer.step()
pretrain(model, config, pretrain_epochs=20)
for epoch in range(1, config['n_epochs'] + 1):
train(model, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
print('test on source_test_loader')
test(extractor, classifier, config['source_test_loader'], epoch)
# print('test on target_train_loader')
# test(model, config['target_train_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(extractor, classifier, config['target_test_loader'], res_dir, epoch, config['models'])
draw_tsne(extractor, classifier, config['source_test_loader'], config['target_test_loader'], res_dir, epoch, config['models'], separate=True)
def train_cnn(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
res_dir = os.path.join(
config['res_dir'],
'slim{}-snr{}-snrp{}-lr{}'.format(config['slim'], config['snr'],
config['snrp'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.Adam(list(extractor.parameters()) +
list(classifier.parameters()),
lr=config['lr'])
def train(extractor, classifier, config, epoch):
extractor.train()
classifier.train()
for step, (features,
labels) in enumerate(config['source_train_loader']):
if torch.cuda.is_available():
features, labels = features.cuda(), labels.cuda()
optimizer.zero_grad()
# if config['aux_classifier'] == 1:
# x1, x2, x3 = extractor(features)
# preds = classifier(x1, x2, x3)
preds, _ = classifier(extractor(features))
# print('preds {}, labels {}'.format(preds.shape, labels.shape))
# print(preds[0])
# preds_l = F.softmax(preds, dim=1)
# print('preds_l {}'.format(preds_l.shape))
# print(preds_l[0])
# print('------')
loss = criterion(preds, labels)
loss.backward()
optimizer.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
print('test on source_test_loader')
test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
config['models'])
draw_tsne(extractor,
classifier,
config['source_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=True)
draw_tsne(extractor,
classifier,
config['source_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=False)
def train_dann(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
critic = Critic2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
critic = critic.cuda()
criterion = torch.nn.CrossEntropyLoss()
loss_class = torch.nn.CrossEntropyLoss()
loss_domain = torch.nn.CrossEntropyLoss()
res_dir = os.path.join(
config['res_dir'],
'VIS-slim{}-targetLabel{}-mmd{}-bnm{}-vat{}-ent{}-ew{}-bn{}-bs{}-lr{}'.
format(config['slim'], config['target_labeling'], config['mmd'],
config['bnm'], config['vat'], config['ent'], config['bnm_ew'],
config['bn'], config['batch_size'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_dann')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(critic)
logging.debug(config)
optimizer = optim.Adam([{
'params': extractor.parameters()
}, {
'params': classifier.parameters()
}, {
'params': critic.parameters()
}],
lr=config['lr'])
vat_loss = VAT(extractor, classifier, n_power=1, radius=3.5).cuda()
def dann(input_data, alpha):
feature = extractor(input_data)
feature = feature.view(feature.size(0), -1)
reverse_feature = ReverseLayerF.apply(feature, alpha)
class_output, _ = classifier(feature)
domain_output = critic(reverse_feature)
return class_output, domain_output, feature
def train(extractor, classifier, critic, config, epoch):
extractor.train()
classifier.train()
critic.train()
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
mmd_loss = MMD_loss()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
if config['slim'] > 0:
iter_target_semi = iter(config['target_train_semi_loader'])
len_target_semi_loader = len(config['target_train_semi_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, label_target = iter_target.next()
if config['slim'] > 0:
data_target_semi, label_target_semi = iter_target_semi.next()
if i % len_target_semi_loader == 0:
iter_target_semi = iter(config['target_train_semi_loader'])
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target, label_target = data_target.cuda(
), label_target.cuda()
if config['slim'] > 0:
data_target_semi, label_target_semi = data_target_semi.cuda(
), label_target_semi.cuda()
optimizer.zero_grad()
class_output_s, domain_output, feature_s = dann(
input_data=data_source, alpha=gamma)
# print('domain_output {}'.format(domain_output.size()))
err_s_label = loss_class(class_output_s, label_source)
domain_label = torch.zeros(data_source.size(0)).long().cuda()
err_s_domain = loss_domain(domain_output, domain_label)
# Training model using target data
domain_label = torch.ones(data_target.size(0)).long().cuda()
class_output_t, domain_output, feature_t = dann(
input_data=data_target, alpha=gamma)
#class_output_t, domain_output, _ = dann(input_data=data_target, alpha=0.5)
err_t_domain = loss_domain(domain_output, domain_label)
err = err_s_label + err_s_domain + err_t_domain
# if config['target_labeling'] == 1:
# err_t_class_healthy = nn.CrossEntropyLoss()(class_output_t, label_target)
# err += err_t_class_healthy
# if i % 100 == 0:
# print('err_t_class_healthy {:.2f}'.format(err_t_class_healthy.item()))
if config['mmd'] == 1:
#err += gamma * mmd_linear(feature_s, feature_t)
err += config['bnm_ew'] * mmd_loss(feature_s, feature_t)
if config['bnm'] == 1 and epoch >= config['startiter']:
err_t_bnm = config['bnm_ew'] * get_loss_bnm(class_output_t)
err += err_t_bnm
if i == 1:
print('epoch {}, loss_t_bnm {:.2f}'.format(
epoch, err_t_bnm.item()))
if config['ent'] == 1 and epoch >= config['startiter']:
err_t_ent = config['bnm_ew'] * get_loss_entropy(class_output_t)
err += err_t_ent
if i == 1:
print('epoch {}, loss_t_ent {:.2f}'.format(
epoch, err_t_ent.item()))
if config['vat'] == 1 and epoch >= config['startiter']:
err_t_vat = config['bnm_ew'] * vat_loss(
data_target, class_output_t)
err += err_t_vat
if i == 1:
print('epoch {}, loss_t_vat {:.2f}'.format(
epoch, err_t_vat.item()))
if config['slim'] > 0:
feature_target_semi = extractor(data_target_semi)
feature_target_semi = feature_target_semi.view(
feature_target_semi.size(0), -1)
preds_target_semi, _ = classifier(feature_target_semi)
err_t_class_semi = loss_class(preds_target_semi,
label_target_semi)
err += err_t_class_semi
if i == 1:
print('epoch {}, err_t_class_semi {:.2f}'.format(
epoch, err_t_class_semi.item()))
if i == 1:
print(
'epoch {}, err_s_label {:.2f}, err_s_domain {:.2f}, err_t_domain {:.2f}, total err {:.2f}'
.format(epoch, err_s_label.item(), err_s_domain.item(),
err_t_domain.item(), err.item()))
err.backward()
optimizer.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, critic, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
title = 'DANN'
if config['bnm'] == 1 and config['vat'] == 1:
title = '(b) Proposed'
elif config['bnm'] == 1:
title = 'BNM'
elif config['vat'] == 1:
title = 'VADA'
elif config['mmd'] == 1:
title = 'DCTLN'
elif config['ent'] == 1:
title = 'EntMin'
# draw_confusion_matrix(extractor, classifier, config['target_test_loader'], res_dir, epoch, config['models'])
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=False)
