def main():
# parameters
learning_rate = 0.001
num_epochs = 100
batch_size = 20
batch_size_test = 20
feature_size = 2048 * 1 * 1
# create the save log file
print("Create the directory")
if not os.path.exists("./save"):
os.makedirs("./save")
if not os.path.exists("./logfile"):
os.makedirs("./logfile")
if not os.path.exists("./logfile/MTL"):
os.makedirs("./logfile/MTL")
# load my Dataset
inf_csv_path = [
"./dataset_public/infograph/infograph_train.csv",
"./dataset_public/infograph/infograph_test.csv"
]
qdr_csv_path = [
"./dataset_public/quickdraw/quickdraw_train.csv",
"./dataset_public/quickdraw/quickdraw_test.csv"
]
skt_csv_path = [
"./dataset_public/sketch/sketch_train.csv",
"./dataset_public/sketch/sketch_test.csv"
]
rel_csv_path = ["./dataset_public/real/real_train.csv"]
test_path = "./dataset_public/test"
inf_train_dataset = Dataset.Dataset(csv_path=inf_csv_path[0], argu=True)
inf_train_loader = DataLoader(inf_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
qdr_train_dataset = Dataset.Dataset(csv_path=qdr_csv_path[0], argu=True)
qdr_train_loader = DataLoader(qdr_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
skt_train_dataset = Dataset.Dataset(csv_path=skt_csv_path[0], argu=True)
skt_train_loader = DataLoader(skt_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
rel_train_dataset = Dataset.Dataset(csv_path=rel_csv_path[0],
sample=True,
argu=True)
rel_train_loader = DataLoader(rel_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
valid_dataset = Dataset.Valid_Dataset(csv_path=rel_csv_path[0],
sample=True)
valid_loader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
test_dataset = Dataset.Dataset(csv_path=test_path,
mode="test",
filename=True)
test_loader = DataLoader(test_dataset,
batch_size=batch_size_test,
shuffle=False,
num_workers=1)
print('the source dataset has %d size.' % (len(rel_train_dataset)))
print('the valid dataset has %d size.' % (len(valid_dataset)))
print('the target dataset has %d size.' % (len(test_dataset)))
print('the batch_size is %d' % (batch_size))
# Pre-train models
modules = list(models.resnet152(pretrained=True).children())[:-1]
encoder = nn.Sequential(*modules)
#encoder = model.Encoder()
classifier_0 = model.Classifier(feature_size)
classifier_1 = model.Classifier(feature_size)
classifier_2 = model.Classifier(feature_size)
moe_classifier = model.Moe_Classifier()
domain_classifier_0 = model.Domain_classifier(feature_size,
number_of_domain)
domain_classifier_1 = model.Domain_classifier(feature_size,
number_of_domain)
domain_classifier_2 = model.Domain_classifier(feature_size,
number_of_domain)
# GPU enable
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print('Device used:', device)
if torch.cuda.is_available():
encoder = encoder.to(device)
classifier_0 = classifier_0.to(device)
classifier_1 = classifier_1.to(device)
classifier_2 = classifier_2.to(device)
moe_classifier = moe_classifier.to(device)
domain_classifier_0 = domain_classifier_0.to(device)
domain_classifier_1 = domain_classifier_1.to(device)
domain_classifier_2 = domain_classifier_2.to(device)
# setup optimizer
"""
optimizer_encoder = optim.SGD(
list(encoder.parameters()),lr=learning_rate,momentum=0.9)
optimizer_classifier = optim.SGD(
list(classifier_0.parameters())+
list(classifier_1.parameters())+
list(classifier_2.parameters())+
list(moe_classifier.parameters()),lr=learning_rate,momentum=0.9)
optimizer_domain = optim.SGD([{'params': domain_classifier_0.parameters()},
{'params': domain_classifier_1.parameters()},
{'params': domain_classifier_2.parameters()}], lr= learning_rate,momentum=0.9)
"""
optimizer_encoder = optim.Adam([
{
'params': encoder.parameters()
},
],
weight_decay=1e-4,
lr=learning_rate)
optimizer_classifier = optim.Adam(
list(classifier_0.parameters()) + list(classifier_1.parameters()) +
list(classifier_2.parameters()) + list(moe_classifier.parameters()),
weight_decay=1e-4,
lr=learning_rate)
optimizer_domain = optim.Adam([{
'params': domain_classifier_0.parameters()
}, {
'params': domain_classifier_1.parameters()
}, {
'params': domain_classifier_2.parameters()
}],
lr=learning_rate,
weight_decay=1e-4)
#Lossfunction
moe_criterion = nn.BCELoss()
mtl_criterion = nn.NLLLoss()
D_loss_list = []
L_loss_list = []
sum_src_acc_list = []
sum_trg_acc_list = []
sum_label_acc_loist = []
sum_test_acc_list = []
print("Starting training...")
best_acc = 0.
