def train_cdan_vat(config):
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()
xi = 1e-06
ip = 1
eps = 15
vat = VirtualAdversarialPerturbationGenerator(extractor, classifier, xi=xi, eps=eps, ip=ip)
cdan_random = config['random_layer']
res_dir = os.path.join(config['res_dir'], 'random{}-bs{}-lr{}'.format(cdan_random, config['batch_size'], 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'])
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()
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)
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)
elif config['models'] == 'CDAN_VAT':
# 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)
d_loss = loss_func.CDAN([feature, softmax_output], ad_net, gamma, None, None, random_layer)
# vat_loss = loss_func.VAT(vat, data_target, extractor, classifier, target_consistency_criterion)
# vat_adv, clean_vat_logits = vat(data_target)
# vat_adv_inputs = data_target + vat_adv
# adv_vat_features = extractor(vat_adv_inputs)
# adv_vat_logits = classifier(adv_vat_features)
# target_vat_loss = target_consistency_criterion(adv_vat_logits, clean_vat_logits)
# vat_loss = target_vat_loss_weight * target_vat_loss
else:
raise ValueError('Method cannot be recognized.')
else:
d_loss = 0
loss = cls_loss + d_loss
loss.backward()
optimizer.step()
vat_adv, clean_vat_logits = vat(data_target)
vat_adv_inputs = data_target + vat_adv
adv_vat_features = extractor(vat_adv_inputs)
adv_vat_logits = classifier(adv_vat_features)
target_vat_loss = target_consistency_criterion(adv_vat_logits, clean_vat_logits)
vat_loss = target_vat_loss_weight * target_vat_loss
vat_loss.backward()
# optimizer.step()
if epoch > start_epoch:
optimizer_ad.step()
if (step) % 20 == 0:
print('Train Epoch {} closs {:.6f}, dloss {:.6f}, vat_loss {:.6f}, Loss {:.6f}'.format(epoch, cls_loss.item(), d_loss.item(), vat_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_test_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)
else:
print('no saved model found')
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_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_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_cdan_ican(config):
BATCH_SIZE = config['batch_size']
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'])
disc_activate = Contrast_ReLU_activate(INI_DISC_WEIGHT_SCALE,
INI_DISC_BIAS)
cdan_random = config['random_layer']
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 = AdversarialNetwork(config['n_flattens'], config['n_hiddens'])
if torch.cuda.is_available():
extractor = extractor.cuda()
classifier = classifier.cuda()
disc_activate = disc_activate.cuda()
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'])
pseudo_optimizer = torch.optim.Adam(disc_activate.parameters(),
lr=config['lr'])
class_criterion = nn.CrossEntropyLoss()
res_dir = os.path.join(
config['res_dir'],
'random{}-bs{}-lr{}'.format(cdan_random, config['batch_size'],
config['lr']))
if not os.path.exists(res_dir):
os.makedirs(res_dir)
print('train_cdan_ican')
print(extractor)
print(classifier)
print(ad_net)
print(config)
set_log_config(res_dir)
logging.debug('train_cdan_ican')
logging.debug(extractor)
logging.debug(classifier)
logging.debug(ad_net)
logging.debug(config)
def select_samples_ican(extractor, classifier, ad_net, disc_activate,
config, epoch, epoch_acc_s):
set_training_mode(extractor, False)
set_training_mode(classifier, False)
set_training_mode(ad_net, False)
set_training_mode(disc_activate, False)
Pseudo_set = []