def train_dctln(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
critic = Critic2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
critic = critic.cuda()
criterion = torch.nn.CrossEntropyLoss()
loss_class = torch.nn.CrossEntropyLoss()
loss_domain = torch.nn.CrossEntropyLoss()
res_dir = os.path.join(
config['res_dir'],
'slim{}-snr{}-lr{}'.format(config['slim'], config['snr'],
config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_dann')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(critic)
logging.debug(config)
optimizer = optim.Adam([{
'params': extractor.parameters()
}, {
'params': classifier.parameters()
}, {
'params': critic.parameters()
}],
lr=config['lr'])
def dann(input_data, alpha):
feature = extractor(input_data)
feature = feature.view(feature.size(0), -1)
reverse_feature = ReverseLayerF.apply(feature, alpha)
class_output = classifier(feature)
domain_output = critic(reverse_feature)
return class_output, domain_output, feature
def train(extractor, classifier, critic, config, epoch):
extractor.train()
classifier.train()
critic.train()
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
if config['slim'] > 0:
iter_target_semi = iter(config['target_train_semi_loader'])
len_target_semi_loader = len(config['target_train_semi_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
data_target_semi, label_target_semi = iter_target_semi.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if i % len_target_semi_loader == 0:
iter_target_semi = iter(config['target_train_semi_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
data_target_semi, label_target_semi = data_target_semi.cuda(
), label_target_semi.cuda()
optimizer.zero_grad()
source = extractor(data_source)
source = source.view(source.size(0), -1)
target = extractor(data_target)
target = target.view(target.size(0), -1)
# loss_mmd = mmd_linear(source, target)
class_output_s, domain_output, _ = dann(input_data=data_source,
alpha=gamma)
# print('domain_output {}'.format(domain_output.size()))
err_s_label = loss_class(class_output_s, label_source)
domain_label = torch.zeros(data_source.size(0)).long().cuda()
err_s_domain = loss_domain(domain_output, domain_label)
# Training model using target data
domain_label = torch.ones(data_target.size(0)).long().cuda()
class_output_t, domain_output, _ = dann(input_data=data_target,
alpha=gamma)
err_t_domain = loss_domain(domain_output, domain_label)
class_output_semi_t, _, _ = dann(input_data=data_target_semi,
alpha=gamma)
err_t_label = loss_class(class_output_semi_t, label_target_semi)
# err = 1.0*err_s_label + err_s_domain + err_t_domain + 0*loss_mmd + err_t_label
err = 1.0 * err_s_label + err_s_domain + err_t_domain + err_t_label
# if i % 200 == 0:
# # print('err_s_label {}, err_s_domain {}, gamma {}, err_t_domain {}, loss_mmd {}, total err {}'.format(err_s_label.item(), err_s_domain.item(), gamma, err_t_domain.item(), loss_mmd.item(), err.item()))
# print('err_s_label {:.2f}, err_t_label {:.2f}, err_s_domain {:.2f}, gamma {:.2f}, err_t_domain {:.2f}, total err {:.2f}'.format(err_s_label.item(), err_t_label.item(), err_s_domain.item(), gamma, err_t_domain.item(), err.item()))
err.backward()
optimizer.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, critic, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
config['models'])
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=True)
def train_cdan(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], bn=config['bn'])
classifier = Classifier(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], n_class=config['n_class'])
vat_loss = VAT(extractor, classifier, n_power=1, radius=3.5).cuda()
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
cdan_random = config['random_layer']
res_dir = os.path.join(config['res_dir'], 'slim{}-targetLabel{}-snr{}-snrp{}-lr{}'.format(config['slim'],
config['target_labeling'],
config['snr'],
config['snrp'],
config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_cdan')
#print(extractor)
#print(classifier)
print(config)
set_log_config(res_dir)
logging.debug('train_cdan')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
if config['models'] == 'DANN':
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
elif cdan_random:
random_layer = RandomLayer([config['n_flattens'], config['n_class']], config['n_hiddens'])
ad_net = AdversarialNetwork(config['n_hiddens'], config['n_hiddens'])
random_layer.cuda()
else:
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'] * config['n_class'], config['n_hiddens'])
ad_net = ad_net.cuda()
optimizer = torch.optim.Adam([
{'params': extractor.parameters(), 'lr': config['lr']},
{'params': classifier.parameters(), 'lr': config['lr']}
])
optimizer_ad = torch.optim.Adam(ad_net.parameters(), lr=config['lr'])
extractor_path = os.path.join(res_dir, "extractor.pth")
classifier_path = os.path.join(res_dir, "classifier.pth")
adnet_path = os.path.join(res_dir, "adnet.pth")
def train(extractor, classifier, ad_net, config, epoch):
start_epoch = 0
extractor.train()
classifier.train()
ad_net.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
if config['slim'] > 0:
iter_target_semi = iter(config['target_train_semi_loader'])
len_target_semi_loader = len(config['target_train_semi_loader'])
num_iter = len_source_loader
for step in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, label_target = iter_target.next()
if config['slim'] > 0:
data_target_semi, label_target_semi = iter_target_semi.next()
if step % len_target_semi_loader == 0:
iter_target_semi = iter(config['target_train_semi_loader'])
if step % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(), label_source.cuda()
data_target, label_target = data_target.cuda(), label_target.cuda()
if config['slim'] > 0:
data_target_semi, label_target_semi = data_target_semi.cuda(), label_target_semi.cuda()
optimizer.zero_grad()
optimizer_ad.zero_grad()
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
h_t = extractor(data_target)
h_t = h_t.view(h_t.size(0), -1)
source_preds = classifier(h_s)
cls_loss = nn.CrossEntropyLoss()(source_preds, label_source)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
target_preds = classifier(h_t)
softmax_output_t = nn.Softmax(dim=1)(target_preds)
if config['target_labeling'] == 1:
cls_loss += nn.CrossEntropyLoss()(target_preds, label_target)
feature = torch.cat((h_s, h_t), 0)
softmax_output = torch.cat((softmax_output_s, softmax_output_t), 0)
if epoch > start_epoch:
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
if config['models'] == 'CDAN-E':
entropy = loss_func.Entropy(softmax_output)
d_loss = loss_func.CDAN([feature, softmax_output], ad_net, gamma, entropy, loss_func.calc_coeff(num_iter*(epoch-start_epoch)+step), random_layer)
elif config['models'] == 'CDAN':
d_loss = loss_func.CDAN([feature, softmax_output], ad_net, gamma, None, None, random_layer)
elif config['models'] == 'DANN':
d_loss = loss_func.DANN(feature, ad_net, gamma)
else:
raise ValueError('Method cannot be recognized.')