valid_acc = 0.0
for epoch in range(num_epochs):
print("Epoch:", epoch + 1)
encoder.train()
classifier_0.train()
classifier_1.train()
classifier_2.train()
moe_classifier.train()
domain_classifier_0.train()
domain_classifier_1.train()
domain_classifier_2.train()
epoch_D_loss = 0.0
epoch_C_loss = 0.0
epoch_C_moe_loss = 0.0
sum_src_acc = 0.0
sum_trg_acc = 0.0
sum_label_acc = 0.0
sum_test_acc = 0.0
tmp_valid_acc = 0.0
if (epoch + 1) in [5, 10, 15, 20, 25, 30, 35, 40, 45]:
for optimizer_t in optimizer_domain.param_groups:
optimizer_t['lr'] /= 1.7
optimizer_encoder.param_groups[0]['lr'] /= 1.7
optimizer_classifier.param_groups[0]['lr'] /= 1.7
train_loader = [
inf_train_loader, qdr_train_loader, skt_train_loader,
rel_train_loader
]
len_loader = min([
len(train_loader[0]),
len(train_loader[1]),
len(train_loader[2]),
len(train_loader[3])
])
for index, (inf, qdr, skt, rel, test) in enumerate(
zip(train_loader[0], train_loader[1], train_loader[2],
train_loader[3], valid_loader)):
optimizer_classifier.zero_grad()
optimizer_encoder.zero_grad()
optimizer_domain.zero_grad()
# colculate the lambda_
p = (index + len_loader * epoch) / (len_loader * num_epochs)
lambda_ = 5.0 / (1. + np.exp(-10 * p)) - 1.0
s1_imgs, s1_labels = skt
s2_imgs, s2_labels = qdr
s3_imgs, s3_labels = inf
t1_imgs, _ = rel
from_s1_labels = Variable(torch.zeros(len(s1_imgs))).to(device)
from_s2_labels = Variable(torch.zeros(len(s2_imgs))).to(device)
from_s3_labels = Variable(torch.zeros(len(s3_imgs))).to(device)
from_t1_labels = Variable(torch.ones(len(t1_imgs))).to(device)
s1_imgs = Variable(s1_imgs).to(device)
s1_labels = Variable(s1_labels.view(-1)).to(device)
s2_imgs = Variable(s2_imgs).to(device)
s2_labels = Variable(s2_labels.view(-1)).to(device)
s3_imgs = Variable(s3_imgs).to(device)
s3_labels = Variable(s3_labels.view(-1)).to(device)
t1_imgs = Variable(t1_imgs).to(device)
s1_feature = encoder(s1_imgs)
s2_feature = encoder(s2_imgs)
s3_feature = encoder(s3_imgs)
t1_feature = encoder(t1_imgs)
# Testing
test_imgs, test_labels = test
test_imgs = Variable(test_imgs).to(device)
test_labels = Variable(test_labels.view(-1)).to(device)
test_feature = encoder(test_imgs)
test_output_0 = classifier_0(test_feature)
test_output_1 = classifier_1(test_feature)
test_output_2 = classifier_2(test_feature)
test_output = moe_classifier(test_output_0, test_output_1,
test_output_2)
test_preds = test_output.argmax(1).cpu()
test_acc = np.mean(
(test_preds.detach().cpu() == test_labels.cpu()).numpy())
tmp_valid_acc += test_acc
# Training Classifier network (loss_mtl)
s1_output = classifier_0(s1_feature)
s2_output = classifier_1(s2_feature)
s3_output = classifier_2(s3_feature)
s1_preds = s1_output.argmax(1).cpu()
s2_preds = s2_output.argmax(1).cpu()
s3_preds = s3_output.argmax(1).cpu()
s1_acc = np.mean(
(s1_preds.detach().cpu() == s1_labels.cpu()).numpy())
s2_acc = np.mean(
(s2_preds.detach().cpu() == s2_labels.cpu()).numpy())
s3_acc = np.mean(
(s3_preds.detach().cpu() == s3_labels.cpu()).numpy())
s1_c_loss = mtl_criterion(s1_output, s1_labels)
s2_c_loss = mtl_criterion(s2_output, s2_labels)
s3_c_loss = mtl_criterion(s3_output, s3_labels)
mtl_loss = s1_c_loss + s2_c_loss + s3_c_loss
# Domain_classifier network with source domain (loss_adv)
s1_domain_output = domain_classifier_0(s1_feature, lambda_)
s2_domain_output = domain_classifier_1(s2_feature, lambda_)
s3_domain_output = domain_classifier_2(s3_feature, lambda_)
s1_domain_acc = np.mean(
(s1_domain_output.detach().cpu() <= 0.5).numpy())
s2_domain_acc = np.mean(
(s2_domain_output.detach().cpu() <= 0.5).numpy())
s3_domain_acc = np.mean(
(s3_domain_output.detach().cpu() <= 0.5).numpy())
s1_d_loss = moe_criterion(s1_domain_output, from_s1_labels)
s2_d_loss = moe_criterion(s2_domain_output, from_s2_labels)
s3_d_loss = moe_criterion(s3_domain_output, from_s3_labels)
D_loss_src = s1_d_loss + s2_d_loss + s3_d_loss
#print(D_loss_src.item())
# Domain_classifier network with target domain (loss_adv)
t1_domain_0_output = domain_classifier_0(t1_feature, lambda_)
t1_domain_1_output = domain_classifier_1(t1_feature, lambda_)
t1_domain_2_output = domain_classifier_2(t1_feature, lambda_)
t1_domain_0_acc = np.mean(
(t1_domain_0_output.detach().cpu() > 0.5).numpy())
t1_domain_1_acc = np.mean(
(t1_domain_1_output.detach().cpu() > 0.5).numpy())
t1_domain_2_acc = np.mean(
(t1_domain_2_output.detach().cpu() > 0.5).numpy())
D0_loss_trg = moe_criterion(t1_domain_0_output, from_t1_labels)
D1_loss_trg = moe_criterion(t1_domain_1_output, from_t1_labels)
D2_loss_trg = moe_criterion(t1_domain_2_output, from_t1_labels)
D_loss_trg = D0_loss_trg + D1_loss_trg + D2_loss_trg
adv_loss = D_loss_src + D_loss_trg
# Moe combination
s1_output_0 = classifier_0(s1_feature)
s1_output_1 = classifier_1(s1_feature)
s1_output_2 = classifier_2(s1_feature)
s2_output_0 = classifier_0(s2_feature)
s2_output_1 = classifier_1(s2_feature)
s2_output_2 = classifier_2(s2_feature)
s3_output_0 = classifier_0(s3_feature)
s3_output_1 = classifier_1(s3_feature)
s3_output_2 = classifier_2(s3_feature)
s1_output_moe = moe_classifier(s1_output_0, s1_output_1,
s1_output_2)
s2_output_moe = moe_classifier(s2_output_0, s2_output_1,
s2_output_2)
s3_output_moe = moe_classifier(s3_output_0, s3_output_1,
s3_output_2)
s1_preds_moe = s1_output_moe.argmax(1).cpu()
s2_preds_moe = s2_output_moe.argmax(1).cpu()
s3_preds_moe = s3_output_moe.argmax(1).cpu()
s1_acc = np.mean(
(s1_preds_moe.detach().cpu() == s1_labels.cpu()).numpy())
s2_acc = np.mean(
(s2_preds_moe.detach().cpu() == s2_labels.cpu()).numpy())
s3_acc = np.mean(
(s3_preds_moe.detach().cpu() == s3_labels.cpu()).numpy())
s1_c_loss_moe = mtl_criterion(s1_output_moe, s1_labels)
s2_c_loss_moe = mtl_criterion(s2_output_moe, s2_labels)
s3_c_loss_moe = mtl_criterion(s3_output_moe, s3_labels)
moe_loss = s1_c_loss_moe + s2_c_loss_moe + s3_c_loss_moe
moe_acc = (s1_acc + s2_acc + s3_acc) / 3.