confid_threshold = 1 / (1 + np.exp(-2.4 * epoch_acc_s))
total_pseudo_errors = 0
# 为什么在target测试集上进行？
# for target_inputs, target_labels in iter(config['target_test_loader']):
for target_inputs, target_labels in iter(
config['target_train_loader']):
target_inputs = target_inputs.cuda()
# 论文中target的domain label是1
domain_labels_t = torch.FloatTensor([0.] *
len(target_inputs)).cuda()
embeddings = extractor(target_inputs)
class_t = classifier(embeddings)
domain_out_t = ad_net(embeddings, training=False)
disc_weight_t, w_t, b_t = disc_activate(domain_out_t,
domain_labels_t)
top_prob, preds_t = torch.max(class_t, 1)
for i in range(len(disc_weight_t)):
if disc_weight_t[i] > b_t and top_prob[i] >= float(
confid_threshold):
s_tuple = (target_inputs[i].cpu(),
(preds_t[i].cpu(), float(disc_weight_t[i])))
Pseudo_set.append(s_tuple)
total_pseudo_errors += preds_t.eq(target_labels.cuda()).cpu().sum()
# 每个pseudo_set样本中包括[features, category_class_predict, domain_weight_predict], [特征，预测的类标，预测的domain权重]
# print("Pseudo error/total = {}/{}, confid_threshold: {:.4f}".format(total_pseudo_errors, len(Pseudo_set),
# confid_threshold))
print(
'Epoch {}, Stage Select_Sample, accuracy {}, confident threshold {}, pseudo number {}, b_t {}'
.format(epoch, epoch_acc_s, confid_threshold, len(Pseudo_set),
b_t))
draw_dict['confid_threshold_point'].append(
float("%.4f" % confid_threshold))
return Pseudo_set
# TODO: 为什么不在上一个函数中直接更新呢？选择pseudo-set之后就更新disc-activate的模型参数，完全可以合并成一步
def update_ican(extractor, classifier, ad_net, disc_activate, config,
Pseudo_set, epoch):
if len(Pseudo_set) == 0:
return
set_training_mode(extractor, False)
set_training_mode(classifier, False)
set_training_mode(ad_net, False)
set_training_mode(disc_activate, True)
pseudo_batch_count = 0
pseudo_sample_count = 0
pseudo_epoch_loss = 0.0
pseudo_epoch_acc = 0
pseudo_epoch_corrects = 0
pseudo_avg_loss = 0.0
# TODO: 每次从pseudo-set中取半个batch-size
pseudo_loader = torch.utils.data.DataLoader(Pseudo_set,
batch_size=int(BATCH_SIZE /
2),
shuffle=True)
for pseudo_inputs, pseudo_labels in pseudo_loader:
pseudo_batch_count += 1
pseudo_sample_count += len(pseudo_inputs)
pseudo_labels, pseudo_weights = pseudo_labels[0], pseudo_labels[1]
pseudo_inputs, pseudo_labels = pseudo_inputs.cuda(
), pseudo_labels.cuda()
domain_labels = torch.FloatTensor([0.] * len(pseudo_inputs)).cuda()
embeddings = extractor(pseudo_inputs)
pseudo_class = classifier(embeddings)
pseudo_domain_out = ad_net(embeddings, training=False)
pseudo_disc_weight, pseudo_ww, pseudo_bb = disc_activate(
pseudo_domain_out, domain_labels)
pseudo_optimizer.zero_grad()
# TODO：为什么不用这个pseudo_preds, 而要用上个函数保存的结果呢？
_, pseudo_preds = torch.max(pseudo_class, 1)
# pseudo_class：未经过softmax的类分类概率
# pseudo_labels: 经过softmax的类标签
# pseudo_disc_weight：样本的领域权重
# TODO：检查pseudo_disc_weight的形状
# pseudo_class_loss = compute_new_loss(pseudo_class, pseudo_labels, pseudo_disc_weight)
pseudo_class_loss = compute_new_loss(pseudo_class, pseudo_preds,
pseudo_disc_weight)
# pseudo_class_loss = class_criterion(pseudo_class, pseudo_preds)
pseudo_epoch_loss += float(pseudo_class_loss)
# 这个正确率没有意义
# pseudo_preds 是pseudo_class的最大值，是target train的预测值
# pseudo_labels 是上一个函数（选择pseudo-set时）计算出来的，同样的公式
pseudo_epoch_corrects += int(
torch.sum(pseudo_preds.squeeze() == pseudo_labels.squeeze()))
pseudo_loss = pseudo_class_loss
pseudo_loss.backward()
pseudo_optimizer.step()
epoch_discrim_lambda = 1.0 / (abs(pseudo_ww)**(1. / 4))
epoch_discrim_bias = pseudo_bb