else:
d_loss = 0
loss = cls_loss + d_loss
err_t_bnm = get_loss_bnm(target_preds)
err_s_vat = vat_loss(data_source, source_preds)
err_t_vat = vat_loss(data_target, target_preds)
loss += 1.0 * err_s_vat + 1.0 * err_t_vat + 1.0 * err_t_bnm
if config['slim'] > 0:
feature_target_semi = extractor(data_target_semi)
feature_target_semi = feature_target_semi.view(feature_target_semi.size(0), -1)
preds_target_semi = classifier(feature_target_semi)
loss += nn.CrossEntropyLoss()(preds_target_semi, label_target_semi)
loss.backward()
optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (step) % 100 == 0:
print('Train Epoch {} closs {:.6f}, dloss {:.6f}, Loss {:.6f}'.format(epoch, cls_loss.item(), d_loss.item(), loss.item()))
if config['testonly'] == 0:
best_accuracy = 0
best_model_index = -1
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, ad_net, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
accuracy = test(extractor, classifier, config['target_test_loader'], epoch)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_model_index = epoch
torch.save(extractor.state_dict(), extractor_path)
torch.save(classifier.state_dict(), classifier_path)
torch.save(ad_net.state_dict(), adnet_path)
print('epoch {} accuracy: {:.6f}, best accuracy {:.6f} on epoch {}'.format(epoch, accuracy, best_accuracy, best_model_index))
if epoch % config['VIS_INTERVAL'] == 0:
title = config['models']
draw_confusion_matrix(extractor, classifier, config['target_test_loader'], res_dir, epoch, title)
draw_tsne(extractor, classifier, config['source_train_loader'], config['target_test_loader'], res_dir, epoch, title, separate=True)
# draw_tsne(extractor, classifier, config['source_test_loader'], config['target_test_loader'], res_dir, epoch, title, separate=False)
else:
if os.path.exists(extractor_path) and os.path.exists(classifier_path) and os.path.exists(adnet_path):
extractor.load_state_dict(torch.load(extractor_path))
classifier.load_state_dict(torch.load(classifier_path))
ad_net.load_state_dict(torch.load(adnet_path))
print('Test only mode, model loaded')
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], -1)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], -1)
title = config['models']
draw_confusion_matrix(extractor, classifier, config['target_test_loader'], res_dir, -1, title)
# draw_tsne(extractor, classifier, config['source_test_loader'], config['target_test_loader'], res_dir, -1, title, separate=True)
draw_tsne(extractor, classifier, config['source_train_loader'], config['target_test_loader'], res_dir, -1, title, separate=True)
else:
print('no saved model found')
def train_deepcoral(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
res_dir = os.path.join(
config['res_dir'], 'normal{}-{}-dilation{}-lr{}-mmdgamma{}'.format(
config['normal'], config['network'], config['dilation'],
config['lr'], config['mmd_gamma']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
criterion = torch.nn.CrossEntropyLoss()
set_log_config(res_dir)
logging.debug('train_deepcoral')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
optimizer = optim.Adam(list(extractor.parameters()) +
list(classifier.parameters()),
lr=config['lr'])
def train(extractor, classifier, config, epoch):
extractor.train()
classifier.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
if config['slim'] > 0:
iter_target_semi = iter(config['target_train_semi_loader'])
len_target_semi_loader = len(config['target_train_semi_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if config['slim'] > 0:
data_target_semi, label_target_semi = iter_target_semi.next()
if i % len_target_semi_loader == 0:
iter_target_semi = iter(config['target_train_semi_loader'])
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
if config['slim'] > 0:
data_target_semi, label_target_semi = data_target_semi.cuda(
), label_target_semi.cuda()
optimizer.zero_grad()
source = extractor(data_source)
source = source.view(source.size(0), -1)
target = extractor(data_target)
target = target.view(target.size(0), -1)
preds = classifier(source)
loss_cls = criterion(preds, label_source)
loss_coral = CORAL(source, target)
# gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
# loss = loss_cls + gamma * loss_coral
loss = loss_cls + config['mmd_gamma'] * loss_coral
if config['slim'] > 0:
feature_target_semi = extractor(data_target_semi)
feature_target_semi = feature_target_semi.view(
feature_target_semi.size(0), -1)
preds_target_semi = classifier(feature_target_semi)
err_t_class_semi = criterion(preds_target_semi,
label_target_semi)
loss += err_t_class_semi
if i % 50 == 0:
print('loss_cls {}, loss_coral {}, gamma {}, total loss {}'.