loss = 0.4 * mtl_loss + 0.4 * moe_loss + 2 * adv_loss
epoch_D_loss += adv_loss.item()
epoch_C_loss += mtl_loss.item()
epoch_C_moe_loss += moe_loss.item()
D_trg_acc = (t1_domain_0_acc + t1_domain_1_acc +
t1_domain_2_acc) / 3.
D_src_acc = (s1_domain_acc + s2_domain_acc + s3_domain_acc) / 3.
loss.backward()
optimizer_domain.step()
optimizer_classifier.step()
optimizer_encoder.step()
if (index + 1) % 10 == 0:
print(
'Epoch [%d/%d], Iter [%d/%d] C_loss %.4f , D_loss %.4f ,LR: %.5f'
% (epoch + 1, num_epochs, index + 1, len_loader,
mtl_loss.item(), adv_loss.item(),
optimizer_encoder.param_groups[0]['lr']))
print(
"====> Acc %.4f %.4f %.4f Domain Acc: %.4f %.4f ,Test Acc: %.4f"
% (s1_acc, s2_acc, s3_acc, D_src_acc, D_trg_acc, test_acc))
print("====> Loss_moe %.4f ,Moe acc %.4f" %
(moe_loss.item(), moe_acc))
print('Validing: Acc %.4f ' % (tmp_valid_acc / len_loader))
if tmp_valid_acc / len_loader > best_acc:
best_acc = tmp_valid_acc / len_loader
print('Find best: Acc %.6f ' % (best_acc))
print('Best: Acc %.6f ' % (best_acc))
if tmp_valid_acc / len_loader >= best_acc and epoch + 1 > 5 and tmp_valid_acc / len_loader > 0.3:
best_acc = tmp_valid_acc / len_loader
print('Find best: Acc %.6f ' % (best_acc))
save_checkpoint(
'./save/encoder-multiD-rel-%.4f-%03i.pth' % (best_acc, epoch),
encoder)
save_checkpoint(
'./save/classifier-multiD-rel-%.4f-%03i.pth' %
(best_acc, epoch), moe_classifier)
save_checkpoint(
'./save/classifier-multiD-rel-%.4f-%03i.pth' %
(best_acc, epoch), classifier_0)
save_checkpoint(
'./save/classifier-multiD-rel-%.4f-%03i.pth' %
(best_acc, epoch), classifier_1)
save_checkpoint(
'./save/classifier-multiD-rel-%.4f-%03i.pth' %
(best_acc, epoch), classifier_2)
test_acc = 0.
test_loss = 0.
encoder.eval()
moe_classifier.eval()
classifier_0.eval()
classifier_1.eval()
classifier_2.eval()
domain_classifier_0.eval()
domain_classifier_1.eval()
domain_classifier_2.eval()
filename_save = []
output_save = []
for index, (imgs, filenames) in enumerate(test_loader):
output_list = []
loss_mtl = []
imgs = Variable(imgs).to(device)
#labels = Variable(labels.view(-1)).to(device)
hidden = encoder(imgs)
tmp_0 = classifier_0(hidden)
tmp_1 = classifier_1(hidden)
tmp_2 = classifier_2(hidden)
output = moe_classifier(tmp_0, tmp_1, tmp_2)
preds = output.argmax(1).cpu()
#acc = np.mean((preds.detach().cpu() == labels.cpu()).numpy())
for filename in filenames:
filename_save.append(filename[-10:])
for out in preds.detach().cpu():
output_save.append(out)
#test_acc += acc
#if index % 500 == 0:
#print(acc)
# save csv
file = open("./submission/multi_01_rel.csv", "w")
file.write("image_name,label\n")
for i, (filename,
data) in enumerate(zip(filename_save, output_save)):
file.write("test/%s,%d\n" % (filename, int(data)))
return 0
def main():
batch_size = 2
batch_size_test = 2
feature_size = 2048 * 1 * 1
print("Check img size!!!!!!!!!!!!")
# load my Dataset
inf_csv_path = [
"./dataset_public/infograph/infograph_train.csv",
"./dataset_public/infograph/infograph_test.csv"
]
qdr_csv_path = [
"./dataset_public/quickdraw/quickdraw_train.csv",
"./dataset_public/quickdraw/quickdraw_test.csv"
]
skt_csv_path = [
"./dataset_public/sketch/sketch_train.csv",
"./dataset_public/sketch/sketch_test.csv"
]
rel_csv_path = ["./dataset_public/real/real_train.csv"]
test_path = "./dataset_public/test"
test_dataset = Dataset.Dataset(csv_path=qdr_csv_path[1], filename=True)
test_loader = DataLoader(test_dataset,
batch_size=batch_size_test,
shuffle=False,
num_workers=1)
print('the target dataset has %d size.' % (len(test_loader)))
print('the batch_size is %d' % (batch_size))
# Pre-train models
modules = list(models.resnet50(pretrained=True).children())[:-1]
encoder = nn.Sequential(*modules)
#encoder = model.Encoder()
classifier = model.Classifier(feature_size)
# GPU enable
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print('Device used:', device)
if torch.cuda.is_available():
encoder = encoder.to(device)
classifier = classifier.to(device)
#loading models
encoder_path = sys.argv[1]
classifier_path = sys.argv[2]
load_checkpoint(encoder_path, encoder)
load_checkpoint(classifier_path, classifier)
encoder.eval()
classifier.eval()
filename_save = []
output_save = []
total_acc = 0.