pseudo_avg_loss = pseudo_epoch_loss / pseudo_batch_count
pseudo_epoch_acc = pseudo_epoch_corrects / pseudo_sample_count
print(
'Epoch {}, Phase: {}, Loss: {:.4f} Acc: {:.4f} Disc_Lam: {:.6f} Disc_bias: {:.4f} '
.format(epoch, 'Pseudo_train', pseudo_avg_loss, pseudo_epoch_acc,
epoch_discrim_lambda, epoch_discrim_bias))
def prepare_dataset(epoch, pseudo_set):
dset_loaders = {}
dset_loaders['source'] = config['source_train_loader']
source_size = len(config['source_train_loader'])
pseudo_size = len(pseudo_set)
# source_batches_per_epoch = np.floor(source_size * 2 / BATCH_SIZE).astype(np.int16)
# total_epochs = config['n_epochs']
if pseudo_size == 0:
dset_loaders['pseudo'] = []
dset_loaders['pseudo_source'] = []
# source_batchsize = int(BATCH_SIZE / 2)
source_batchsize = BATCH_SIZE
pseudo_batchsize = 0
else:
# source_batchsize = int(int(BATCH_SIZE / 2) * source_size
# / (source_size + pseudo_size))
# if source_batchsize == int(BATCH_SIZE / 2):
# source_batchsize -= 1
# if source_batchsize < int(int(BATCH_SIZE / 2) / 2):
# source_batchsize = int(int(BATCH_SIZE / 2) / 2)
# pseudo_batchsize = int(BATCH_SIZE / 2) - source_batchsize
# print('source_batchsize {}, pseudo_batchsize {}'.format(source_batchsize, pseudo_batchsize))
# dset_loaders['pseudo'] = torch.utils.data.DataLoader(pseudo_set,
# batch_size=pseudo_batchsize, shuffle=True)
# dset_loaders['pseudo_source'] = config['source_train_loader']
#
# 重新修改，按照source_train中每个epoch的batch数量，计算pseudo-set的batchsize
pseudo_batchsize = int(
np.floor(pseudo_size / len(config['source_train_loader'])))
dset_loaders['pseudo'] = torch.utils.data.DataLoader(
pseudo_set,
batch_size=pseudo_batchsize,
shuffle=True,
drop_last=False)
dset_loaders['pseudo_source'] = config['source_train_loader']
source_batchsize = BATCH_SIZE
print(
'Epoch {}, Stage prepare_dataset, pseudo_size {}, num batch each epoch: {}, pseudo_batchsize {}'
.format(epoch, pseudo_size, source_size, pseudo_batchsize))
target_dict = [(i, j) for (i, j) in config['target_train_loader']]
if pseudo_size > 0:
pseudo_dict = [(i, j) for (i, j) in dset_loaders['pseudo']]
pseudo_source_dict = [(i, j)
for (i, j) in dset_loaders['pseudo_source']]
else:
pseudo_dict = []
pseudo_source_dict = []
# total_iters = source_batches_per_epoch * pre_epochs + \
# source_batches_per_epoch * (total_epochs - pre_epochs) * \
# BATCH_SIZE / (source_batchsize * 2)
# total_iters = source_batches_per_epoch * (total_epochs) * BATCH_SIZE / (source_batchsize * 2)
return dset_loaders, target_dict, pseudo_dict, pseudo_source_dict, source_batchsize, pseudo_batchsize
def train(extractor, classifier, ad_net, disc_activate, config, epoch):
start_epoch = 0
extractor.train()
classifier.train()
ad_net.train()
disc_activate.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()
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)
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)
elif config['models'] == 'CDAN_ICAN':
d_loss = loss_func.CDAN([feature, softmax_output], ad_net,
gamma, None, None, random_layer)
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()))
# def do_forward(extractor, classifier, ad_net, disc_activate, src_features, all_features, labels):
# # 预测source features的class labels
# bottle = extractor(src_features)
# class_pred = classifier(bottle)
# dom_pred = ad_net(bottle)
# return class_pred, dom_pred.squeeze(1)
def do_training(dset_loaders, target_dict, source_batchsize,
pseudo_batchsize, pseudo_dict, pseudo_source_dict):
batch_count = 0
target_pointer = 0