format(loss_cls.item(), loss_coral.item(),
config['mmd_gamma'], loss.item()))
loss.backward()
optimizer.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
print('test on source_test_loader')
test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
config['models'])
draw_tsne(extractor,
classifier,
config['source_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=True)
draw_tsne(extractor,
classifier,
config['source_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=False)
def train_tcl_vat(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
#summary(extractor, (1, 5120))
res_dir = os.path.join(
config['res_dir'],
'slim{}-snr{}-snrp{}-Lythred{}-Ldthred{}-lambdad{}-lr{}'.format(
config['slim'], config['snr'], config['snrp'], config['Lythred'],
config['Ldthred'], config['lambdad'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_tcl')
#print(extractor)
#print(classifier)
print(config)
set_log_config(res_dir)
logging.debug('train_tcl')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
ad_net = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
ad_net = ad_net.cuda()
optimizer = torch.optim.Adam([{
'params': extractor.parameters(),
'lr': config['lr']
}, {
'params': classifier.parameters(),
'lr': config['lr']
}],
weight_decay=0.0001)
optimizer_ad = torch.optim.Adam(ad_net.parameters(),
lr=config['lr'],
weight_decay=0.0001)
print(ad_net)
extractor_path = os.path.join(res_dir, "extractor.pth")
classifier_path = os.path.join(res_dir, "classifier.pth")
adnet_path = os.path.join(res_dir, "adnet.pth")
def cal_Ly(source_y_softmax, source_d, label):
#
# source_y_softmax, category预测结果带softmax
# source_d，domain预测结果
# label: 实际category标签
#
agey = -math.log(config['Lythred'])
aged = -math.log(1.0 - config['Ldthred'])
age = agey + config['lambdad'] * aged
# print('agey {}, labmdad {}, aged {}, age {}'.format(agey, config['lambdad'], aged, age))
y_softmax = source_y_softmax
the_index = torch.LongTensor(np.array(range(
config['batch_size']))).cuda()
# 这是什么意思？对于每个样本，只取出实际label对应的softmax值
# 与softmax loss有什么区别？
y_label = y_softmax[the_index, label]
# print('y_softmax {}, the_index {}, y_label shape {}'.format(y_softmax.shape, the_index.shape, y_label.shape))
y_loss = -torch.log(y_label + 1e-8)
d_loss = -torch.log(1.0 - source_d)
d_loss = d_loss.view(config['batch_size'])
weight_loss = y_loss + config['lambdad'] * d_loss
# print('y_loss {}'.format(torch.mean(y_loss)))
# print('lambdad {}'.format(config['lambdad']))
# print('d_loss {}'.format(torch.mean(d_loss)))
# print('y_loss {}'.format(y_loss.item()))
# print('lambdad {}'.format(config['lambdad']))
# print('d_loss {}'.format(d_loss.item()))
weight_var = (weight_loss < age).float().detach()
Ly = torch.mean(y_loss * weight_var)
source_weight = weight_var.data.clone()
source_num = float((torch.sum(source_weight)))
return Ly, source_weight, source_num
def cal_Lt(target_y_softmax):
# 这是entropy loss吧？
Gt_var = target_y_softmax
Gt_en = -torch.sum((Gt_var * torch.log(Gt_var + 1e-8)), 1)
Lt = torch.mean(Gt_en)
return Lt
def train(extractor, classifier, ad_net, config, epoch):
start_epoch = 0
extractor.train()
classifier.train()
ad_net.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
if config['slim'] > 0:
iter_target_semi = iter(config['target_train_semi_loader'])
len_target_semi_loader = len(config['target_train_semi_loader'])
for step in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if config['slim'] > 0:
data_target_semi, label_target_semi = iter_target_semi.next()
if step % len_target_semi_loader == 0:
iter_target_semi = iter(config['target_train_semi_loader'])
if step % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
if config['slim'] > 0:
data_target_semi, label_target_semi = data_target_semi.cuda(
), label_target_semi.cuda()
source_domain_label = torch.FloatTensor(config['batch_size'], 1)
target_domain_label = torch.FloatTensor(config['batch_size'], 1)
source_domain_label.fill_(1)
target_domain_label.fill_(0)
domain_label = torch.cat(
[source_domain_label, target_domain_label], 0)
domain_label = domain_label.cuda()
inputs = torch.cat([data_source, data_target], 0)
features = extractor(inputs)
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
y_var = classifier(features)
features = features.view(features.size(0), -1)
d_var = ad_net(features, gamma)
y_softmax_var = nn.Softmax(dim=1)(y_var)
source_y, target_y = y_var.chunk(2, 0)
source_y_softmax, target_y_softmax = y_softmax_var.chunk(2, 0)
source_d, target_d = d_var.chunk(2, 0)
# h_s = extractor(data_source)
# h_s = h_s.view(h_s.size(0), -1)
# h_t = extractor(data_target)
# h_t = h_t.view(h_t.size(0), -1)
# source_preds = classifier(h_s)
# softmax_output_s = nn.Softmax(dim=1)(source_preds)
# target_preds = classifier(h_t)
# softmax_output_t = nn.Softmax(dim=1)(target_preds)
# source_d, d_loss_source = loss_func.DANN_logits(h_s, ad_net, gamma)
# target_d, d_loss_target = loss_func.DANN_logits(h_t, ad_net, gamma)
# source_d = ad_net(h_s, gamma)
# target_d = ad_net(h_t, gamma)
#calculate Ly
if epoch < config['startiter']:
#也就是cls_loss，不考虑权重
Ly = nn.CrossEntropyLoss()(source_y, label_source)
else:
Ly, source_weight, source_num = cal_Ly(source_y_softmax,
source_d, label_source)
# print('source_num {}'.format(source_num))
target_weight = torch.ones(source_weight.size()).cuda()
#calculate Lt
# 计算target category的熵
Lt = cal_Lt(target_y_softmax)
#calculate Ld
if epoch < config['startiter']:
Ld = nn.BCELoss()(d_var, domain_label)
else:
domain_weight = torch.cat([source_weight, target_weight], 0)
domain_weight = domain_weight.view(-1, 1)
# print('domain_weight {}'.format(domain_weight.shape))
# print('domain_weight {}'.format(domain_weight))
# print('d_var {}'.format(d_var))
domain_criterion = nn.BCELoss(weight=domain_weight).cuda()
# domain_criterion = nn.BCELoss().cuda()
# print('max {}'.format(torch.max(d_var)))
# print('min {}'.format(torch.min(d_var)))
# print(d_var)
Ld = domain_criterion(d_var, domain_label)
loss = Ly + config['traded'] * Ld + config['tradet'] * Lt
if config['slim'] > 0:
feature_target_semi = extractor(data_target_semi)
feature_target_semi = feature_target_semi.view(
feature_target_semi.size(0), -1)
preds_target_semi = classifier(feature_target_semi)
loss += nn.CrossEntropyLoss()(preds_target_semi,
label_target_semi)
optimizer.zero_grad()
optimizer_ad.zero_grad()
# net.zero_grad()
loss.backward()
optimizer.step()
optimizer_ad.step()
# if (step) % 20 == 0:
# print('Train Epoch {} closs {:.6f}, dloss {:.6f}, coral_loss {:.6f}, Loss {:.6f}'.format(epoch, cls_loss.item(), d_loss.item(), coral_loss.item(), loss.item()))
# print('Train Epoch {} closs {:.6f}, dloss {:.6f}, Loss {:.6f}'.format(epoch, cls_loss.item(), d_loss.item(), loss.item()))
best_accuracy = 0
best_model_index = -1
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, ad_net, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
accuracy = test(extractor, classifier,
config['target_test_loader'], epoch)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_model_index = epoch
torch.save(extractor.state_dict(), extractor_path)
torch.save(classifier.state_dict(), classifier_path)
torch.save(ad_net.state_dict(), adnet_path)
print(
'epoch {} accuracy: {:.6f}, best accuracy {:.6f} on epoch {}'.