with torch.no_grad():
for index, (imgs, labels, filenames) in enumerate(test_loader):
print("\r%d/%d" % (index + 1, len(test_loader)), end="")
output_list = []
imgs = Variable(imgs).to(device)
labels = Variable(labels.view(-1)).to(device)
hidden = encoder(imgs)
output = classifier(hidden)
preds = output.argmax(1).cpu()
acc = np.mean((preds.detach().cpu() == labels.cpu()).numpy())
total_acc += acc
for filename in filenames:
filename_save.append(filename[-10:])
for out in preds.detach().cpu():
output_save.append(out)
total_acc = total_acc / len(test_loader)
print("Acc = ", total_acc)
print(len(filename_save))
print(len(output_save))
# save csv
file = open("./submission/pred_qdr.csv", "w")
file.write("image_name,label\n")
for i, (filename, data) in enumerate(zip(filename_save, output_save)):
file.write("test/%s,%d\n" % (filename, int(data)))
return 0
def main():
# parameters
learning_rate = 0.001
num_epochs = 30
batch_size = 30
batch_size_test = 30
feature_size = 2048 * 1 * 1
#learning_rate = float(sys.argv[1])
# create the save log file
print("Create the directory")
if not os.path.exists("./save"):
os.makedirs("./save")
if not os.path.exists("./logfile"):
os.makedirs("./logfile")
if not os.path.exists("./logfile/MTL"):
os.makedirs("./logfile/MTL")
# load my Dataset
inf_csv_path = [
"./dataset_public/infograph/infograph_train.csv",
"./dataset_public/infograph/infograph_test.csv"
]
qdr_csv_path = [
"./dataset_public/quickdraw/quickdraw_train.csv",
"./dataset_public/quickdraw/quickdraw_test.csv"
]
skt_csv_path = [
"./dataset_public/sketch/sketch_train.csv",
"./dataset_public/sketch/sketch_test.csv"
]
rel_csv_path = ["./dataset_public/real/real_train.csv"]
test_path = "./dataset_public/test"
inf_train_dataset = Dataset.Dataset(csv_path=inf_csv_path[0], argu=True)
inf_train_loader = DataLoader(inf_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
qdr_train_dataset = Dataset.Dataset(csv_path=qdr_csv_path[0], argu=True)
qdr_train_loader = DataLoader(qdr_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
skt_train_dataset = Dataset.Dataset(csv_path=skt_csv_path[0], argu=True)
skt_train_loader = DataLoader(skt_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
rel_train_dataset = Dataset.Dataset(csv_path=rel_csv_path[0],
sample=True,
argu=True)
rel_train_loader = DataLoader(rel_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
valid_dataset = Dataset.Valid_Dataset(csv_path=rel_csv_path[0],
sample=True)
valid_loader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
test_dataset = Dataset.Dataset(csv_path=test_path,
mode="test",
filename=True)
test_loader = DataLoader(test_dataset,
batch_size=batch_size_test,
shuffle=False,
num_workers=4)
print('the source dataset has %d size.' % (len(rel_train_dataset)))
print('the valid dataset has %d size.' % (len(valid_dataset)))
print('the target dataset has %d size.' % (len(test_loader)))
print('the batch_size is %d' % (batch_size))
# Pre-train models
modules = list(models.resnet152(pretrained=True).children())[:-1]
encoder = nn.Sequential(*modules)
#encoder = model.Encoder()
classifier = model.Classifier(feature_size)
domain_classifier_0 = model.Domain_classifier_0(feature_size,
number_of_domain)
domain_classifier_1 = model.Domain_classifier_1(feature_size,
number_of_domain)
domain_classifier_2 = model.Domain_classifier_2(feature_size,
number_of_domain)
# GPU enable
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print('Device used:', device)
if torch.cuda.is_available():
encoder = encoder.to(device)
classifier = classifier.to(device)
domain_classifier_0 = domain_classifier_0.to(device)
domain_classifier_1 = domain_classifier_1.to(device)
domain_classifier_2 = domain_classifier_2.to(device)
# setup optimizer
optimizer_encoder = optim.SGD(list(encoder.parameters()) +
list(classifier.parameters()),
lr=learning_rate,
momentum=0.9)
optimizer_classifier = optim.SGD(list(classifier.parameters()),
lr=learning_rate,
momentum=0.9)
optimizer_domain = optim.Adam([{
'params': domain_classifier_0.parameters()
}, {
'params': domain_classifier_1.parameters()
}, {
'params': domain_classifier_2.parameters()
}],
lr=learning_rate,
weight_decay=1e-4)
"""
optimizer_encoder = optim.Adam([{'params': encoder.parameters()},
{'params': classifier.parameters()}], weight_decay = 1e-4 , lr= learning_rate)
optimizer_classifier = optim.Adam(classifier.parameters(), weight_decay = 1e-4 , lr= learning_rate)
optimizer_domain = optim.Adam([{'params': domain_classifier_0.parameters()},
{'params': domain_classifier_1.parameters()},
{'params': domain_classifier_2.parameters()}], lr= learning_rate, weight_decay = 1e-4 )
"""
#loading models
#if input("Loading pre- file?(T/F)") == "T":
# encoder_path = sys.argv[2]
# classifier_path = sys.argv[3]
# load_checkpoint(encoder_path,encoder)
# load_checkpoint(classifier_path,classifier)
#Lossfunction
moe_criterion = nn.CrossEntropyLoss()
mtl_criterion = nn.NLLLoss()
D_loss_list = []
L_loss_list = []
sum_src_acc_list = []
sum_trg_acc_list = []
sum_label_acc_loist = []
sum_test_acc_list = []
print("Starting training...")
best_acc = 0.