target_pointer = 0
pseudo_pointer = 0
pseudo_source_pointer = 0
INI_MAIN_THRESH = -0.8
# pre_epochs = 10
pre_epochs = 0
set_training_mode(extractor, True)
set_training_mode(classifier, True)
set_training_mode(ad_net, True)
set_training_mode(disc_activate, False)
# class_count = 0
# epoch_loss = 0.0
# epoch_corrects = 0
# domain_epoch_loss = 0.0
# ini_w_main = torch.FloatTensor([float(INI_MAIN_THRESH)]).cuda()
# epoch_batch_count = 0
# total_epoch_loss = 0.0
# domain_epoch_corrects = 0
# domain_counts = 0
for data in dset_loaders['source']:
inputs, labels = data
batch_count += 1
# ---------------- reset exceeded datasets --------------------
if target_pointer >= len(target_dict) - 1:
target_pointer = 0
target_dict = [(i, j)
for (i, j) in config['target_train_loader']]
target_inputs = target_dict[target_pointer][0]
if epoch <= pre_epochs:
# 训练CAN，使用source_train和target_train，target_train不经筛选，全部使用
# -------------------- pretrain model -----------------------
domain_inputs = torch.cat((inputs, target_inputs), 0)
# domain_labels = torch.FloatTensor([1.]*BATCH_SIZE + [0.]*BATCH_SIZE)
domain_labels = torch.FloatTensor([1.] * inputs.size(0) +
[0.] * target_inputs.size(0))
domain_inputs, domain_labels = domain_inputs.cuda(
), domain_labels.cuda()
inputs, labels = inputs.cuda(), labels.cuda()
# print('inputs {}, target_inputs {}, domain_inputs {}, domain_labels {}'.format(inputs.size(), target_inputs.size(), domain_inputs.size(), domain_labels.size()))
# source数据集上的分类结果
class_outputs = classifier(extractor(inputs))
# 在source和target数据集上判断domain分类
domain_outputs = ad_net(extractor(domain_inputs)).squeeze()
target_pointer += 1
# epoch_discrim_bias = 0.5
# ------------ training classification statistics --------------
criterion = nn.CrossEntropyLoss()
class_loss = criterion(class_outputs, labels)
else:
# -------------- train with pseudo sample model -------------
# target域使用经过筛选的pseudo-set数据
pseudo_weights = torch.FloatTensor([])
pseudo_size = len(pseudo_dict)
# 重置索引位置
if (pseudo_pointer >=
len(pseudo_dict) - 1) and (len(pseudo_dict) != 0):
pseudo_pointer = 0
pseudo_dict = [(i, j) for (i, j) in dset_loaders['pseudo']]
if (pseudo_source_pointer >= len(pseudo_source_dict) - 1) and (
len(pseudo_source_dict) != 0):
pseudo_source_pointer = 0
pseudo_source_dict = [
(i, j) for (i, j) in dset_loaders['pseudo_source']
]
if pseudo_size == 0:
# 如果pseudo-set为空，那还是使用全部source_train和target_train
domain_inputs = torch.cat((inputs, target_inputs), 0)
# domain_labels = torch.FloatTensor([1.]*int(BATCH_SIZE / 2)+
# [0.]*int(BATCH_SIZE / 2))
domain_labels = torch.FloatTensor([1.] * inputs.size(0) +
[0.] *
target_inputs.size(0))
fuse_inputs = inputs
fuse_labels = labels
else:
pseudo_inputs, pseudo_labels, pseudo_weights = pseudo_dict[pseudo_pointer][0], \
pseudo_dict[pseudo_pointer][1][0], pseudo_dict[pseudo_pointer][1][1]
pseudo_source_inputs = pseudo_source_dict[
pseudo_source_pointer][0]
# TODO: 为什么要这么干？source + pseudo + target + source
# domain_inputs = torch.cat((inputs, pseudo_inputs, target_inputs, pseudo_source_inputs),0)
# domain_labels = torch.FloatTensor([1.]*inputs.size(0) + [0.]*pseudo_inputs.size(0) +
# [0.]*target_inputs.size(0)+[1.]*pseudo_source_inputs.size(0))
domain_inputs = torch.cat((inputs, pseudo_inputs), 0)
domain_labels = torch.FloatTensor([1.] * inputs.size(0) +
[0.] *
pseudo_inputs.size(0))
fuse_inputs = torch.cat((inputs, pseudo_inputs), 0)
fuse_labels = torch.cat((labels, pseudo_labels), 0)