format(epoch, accuracy, best_accuracy, best_model_index))
if epoch % config['VIS_INTERVAL'] == 0:
title = config['models']
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
title)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=False)
def train_cdan_vat(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
bn=config['bn'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
cdan_random = config['random_layer']
res_dir = os.path.join(
config['res_dir'],
'normal{}-{}-dilation{}-lr{}'.format(config['normal'],
config['network'],
config['dilation'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_cdan_vat')
#print(extractor)
#print(classifier)
print(config)
set_log_config(res_dir)
logging.debug('train_cdan_vat')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
if config['models'] == 'DANN_VAT':
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
elif cdan_random:
random_layer = RandomLayer([config['n_flattens'], config['n_class']],
config['n_hiddens'])
ad_net = AdversarialNetwork(config['n_hiddens'], config['n_hiddens'])
random_layer.cuda()
else:
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'] * config['n_class'],
config['n_hiddens'])
ad_net = ad_net.cuda()
optimizer = torch.optim.Adam([{
'params': extractor.parameters(),
'lr': config['lr']
}, {
'params': classifier.parameters(),
'lr': config['lr']
}])
optimizer_ad = torch.optim.Adam(ad_net.parameters(), lr=config['lr'])
vat_loss = VAT(extractor, classifier, n_power=1, radius=3.5).cuda()
def train(extractor, classifier, ad_net, config, epoch):
start_epoch = 0
extractor.train()
classifier.train()
ad_net.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for step in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, label_target = iter_target.next()
if step % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target, label_target = data_target.cuda(
), label_target.cuda()
with torch.no_grad():
if 'CDAN' in config['models']:
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
source_preds = classifier(h_s)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
op_out = torch.bmm(softmax_output_s.unsqueeze(2),
h_s.unsqueeze(1))
gamma = 2 / (1 +
math.exp(-10 *
(epoch) / config['n_epochs'])) - 1
ad_out = ad_net(op_out.view(
-1,
softmax_output_s.size(1) * h_s.size(1)),
gamma,
training=False)
dom_entropy = 1 - (torch.abs(0.5 - ad_out))**config['iw']
dom_weight = dom_entropy
elif 'DANN' in config['models']:
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
gamma = 2 / (1 +
math.exp(-10 *
(epoch) / config['n_epochs'])) - 1
ad_out = ad_net(h_s, gamma, training=False)
dom_entropy = 1 - (torch.abs(0.5 - ad_out))**config['iw']
dom_weight = dom_entropy
optimizer.zero_grad()
optimizer_ad.zero_grad()
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
h_t = extractor(data_target)
h_t = h_t.view(h_t.size(0), -1)
source_preds = classifier(h_s)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
if config['iw'] > 0:
cls_loss = nn.CrossEntropyLoss(reduction='none')(source_preds,
label_source)
cls_loss = torch.mean(dom_weight * cls_loss)
# print('dom_weight mean {}'.format(torch.mean(dom_weight)))
else:
cls_loss = nn.CrossEntropyLoss()(source_preds, label_source)
target_preds = classifier(h_t)
softmax_output_t = nn.Softmax(dim=1)(target_preds)
if config['target_labeling'] == 1:
cls_loss += nn.CrossEntropyLoss()(target_preds, label_target)
feature = torch.cat((h_s, h_t), 0)
softmax_output = torch.cat((softmax_output_s, softmax_output_t), 0)
if epoch > start_epoch:
gamma = 2 / (1 + math.exp(-10 *
(epoch) / config['n_epochs'])) - 1
if config['models'] == 'CDAN-E':
entropy = loss_func.Entropy(softmax_output)
d_loss = loss_func.CDAN(
[feature, softmax_output], ad_net, gamma, entropy,
loss_func.calc_coeff(num_iter * (epoch - start_epoch) +
step), random_layer)
elif config['models'] == 'CDAN_VAT':
d_loss = loss_func.CDAN([feature, softmax_output], ad_net,
gamma, None, None, random_layer)
elif config['models'] == 'DANN_VAT':
d_loss = loss_func.DANN(feature, ad_net, gamma)
else:
raise ValueError('Method cannot be recognized.')
else:
d_loss = 0
# target entropy loss
err_t_entropy = get_loss_entropy(softmax_output_t)
# virtual adversarial loss.
err_s_vat = vat_loss(data_source, source_preds)
err_t_vat = vat_loss(data_target, target_preds)
# loss = cls_loss + d_loss
loss = cls_loss + d_loss + err_t_entropy + err_s_vat + err_t_vat
loss.backward()
optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (step) % 20 == 0:
print('Train Epoch {} closs {:.6f}, dloss {:.6f}, Loss {:.6f}'.
format(epoch, cls_loss.item(), d_loss.item(),
loss.item()))
best_accuracy = 0
best_model_index = -1
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, ad_net, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
accuracy = test(extractor, classifier,
config['target_test_loader'], epoch)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_model_index = epoch
print(
'epoch {} accuracy: {:.6f}, best accuracy {:.6f} on epoch {}'.