valid_acc = 0.0
for epoch in range(num_epochs):
print("Epoch:", epoch + 1)
encoder.train()
classifier.train()
domain_classifier_0.train()
domain_classifier_1.train()
domain_classifier_2.train()
epoch_D_loss = 0.0
epoch_C_loss = 0.0
domain_acc_src = 0.0
domain_acc_trg = 0.0
sum_trg_acc = 0.0
sum_label_acc = 0.0
sum_test_acc = 0.0
tmp_valid_acc = 0.0
if (epoch + 1) in [5, 10, 15, 20, 25, 30, 35, 40, 45]:
for optimizer_t in optimizer_domain.param_groups:
optimizer_t['lr'] /= 1.7
optimizer_encoder.param_groups[0]['lr'] /= 1.7
optimizer_classifier.param_groups[0]['lr'] /= 1.7
train_loader = [
inf_train_loader, qdr_train_loader, skt_train_loader,
rel_train_loader
]
len_loader = min([
len(train_loader[0]),
len(train_loader[1]),
len(train_loader[2]),
len(train_loader[3])
])
sum_acc_1 = 0.
sum_acc_2 = 0.
sum_acc_3 = 0.
for index, (inf, qdr, skt, rel, test) in enumerate(
zip(train_loader[0], train_loader[1], train_loader[2],
train_loader[3], valid_loader)):
optimizer_classifier.zero_grad()
optimizer_encoder.zero_grad()
optimizer_domain.zero_grad()
# colculate the lambda_
p = (index + len_loader * epoch) / (len_loader * num_epochs)
lambda_ = 2.0 / (1. + np.exp(-10 * p)) - 1.0
s1_imgs, s1_labels = skt
s2_imgs, s2_labels = inf
s3_imgs, s3_labels = qdr
t1_imgs, _ = rel
from_s1_labels = Variable(
torch.LongTensor([0 for i in range(len(s1_imgs))])).to(device)
from_s2_labels = Variable(
torch.LongTensor([0 for i in range(len(s2_imgs))])).to(device)
from_s3_labels = Variable(
torch.LongTensor([0 for i in range(len(s3_imgs))])).to(device)
from_t1_labels = Variable(
torch.LongTensor([1 for i in range(len(t1_imgs))])).to(device)
s1_imgs = Variable(s1_imgs).to(device)
s1_labels = Variable(s1_labels.view(-1)).to(device)
s2_imgs = Variable(s2_imgs).to(device)
s2_labels = Variable(s2_labels.view(-1)).to(device)
s3_imgs = Variable(s3_imgs).to(device)
s3_labels = Variable(s3_labels.view(-1)).to(device)
t1_imgs = Variable(t1_imgs).to(device)
s1_feature = encoder(s1_imgs)
s2_feature = encoder(s2_imgs)
s3_feature = encoder(s3_imgs)
t1_feature = encoder(t1_imgs)
if len(s1_feature) < 2 or len(s2_feature) < 2 or len(
s3_feature) < 2 or len(t1_feature) < 2:
break
# Training Classifier network (loss_mtl)
s1_output = classifier(s1_feature)
s2_output = classifier(s2_feature)
s3_output = classifier(s3_feature)
s1_preds = s1_output.argmax(1).cpu()
s2_preds = s2_output.argmax(1).cpu()
s3_preds = s3_output.argmax(1).cpu()
s1_acc = np.mean(
(s1_preds.detach().cpu() == s1_labels.cpu()).numpy())
s2_acc = np.mean(
(s2_preds.detach().cpu() == s2_labels.cpu()).numpy())
s3_acc = np.mean(
(s3_preds.detach().cpu() == s3_labels.cpu()).numpy())
s1_c_loss = mtl_criterion(s1_output, s1_labels)
s2_c_loss = mtl_criterion(s2_output, s2_labels)
s3_c_loss = mtl_criterion(s3_output, s3_labels)
mtl_loss = 1 * (1 * s1_c_loss + 1. * s2_c_loss + 1. * s3_c_loss)
s_acc = s1_acc + s2_acc + s3_acc
# Domain_classifier network with source domain (loss_adv)
s1_domain_output = domain_classifier_0(s1_feature, lambda_)
s2_domain_output = domain_classifier_1(s2_feature, lambda_)
s3_domain_output = domain_classifier_2(s3_feature, lambda_)
#if index == 10:
# print(s1_domain_preds)
s1_domain_acc = np.mean(
(s1_domain_output.argmax(1).cpu() == from_s1_labels.cpu()
).numpy())
s2_domain_acc = np.mean(
(s2_domain_output.argmax(1).cpu() == from_s2_labels.cpu()
).numpy())
s3_domain_acc = np.mean(
(s3_domain_output.argmax(1).cpu() == from_s3_labels.cpu()
).numpy())
#print(s1_domain_output.shape)
#print(s1_domain_output[0])
s1_d_loss = moe_criterion(s1_domain_output, from_s1_labels)
s2_d_loss = moe_criterion(s2_domain_output, from_s2_labels)
s3_d_loss = moe_criterion(s3_domain_output, from_s3_labels)
#D_loss_src = 1 * s1_d_loss + s2_d_loss + 1* s3_d_loss
#print(D_loss_src.item())
# Domain_classifier network with target domain (loss_adv)
t1_domain_0_output = domain_classifier_0(t1_feature, lambda_)
t1_domain_1_output = domain_classifier_1(t1_feature, lambda_)
t1_domain_2_output = domain_classifier_2(t1_feature, lambda_)
t1_domain_0_acc = np.mean(
(t1_domain_0_output.argmax(1).cpu() == from_t1_labels.cpu()
).numpy())
t1_domain_1_acc = np.mean(
(t1_domain_1_output.argmax(1).cpu() == from_t1_labels.cpu()
).numpy())
t1_domain_2_acc = np.mean(
(t1_domain_2_output.argmax(1).cpu() == from_t1_labels.cpu()
).numpy())
D0_loss_trg = moe_criterion(t1_domain_0_output, from_t1_labels)
D1_loss_trg = moe_criterion(t1_domain_1_output, from_t1_labels)
D2_loss_trg = moe_criterion(t1_domain_2_output, from_t1_labels)
D_loss_trg = (1 * D0_loss_trg + D1_loss_trg +
1 * D2_loss_trg) / 2.9
D_s1t1_loss = s1_d_loss
D_s2t1_loss = s2_d_loss
D_s3t1_loss = s3_d_loss
if D_s1t1_loss > D_s2t1_loss and D_s1t1_loss > D_s3t1_loss:
adv_loss = D_s1t1_loss + D_loss_trg
if D_s2t1_loss > D_s1t1_loss and D_s2t1_loss > D_s3t1_loss:
adv_loss = D_s2t1_loss + D_loss_trg
if D_s3t1_loss > D_s1t1_loss and D_s3t1_loss > D_s2t1_loss:
adv_loss = D_s3t1_loss + D_loss_trg
#adv_loss = D_loss_src + D_loss_trg
loss = 1. * mtl_loss + adv_loss
epoch_D_loss += adv_loss.item()
epoch_C_loss += mtl_loss.item()
D_trg_acc = (t1_domain_0_acc + t1_domain_1_acc +
t1_domain_2_acc) / 3.