# print('inputs {}, pseudo_inputs {}, target_inputs {}, domain_inputs {}'.format(inputs.size(), pseudo_inputs.size(), target_inputs.size(), domain_inputs.size()))
# print('domain_labels {}, fuse_inputs {}, fuse_labels {}'.format(domain_labels.size(), fuse_inputs.size(), fuse_labels.size()))
inputs, labels = fuse_inputs.cuda(), fuse_labels.cuda()
domain_inputs, domain_labels = domain_inputs.cuda(
), domain_labels.cuda()
source_weight_tensor = torch.FloatTensor([1.] *
source_batchsize)
pseudo_weights_tensor = pseudo_weights.float()
class_weights_tensor = torch.cat(
(source_weight_tensor, pseudo_weights_tensor), 0)
dom_weights_tensor = torch.FloatTensor([0.] *
source_batchsize +
[1.] * pseudo_batchsize)
ini_weight = torch.cat(
(class_weights_tensor, dom_weights_tensor),
0).squeeze().cuda()
class_outputs = classifier(extractor(inputs))
domain_outputs = ad_net(extractor(domain_inputs)).squeeze()
# ------------ training classification statistics --------------
# _, preds = torch.max(class_outputs, 1)
# class_count += len(preds)
class_loss = compute_new_loss(class_outputs, labels,
ini_weight)
# epoch_loss += float(class_loss)
# epoch_corrects += int(torch.sum(preds.squeeze() == labels.squeeze()))
target_pointer += 1
pseudo_pointer += 1
pseudo_source_pointer += 1
# zero the parameter gradients
optimizer.zero_grad()
optimizer_ad.zero_grad()
# ----------- calculate pred domain labels and losses -----------
domain_criterion = nn.BCEWithLogitsLoss()
domain_labels = domain_labels.squeeze()
domain_loss = domain_criterion(domain_outputs, domain_labels)
# domain_epoch_loss += float(domain_loss)
# ------ calculate pseudo predicts and losses with weights and threshold lambda -------
total_loss = class_loss + 1.0 * domain_loss
# total_epoch_loss += float(total_loss)
print('class_loss {}, domain_loss {}'.format(
class_loss.item(), domain_loss.item()))
# ------- backward + optimize in training and Pseudo-training phase -------
total_loss.backward()
optimizer.step()
optimizer_ad.step()
def train_ican(extractor, classifier, ad_net, disc_activate, config,
epoch):
# start_epoch = 0
# 1. 计算在source上的准确度，用于选择伪标签
accuracy_s = test(extractor, classifier, config['source_test_loader'],
epoch)
# 2. 计算伪标签数据集
pseu_set = select_samples_ican(extractor, classifier, ad_net,
disc_activate, config, epoch,
accuracy_s)
# 3. 使用伪数据集训练disc_activate，更新disc threshold
update_ican(extractor, classifier, ad_net, disc_activate, config,
pseu_set, epoch)
# 4. 准备最终训练ican所用的数据集，将source dataset和pseudo set合并
dset_loaders, target_dict, pseudo_dict, pseudo_source_dict, source_batchsize, pseudo_batchsize = prepare_dataset(
epoch, pseu_set)
# 5. train
# do_training()
do_training(dset_loaders, target_dict, source_batchsize,
pseudo_batchsize, pseudo_dict, pseudo_source_dict)
# 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()
# 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)
# 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)
# elif config['models'] == 'CDAN_ICAN':
# d_loss = loss_func.CDAN([feature, softmax_output], ad_net, gamma, None, None, random_layer)
# 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()))
# function done
for epoch in range(1, config['n_epochs'] + 1):
train_ican(extractor, classifier, ad_net, disc_activate, 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 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_test_loader'],
config['target_test_loader'],
res_dir,
epoch,
title,
separate=True)
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)