format(epoch, accuracy, best_accuracy, best_model_index))
if epoch % config['VIS_INTERVAL'] == 0:
title = config['models']
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
title)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
draw_tsne(extractor,
classifier,
config['source_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=False)
def train_mcd_a(config):
if config['network'] == 'inceptionv1':
G = InceptionV1(num_classes=32, dilation=config['dilation'])
elif config['network'] == 'inceptionv1s':
G = InceptionV1s(num_classes=32, dilation=config['dilation'])
else:
G = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
C1 = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
C2 = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
G = G.cuda()
C1 = C1.cuda()
C2 = C2.cuda()
# opt_g = optim.Adam(G.parameters(), lr=config['lr'], weight_decay=0.0005)
# opt_c1 = optim.Adam(C1.parameters(), lr=config['lr'], weight_decay=0.0005)
# opt_c2 = optim.Adam(C2.parameters(), lr=config['lr'], weight_decay=0.0005)
opt_g = optim.Adam(G.parameters(), lr=config['lr'])
opt_c1 = optim.Adam(C1.parameters(), lr=config['lr'])
opt_c2 = optim.Adam(C2.parameters(), lr=config['lr'])
criterion = torch.nn.CrossEntropyLoss()
res_dir = os.path.join(
config['res_dir'],
'normal{}-{}-lr{}'.format(config['normal'], config['network'],
config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_mcd')
logging.debug(G)
logging.debug(C1)
logging.debug(C2)
logging.debug(config)
def train(G, C1, C2, config, epoch):
G.train()
C1.train()
C2.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
# step 1.
#
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
# 源分类误差
feat_s = G(data_source)
feat_t = G(data_target)
output_s1 = C1(feat_s)
output_s2 = C2(feat_s)
loss_s1 = criterion(output_s1, label_source)
loss_s2 = criterion(output_s2, label_source)
loss = loss_s1 + loss_s2
if i == 0:
print('Epoch: {} Stage1, Loss1: {:.6f}, Loss2: {:.6f}'.format(
epoch, loss_s1.item(), loss_s2.item()))
if config['mcd_vat']:
# vat误差
source_vat_c1 = vat(C1, feat_s, 0.5)
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
target_vat_c1 = vat(C1, feat_t, 0.5)
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
source_vat_c2 = vat(C2, feat_s, 0.5)
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
target_vat_c2 = vat(C2, feat_t, 0.5)
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
source_vat_loss1 = get_loss_vat(C1, feat_s, source_vat_c1)
source_vat_loss2 = get_loss_vat(C2, feat_s, source_vat_c2)
target_vat_loss1 = get_loss_vat(C1, feat_t, target_vat_c1)
target_vat_loss2 = get_loss_vat(C2, feat_t, target_vat_c2)
loss += source_vat_loss1 + source_vat_loss2 + target_vat_loss1 + target_vat_loss2
if i == 0:
print(
'Epoch: {} Stage1, source_vat_loss1: {:.6f}, source_vat_loss2: {:.6f}, target_vat_loss1: {:.6f}, target_vat_loss2: {:.6f}'
.format(epoch, source_vat_loss1.item(),
source_vat_loss2.item(),
target_vat_loss1.item(),
target_vat_loss2.item()))
if config['mcd_entropy'] == 1:
entropy_loss_c1 = get_loss_entropy(C1, feat_t)
entropy_loss_c2 = get_loss_entropy(C2, feat_t)
loss += entropy_loss_c1 + entropy_loss_c2
if i == 0:
print(
'Epoch: {} Stage1, entropy c1: {:.6f}, entropy c2: {:.6f}'
.format(epoch, entropy_loss_c1.item(),
entropy_loss_c2.item()))
loss.backward()
opt_g.step()
opt_c1.step()
opt_c2.step()
# step 2.
#
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
feat_s = G(data_source)
output_s1 = C1(feat_s)
output_s2 = C2(feat_s)
loss_dis_s = discrepancy(output_s1, output_s2)
feat_t = G(data_target)
output_t1 = C1(feat_t)
output_t2 = C2(feat_t)
loss_dis_t = discrepancy(output_t1, output_t2)
loss_s1 = criterion(output_s1, label_source)
loss_s2 = criterion(output_s2, label_source)
# 源分类误差 - 源和目的特征差异
loss = loss_s1 + loss_s2 - loss_dis_t - loss_dis_s
loss.backward()
opt_c1.step()
opt_c2.step()
if i == 0:
print(
'Epoch: {} Stage2, Loss1: {:.6f}, Loss2: {:.6f} Disc S: {:.6f}, Disc T: {:.6f}'
.format(epoch, loss_s1.item(), loss_s2.item(),
loss_dis_s.item(), loss_dis_t.item()))
# 更新特征提取器
for _ in range(1):
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
feat_s = G(data_source)
output_s1 = C1(feat_s)
output_s2 = C2(feat_s)
loss_dis_s = discrepancy(output_s1, output_s2)
feat_t = G(data_target)
output_t1 = C1(feat_t)
output_t2 = C2(feat_t)
loss_dis_t = discrepancy(output_t1, output_t2)
loss_s1 = criterion(output_s1, label_source)
loss_s2 = criterion(output_s2, label_source)
loss = loss_s1 + loss_s2 + loss_dis_s + loss_dis_t
loss.backward()
opt_g.step()
if i == 0:
print(
'Epoch: {} Stage3, Loss1: {:.6f}\t Loss2: {:.6f}\t Discrepancy S: {:.6f}\t Discrepancy T: {:.6f}'
.format(epoch, loss_s1.item(), loss_s2.item(),
loss_dis_s.item(), loss_dis_t.item()))
# logging.debug('Train Epoch: {} Loss1: {:.6f}\t Loss2: {:.6f}\t Discrepancy: {:.6f}'.format(
# epoch, loss_s1.item(), loss_s2.item(), loss_dis.item()))
def train_onestep(G, C1, C2, config, epoch):
criterion = nn.CrossEntropyLoss().cuda()
G.train()
C1.train()
C2.train()
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
opt_g.zero_grad()
opt_c1.zero_grad()
opt_c2.zero_grad()
set_requires_grad(G, requires_grad=True)
set_requires_grad(C1, requires_grad=True)
set_requires_grad(C2, requires_grad=True)
feat_s = G(data_source)
output_s1 = C1(feat_s)
output_s2 = C2(feat_s)
loss_s1 = criterion(output_s1, label_source)
loss_s2 = criterion(output_s2, label_source)
loss_s = loss_s1 + loss_s2
# loss_s.backward(retain_variables=True)
##loss_s.backward()
set_requires_grad(G, requires_grad=False)
set_requires_grad(C1, requires_grad=True)
set_requires_grad(C2, requires_grad=True)
with torch.no_grad():
feat_t = G(data_target)
reverse_feature_t = ReverseLayerF.apply(feat_t, gamma)
output_t1 = C1(reverse_feature_t)
output_t2 = C2(reverse_feature_t)
loss_dis = -discrepancy(output_t1, output_t2)
##loss_dis.backward()