D_src_acc = (s1_domain_acc + s2_domain_acc + s3_domain_acc) / 3.
sum_acc_1 += s1_acc
sum_acc_2 += s2_acc
sum_acc_3 += s3_acc
domain_acc_src += D_src_acc
domain_acc_trg += D_trg_acc
#mtl_loss.backward()
#adv_loss.backward()
loss.backward()
optimizer_classifier.step()
optimizer_encoder.step()
optimizer_domain.step()
if (index + 1) % 10 == 0:
print(
'Epoch [%d/%d], Iter [%d/%d] C_loss %.4f , D_loss %.4f ,LR: %.5f'
% (epoch + 1, num_epochs, index + 1, len_loader,
epoch_C_loss / (index + 1), epoch_D_loss /
(index + 1), optimizer_domain.param_groups[0]['lr']))
print(
"====> Acc %.4f %.4f %.4f Domain Acc: %.4f %.4f ,Test Acc: %.4f"
% (sum_acc_1 / (index + 1), sum_acc_2 /
(index + 1), sum_acc_3 / (index + 1), domain_acc_src /
(index + 1), domain_acc_trg /
(index + 1), tmp_valid_acc / (index + 1)))
# Testing
encoder.eval()
classifier.eval()
with torch.no_grad():
test_imgs, test_labels = test
test_imgs = Variable(test_imgs).to(device)
test_labels = Variable(test_labels.view(-1)).to(device)
test_feature = encoder(test_imgs)
test_output = classifier(test_feature)
test_preds = test_output.argmax(1).cpu()
test_acc = np.mean(
(test_preds.detach().cpu() == test_labels.cpu()).numpy())
tmp_valid_acc += test_acc
encoder.train()
classifier.train()
s1_avg_acc = sum_acc_1 / len_loader
s2_avg_acc = sum_acc_2 / len_loader
s3_avg_acc = sum_acc_3 / len_loader
print('Validing: Acc %.4f ' % (tmp_valid_acc / len_loader))
print('Avg sur: Acc %.4f,%.4f,%.4f ' %
(s1_avg_acc, s2_avg_acc, s3_avg_acc))
if tmp_valid_acc / len_loader > best_acc:
best_acc = tmp_valid_acc / len_loader
print('Find best: Acc %.6f ' % (best_acc))
save_checkpoint(
'./save/encoder-%.4f-%.4f-%.4f.pth' %
(s1_avg_acc, s2_avg_acc, s3_avg_acc), encoder)
save_checkpoint(
'./save/domain-%.4f-%.4f.pth' %
(domain_acc_src / (index + 1), domain_acc_trg / (index + 1)),
domain_classifier_0)
save_checkpoint(
'./save/classifier-%.4f-%.4f-%.4f.pth' %
(s1_avg_acc, s2_avg_acc, s3_avg_acc), classifier)
print('Best: Acc %.6f Avg sur: Acc %.4f,%.4f,%.4f' %
(best_acc, s1_avg_acc, s2_avg_acc, s3_avg_acc))
if tmp_valid_acc / len_loader > 0.3:
best_acc = tmp_valid_acc / len_loader
print('Find best: Acc %.6f ' % (best_acc))
save_checkpoint('./save/encoder--%.4f.pth' % (best_acc), encoder)
save_checkpoint('./save/classifier--%.4f.pth' % (best_acc),
classifier)
test_acc = 0.
test_loss = 0.
encoder.eval()
classifier.eval()
domain_classifier_0.eval()
domain_classifier_1.eval()
domain_classifier_2.eval()
filename_save = []
output_save = []
with torch.no_grad():
for index, (imgs, filenames) in enumerate(test_loader):
output_list = []
loss_mtl = []
imgs = Variable(imgs).to(device)
#labels = Variable(labels.view(-1)).to(device)
hidden = encoder(imgs)
output = classifier(hidden)
preds = output.argmax(1).cpu()
#acc = np.mean((preds.detach().cpu() == labels.cpu()).numpy())
for filename in filenames:
filename_save.append(filename[-10:])
for out in preds.detach().cpu():
output_save.append(out)
#test_acc += acc
#if index % 500 == 0:
#print(acc)
# save csv
file = open("./submission/multi_01_rel.csv", "w")
file.write("image_name,label\n")
for i, (filename,
data) in enumerate(zip(filename_save, output_save)):
file.write("test/%s,%d\n" % (filename, int(data)))
return 0
def main():
# parameters
learning_rate = 0.001
num_epochs = 50
batch_size_train = 200
batch_size_test = 100
feature_size = 512 * 1 * 1
# load my Dataset
inf_csv_path = [
"./dataset_public/infograph/infograph_train.csv",
"./dataset_public/infograph/infograph_test.csv"
]
qdr_csv_path = [
"./dataset_public/quickdraw/quickdraw_train.csv",
"./dataset_public/quickdraw/quickdraw_test.csv"
]
skt_csv_path = [
"./dataset_public/sketch/sketch_train.csv",
"./dataset_public/sketch/sketch_test.csv"
]
rel_csv_path = ["./dataset_public/real/real_train.csv"]
test_path = "./dataset_public/test"
source = rel_csv_path[0]
test = rel_csv_path[0]
print("source from : ", source)
print("test from : ", test)
inf_train_dataset = Dataset.Dataset(csv_path=source, sample=True)
s_train_loader = DataLoader(inf_train_dataset,
batch_size=batch_size_train,
shuffle=True,
num_workers=1)
test_dataset = Dataset.Valid_Dataset(csv_path=test)
#test_dataset = Dataset.Dataset(csv_path = test)
test_loader = DataLoader(test_dataset,
batch_size=batch_size_test,
shuffle=True,
num_workers=1)
print('the source dataset has %d size.' % (len(inf_train_dataset)))
print('the target dataset has %d size.' % (len(test_dataset)))
print('the batch_size is %d' % (batch_size_train))
# models setting
modules = list(models.resnet18(pretrained=True).children())[:-1]
feature_extractor = nn.Sequential(*modules)
#feature_extractor = model.Encoder()
label_predictor = model.Classifier(feature_size)
# GPU enable
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print('Device used:', device)
if torch.cuda.is_available():
feature_extractor = feature_extractor.to(device)
label_predictor = label_predictor.to(device)
# setup optimizer
optimizer = optim.Adam([{
'params': label_predictor.parameters()
}, {
'params': feature_extractor.parameters()
}],
lr=learning_rate)
#Lossfunction
L_criterion = nn.NLLLoss()
print("Starting training...")
best_acc = 0.
for epoch in range(num_epochs):
feature_extractor.train()
label_predictor.train()
print("Epoch:", epoch + 1)
len_dataloader = len(s_train_loader)
epoch_L_loss = 0.0
if (epoch + 1) == 8:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 2
# F_optimizer.param_groups[0]['lr'] /= 2
# L_optimizer.param_groups[0]['lr'] /= 2
# D_optimizer.param_groups[0]['lr'] /= 2
if (epoch + 1) == 15:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 2
if (epoch + 1) == 20:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 2
if (epoch + 1) == 25:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 2
if (epoch + 1) == 30:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 2
# F_optimizer.param_groups[0]['lr'] /= 2
# L_optimizer.param_groups[0]['lr'] /= 2
# D_optimizer.param_groups[0]['lr'] /= 2
for i, source_data in enumerate(s_train_loader):
source_img, source_label = source_data
source_img = Variable(source_img).to(device)
source_label = Variable(source_label).to(device)
# train the feature_extractor
optimizer.zero_grad()
#optimizer_label.zero_grad()
source_feature = feature_extractor(source_img)
# Label_Predictor network
src_label_output = label_predictor(source_feature)
_, src_pred_arg = torch.max(src_label_output, 1)
src_acc = np.mean(
np.array(src_pred_arg.cpu()) == np.array(
source_label.view(-1).cpu()))
loss = L_criterion(src_label_output, source_label.view(-1))
epoch_L_loss += loss.item()
loss.backward()
#optimizer_label.step()
optimizer.step()
if (i % 20 == 0):
print(
'Epoch [%d/%d], Iter [%d/%d] loss %.4f , LR = %.6f , Acc = %.4f'
% (epoch + 1, num_epochs, i + 1, len_dataloader,
loss.item(), optimizer.param_groups[0]['lr'], src_acc))
# epoch done
print('-' * 80)
feature_extractor.eval()
label_predictor.eval()
total_acc = 0.
for i, test_data in enumerate(test_loader):
imgs, labels = test_data
imgs = Variable(imgs).to(device)
labels = Variable(labels).to(device)
feature = feature_extractor(imgs)
# Label_Predictor network
output = label_predictor(feature)
_, pred_arg = torch.max(output, 1)
acc = np.mean(
np.array(pred_arg.cpu()) == np.array(labels.view(-1).cpu()))
if i % 100 == 0:
print(acc)
total_acc += acc
total_acc = total_acc / len(test_loader)
if total_acc > best_acc:
best_acc = total_acc
print("Best accuracy : ", best_acc)
print("<" + "=" * 40 + ">")
print("Best accuracy : ", best_acc)
def main():
# parameters
learning_rate = 0.001
num_epochs = 100
batch_size = 50
batch_size_test = 50
feature_size = 2048 * 1 * 1
# create the save log file
print("Create the directory")
if not os.path.exists("./save"):
os.makedirs("./save")
if not os.path.exists("./logfile"):
os.makedirs("./logfile")
if not os.path.exists("./logfile/MTL"):
os.makedirs("./logfile/MTL")
# load my Dataset
inf_csv_path = ["./dataset_public/infograph/infograph_train.csv","./dataset_public/infograph/infograph_test.csv"]
qdr_csv_path = ["./dataset_public/quickdraw/quickdraw_train.csv","./dataset_public/quickdraw/quickdraw_test.csv"]
skt_csv_path = ["./dataset_public/sketch/sketch_train.csv","./dataset_public/sketch/sketch_test.csv"]
rel_csv_path = ["./dataset_public/real/real_train.csv"]
test_path = "./dataset_public/test"
inf_train_dataset = Dataset.Dataset(csv_path = inf_csv_path[0],argu = True)
inf_train_loader = DataLoader(inf_train_dataset ,batch_size = batch_size ,shuffle=True ,num_workers=1)
qdr_train_dataset = Dataset.Dataset(csv_path = qdr_csv_path[0],argu = True)
qdr_train_loader = DataLoader(qdr_train_dataset ,batch_size = batch_size ,shuffle=True ,num_workers=1)
skt_train_dataset = Dataset.Dataset(csv_path = skt_csv_path[0],argu = True)
skt_train_loader = DataLoader(skt_train_dataset ,batch_size = batch_size ,shuffle=True ,num_workers=1)
rel_train_dataset = Dataset.Dataset(csv_path = rel_csv_path[0],sample = True,argu = True)
rel_train_loader = DataLoader(rel_train_dataset ,batch_size = batch_size ,shuffle=True ,num_workers=1)
valid_dataset = Dataset.Valid_Dataset(csv_path = rel_csv_path[0],sample = True)
valid_loader = DataLoader(valid_dataset ,batch_size = batch_size ,shuffle=True ,num_workers=1)
test_dataset = Dataset.Dataset(csv_path = test_path,mode = "test",filename = True)
test_loader = DataLoader(test_dataset ,batch_size = batch_size_test ,shuffle=False ,num_workers=1)
print('the source dataset has %d size.' % (len(rel_train_dataset)))
print('the valid dataset has %d size.' % (len(valid_dataset)))
print('the target dataset has %d size.' % (len(test_dataset)))
print('the batch_size is %d' % (batch_size))