loss = loss_s + loss_dis
loss.backward()
opt_c1.step()
opt_c2.step()
opt_g.step()
if i % 20 == 0:
print(
'Train Epoch: {}, Loss1: {:.6f}\t Loss2: {:.6f}\t Discrepancy: {:.6f}'
.format(epoch, loss_s1.item(), loss_s2.item(),
loss_dis.item()))
for epoch in range(1, config['n_epochs'] + 1):
if config['mcd_onestep'] == 1:
train_onestep(G, C1, C2, config, epoch)
else:
train(G, C1, C2, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
#print('C1 on source_test_loader')
#logging.debug('C1 on source_test_loader')
#test(G, C1, config['source_test_loader'], epoch)
#print('C2 on source_test_loader')
#logging.debug('C2 on source_test_loader')
#test(G, C2, config['source_test_loader'], epoch)
print('C1 on target_test_loader')
logging.debug('C1 on target_test_loader')
test(G, C1, config['target_test_loader'], epoch)
print('C2 on target_test_loader')
logging.debug('C2 on target_test_loader')
test(G, C2, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(G, C1, config['target_test_loader'], res_dir,
epoch, config['models'])
draw_tsne(G,
C1,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=True)
draw_tsne(G,
C1,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=False)
def train_dann_mm2(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
bn=config['bn'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
# classifier = Predictor_deep(n_flattens=config['n_flattens'], n_hiddens=config['n_hiddens'], num_class=config['n_class'])
critic = Critic2(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
critic = critic.cuda()
summary(extractor, (1, 5120))
res_dir = os.path.join(config['res_dir'],
'snr{}-lr{}'.format(config['snr'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
set_log_config(res_dir)
logging.debug('train_dann_mm2')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(critic)
logging.debug(config)
criterion = torch.nn.CrossEntropyLoss()
optimizer_e = optim.Adam(extractor.parameters(), lr=config['lr'])
optimizer_cls = optim.Adam(classifier.parameters(), lr=config['lr'])
optimizer_critic = optim.Adam(critic.parameters(), lr=config['lr'])
def dann(input_data, alpha):
feature = extractor(input_data)
feature = feature.view(feature.size(0), -1)
reverse_feature = ReverseLayerF.apply(feature, alpha)
class_output = classifier(feature)
domain_output = critic(reverse_feature)
return class_output, domain_output, feature
def entropy(F1, feat, lamda, eta=1.0):
out_t1 = F1(feat, reverse=True, eta=-eta)
out_t1 = F.softmax(out_t1, dim=1)
loss_ent = -lamda * torch.mean(
torch.sum(out_t1 * (torch.log(out_t1 + 1e-5)), 1))
return loss_ent
def adentropy(F1, feat, lamda, eta=1.0):
out_t1 = F1(feat, reverse=True, eta=eta)
out_t1 = F.softmax(out_t1, dim=1)
loss_adent = lamda * torch.mean(
torch.sum(out_t1 * (torch.log(out_t1 + 1e-5)), 1))
return loss_adent
def entropy_softmax(output, lamda):
loss_ent = -lamda * torch.mean(
torch.sum(output * (torch.log(output + 1e-5)), 1))
return loss_ent
def adentropy_softmax(output, lamda):
loss_adent = lamda * torch.mean(
torch.sum(output * (torch.log(output + 1e-5)), 1))
return loss_adent
def train(extractor, classifier, critic, config, epoch):
extractor.train()
classifier.train()
critic.train()
gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for i in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if i % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
optimizer_e.zero_grad()
optimizer_cls.zero_grad()
optimizer_critic.zero_grad()
class_output_s, domain_output, _ = dann(input_data=data_source,
alpha=gamma)
err_s_label = criterion(class_output_s, label_source)
domain_label = torch.zeros(data_source.size(0)).long().cuda()
err_s_domain = criterion(domain_output, domain_label)
# Training model using target data
domain_label = torch.ones(data_target.size(0)).long().cuda()
class_output_t, domain_output, _ = dann(input_data=data_target,
alpha=gamma)
err_t_domain = criterion(domain_output, domain_label)
err = err_s_label + err_s_domain + err_t_domain
if i % 100 == 0:
print(
'err_s_label {:.2f}, err_s_domain {:.2f}, gamma {:.2f}, err_t_domain {:.2f}, total err {:.2f}'
.format(err_s_label.item(), err_s_domain.item(), gamma,
err_t_domain.item(), err.item()))
err.backward()
optimizer_e.step()
optimizer_cls.step()
optimizer_critic.step()
# minmax
optimizer_e.zero_grad()
optimizer_cls.zero_grad()
feature_t = extractor(data_target)
feature_t = feature_t.view(feature_t.size(0), -1)
# entropy_loss = adentropy(classifier, feature_t, 1)
entropy_loss = entropy(classifier, feature_t, 1)
entropy_loss.backward()
optimizer_e.step()
optimizer_cls.step()
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, critic, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
test(extractor, classifier, config['target_test_loader'], epoch)
if epoch % config['VIS_INTERVAL'] == 0:
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
config['models'])
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=True)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
config['models'],
separate=False)
def train_cdan_iw(config):
if config['network'] == 'inceptionv1':
extractor = InceptionV1(num_classes=32)
elif config['network'] == 'inceptionv1s':
extractor = InceptionV1s(num_classes=32)
else:
extractor = Extractor(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'])
classifier = Classifier(n_flattens=config['n_flattens'],
n_hiddens=config['n_hiddens'],
n_class=config['n_class'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
#summary(extractor, (1, 5120))
cdan_random = config['random_layer']
res_dir = os.path.join(
config['res_dir'],
'normal{}-{}-dilation{}-iw{}-lr{}'.format(config['normal'],
config['network'],
config['dilation'],
config['iw'], config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_cdan_iw')
#print(extractor)
#print(classifier)
print(config)
set_log_config(res_dir)
logging.debug('train_cdan')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(config)
if config['models'] == 'DANN_IW':
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
elif cdan_random:
random_layer = RandomLayer([config['n_flattens'], config['n_class']],
config['n_hiddens'])
ad_net = AdversarialNetwork(config['n_hiddens'], config['n_hiddens'])
random_layer.cuda()
else:
random_layer = None
ad_net = AdversarialNetwork(config['n_flattens'] * config['n_class'],
config['n_hiddens'])
ad_net = ad_net.cuda()
optimizer = torch.optim.Adam([{
'params': extractor.parameters(),
'lr': config['lr']
}, {
'params': classifier.parameters(),
'lr': config['lr']
}],
weight_decay=0.0001)
optimizer_ad = torch.optim.Adam(ad_net.parameters(),
lr=config['lr'],
weight_decay=0.0001)
print(ad_net)
extractor_path = os.path.join(res_dir, "extractor.pth")
classifier_path = os.path.join(res_dir, "classifier.pth")
adnet_path = os.path.join(res_dir, "adnet.pth")
def train(extractor, classifier, ad_net, config, epoch):
start_epoch = 0
extractor.train()
classifier.train()
ad_net.train()
iter_source = iter(config['source_train_loader'])
iter_target = iter(config['target_train_loader'])
len_source_loader = len(config['source_train_loader'])
len_target_loader = len(config['target_train_loader'])
num_iter = len_source_loader
for step in range(1, num_iter + 1):
data_source, label_source = iter_source.next()
data_target, _ = iter_target.next()
if step % len_target_loader == 0:
iter_target = iter(config['target_train_loader'])
if torch.cuda.is_available():
data_source, label_source = data_source.cuda(
), label_source.cuda()
data_target = data_target.cuda()
"""
add code start
"""
with torch.no_grad():
if config['models'] == 'CDAN_IW':
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
h_t = extractor(data_target)
h_t = h_t.view(h_t.size(0), -1)
source_preds = classifier(h_s)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
# print(softmax_output_s.shape)
# print(softmax_output_s.unsqueeze(2).shape)
# print(softmax_output_s)
# target_preds = classifier(h_t)
# softmax_output_t = nn.Softmax(dim=1)(target_preds)
# feature = torch.cat((h_s, h_t), 0)
# softmax_output = torch.cat((softmax_output_s, softmax_output_t), 0)
weights = torch.ones(softmax_output_s.shape).cuda()
weights = 1.0 * weights
weights = weights.unsqueeze(2)
# op_out = torch.bmm(softmax_output_s.unsqueeze(2), h_s.unsqueeze(1))
op_out = torch.bmm(weights, h_s.unsqueeze(1))
# gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
gamma = 1
ad_out = ad_net(op_out.view(
-1,
softmax_output_s.size(1) * h_s.size(1)),
gamma,
training=False)
# dom_entropy = loss_func.Entropy(ad_out)
dom_entropy = 1 + (torch.abs(0.5 - ad_out))**config['iw']
# dom_weight = dom_entropy / torch.sum(dom_entropy)
dom_weight = dom_entropy
elif config['models'] == 'DANN_IW':
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
# gamma = 2 / (1 + math.exp(-10 * (epoch) / config['n_epochs'])) - 1
gamma = 1
ad_out = ad_net(h_s, gamma, training=False)
# dom_entropy = 1-((torch.abs(0.5-ad_out))**config['iw'])
# dom_weight = dom_entropy
dom_weight = torch.ones(ad_out.shape).cuda()
#dom_entropy = loss_func.Entropy(dom_entropy)
# dom_weight = dom_entropy / torch.sum(dom_entropy)
"""
add code end
"""
optimizer.zero_grad()
optimizer_ad.zero_grad()
h_s = extractor(data_source)
h_s = h_s.view(h_s.size(0), -1)
h_t = extractor(data_target)
h_t = h_t.view(h_t.size(0), -1)
source_preds = classifier(h_s)
softmax_output_s = nn.Softmax(dim=1)(source_preds)
target_preds = classifier(h_t)
softmax_output_t = nn.Softmax(dim=1)(target_preds)
feature = torch.cat((h_s, h_t), 0)
softmax_output = torch.cat((softmax_output_s, softmax_output_t), 0)
cls_loss = nn.CrossEntropyLoss(reduction='none')(source_preds,
label_source)
cls_loss = torch.mean(dom_weight * cls_loss)
if epoch > start_epoch:
gamma = 2 / (1 + math.exp(-10 *
(epoch) / config['n_epochs'])) - 1
if config['models'] == 'CDAN_EIW':
entropy = loss_func.Entropy(softmax_output)
# print('softmax_output {}, entropy {}'.format(softmax_output.size(), entropy.size()))
d_loss = loss_func.CDAN(
[feature, softmax_output], ad_net, gamma, entropy,
loss_func.calc_coeff(num_iter * (epoch - start_epoch) +
step), random_layer)
elif config['models'] == 'CDAN_IW':
d_loss = loss_func.CDAN([feature, softmax_output], ad_net,
gamma, None, None, random_layer)
elif config['models'] == 'DANN_IW':
d_loss = loss_func.DANN(feature, ad_net, gamma)
else:
raise ValueError('Method cannot be recognized.')
else:
d_loss = 0
loss = cls_loss + d_loss
loss.backward()
optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (step) % 20 == 0:
print('Train Epoch {} closs {:.6f}, dloss {:.6f}, Loss {:.6f}'.
format(epoch, cls_loss.item(), d_loss.item(),
loss.item()))
best_accuracy = 0
best_model_index = -1
for epoch in range(1, config['n_epochs'] + 1):
train(extractor, classifier, ad_net, config, epoch)
if epoch % config['TEST_INTERVAL'] == 0:
# print('test on source_test_loader')
# test(extractor, classifier, config['source_test_loader'], epoch)
print('test on target_test_loader')
accuracy = test(extractor, classifier,
config['target_test_loader'], epoch)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_model_index = epoch
torch.save(extractor.state_dict(), extractor_path)
torch.save(classifier.state_dict(), classifier_path)
torch.save(ad_net.state_dict(), adnet_path)
print(
'epoch {} accuracy: {:.6f}, best accuracy {:.6f} on epoch {}'.
format(epoch, accuracy, best_accuracy, best_model_index))
if epoch % config['VIS_INTERVAL'] == 0:
title = config['models']
draw_confusion_matrix(extractor, classifier,
config['target_test_loader'], res_dir, epoch,
title)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
draw_tsne(extractor,
classifier,
config['source_train_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=False)