# Pre-train models
modules = list(models.resnet50(pretrained = True).children())[:-1]
encoder = nn.Sequential(*modules)
#encoder = model.Encoder()
classifier = model.Classifier(feature_size)
domain_classifier = model.Domain_classifier_0(feature_size,number_of_domain)
# GPU enable
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print('Device used:', device)
if torch.cuda.is_available():
encoder = encoder.to(device)
classifier = classifier.to(device)
domain_classifier = domain_classifier.to(device)
# setup optimizer
"""
optimizer_encoder = optim.SGD(
list(encoder.parameters()) +
list(classifier.parameters()),lr=learning_rate,momentum=0.9)
optimizer_classifier = optim.SGD(list(classifier.parameters()),lr=learning_rate,momentum=0.9)
optimizer = optim.SGD([{'params': domain_classifier.parameters()}], lr= learning_rate,momentum=0.9)
"""
optimizer_encoder = optim.Adam([{'params': encoder.parameters()},
{'params': classifier.parameters()}], weight_decay = 1e-4 , lr= learning_rate)
optimizer_classifier = optim.Adam(classifier.parameters(), weight_decay = 1e-4 , lr= learning_rate)
optimizer = optim.Adam([{'params': domain_classifier.parameters()}], lr= learning_rate, weight_decay = 1e-4 )
#Lossfunction
moe_criterion = nn.CrossEntropyLoss()
mtl_criterion = nn.NLLLoss()
D_loss_list = []
L_loss_list = []
sum_src_acc_list = []
sum_trg_acc_list = []
sum_label_acc_loist = []
sum_test_acc_list = []
print("Starting training...")
best_acc = 0.
valid_acc = 0.0
for epoch in range(num_epochs):
print("Epoch:", epoch+1)
encoder.train()
classifier.train()
domain_classifier.train()
epoch_D_loss = 0.0
epoch_C_loss = 0.0
sum_src_acc = 0.0
sum_trg_acc = 0.0
sum_label_acc = 0.0
sum_test_acc = 0.0
tmp_valid_acc = 0.0
if (epoch+1) in [5,10,15,20,25,30,35,40,45]:
for optimizer_t in optimizer.param_groups:
optimizer_t['lr'] /= 1.7
train_loader = [inf_train_loader,qdr_train_loader,skt_train_loader,rel_train_loader]
len_loader = min([len(train_loader[0]),len(train_loader[1]),len(train_loader[2]),len(train_loader[3])])
for index, (inf,qdr,skt,rel,test) in enumerate(zip(train_loader[0],train_loader[1],train_loader[2],train_loader[3],valid_loader)):
optimizer_encoder.zero_grad()
optimizer_classifier.zero_grad()
optimizer.zero_grad()
# colculate the lambda_
p = (index + len_loader * epoch)/(len_loader * num_epochs)
lambda_ = 2.0 / (1. + np.exp(-10 * p)) - 1.0
s1_imgs , s1_labels = skt
s2_imgs , s2_labels = inf
s3_imgs , s3_labels = inf
test_imgs , test_labels = test
t1_imgs , _ = rel
from_s2_labels = Variable(torch.LongTensor([0 for i in range(len(s3_imgs))])).to(device)
from_t1_labels = Variable(torch.LongTensor([1 for i in range(len(t1_imgs))])).to(device)
s1_imgs = Variable(s1_imgs).to(device) ; s1_labels = Variable(s1_labels.view(-1)).to(device)
s2_imgs = Variable(s2_imgs).to(device) ; s2_labels = Variable(s2_labels.view(-1)).to(device)
s3_imgs = Variable(s3_imgs).to(device) ; s3_labels = Variable(s3_labels.view(-1)).to(device)
test_imgs = Variable(test_imgs).to(device) ; test_labels = Variable(test_labels.view(-1)).to(device)
t1_imgs = Variable(t1_imgs).to(device)
s2_feature = encoder(s2_imgs)
test_feature = encoder(test_imgs)
t1_feature = encoder(t1_imgs)
test_output = classifier(test_feature)
test_preds = test_output.argmax(1).cpu()
test_acc = np.mean((test_preds.detach().cpu() == test_labels.cpu()).numpy())
s2_output = classifier(s2_feature)
s2_preds = s2_output.argmax(1).cpu()
s2_acc = np.mean((s2_preds.detach().cpu() == s2_labels.cpu()).numpy())
s2_c_loss = mtl_criterion(s2_output,s2_labels)
mtl_loss = s2_c_loss
# Domain_classifier network with source domain (loss_adv)
s2_domain_output = domain_classifier(s2_feature,lambda_)
s2_domain_acc = np.mean((s2_domain_output.argmax(1).cpu() <= 0.5).numpy())
s2_d_loss = moe_criterion(s2_domain_output,from_s2_labels)
D_loss_src = s2_d_loss
#print(D_loss_src.item())
# Domain_classifier network with target domain (loss_adv)
t1_domain_0_output = domain_classifier(t1_feature,lambda_)
t1_domain_0_acc = np.mean((t1_domain_0_output.argmax(1).cpu() > 0.5).numpy())
D0_loss_trg = moe_criterion(t1_domain_0_output,from_t1_labels)
D_loss_trg = D0_loss_trg
if (index+1) % 100 == 0:
print(s2_domain_output.argmax(1).cpu())
print(t1_domain_0_output.argmax(1).cpu())
adv_loss = D_loss_src + D_loss_trg
loss = 1 * mtl_loss + 1 * adv_loss
D_trg_acc = t1_domain_0_acc
D_src_acc = s2_domain_acc
#mtl_loss.backward()
#adv_loss.backward()
loss.backward()
optimizer.step()
optimizer_encoder.step()
optimizer_classifier.step()
if (index+1) % 10 == 0:
print('Epoch [%d/%d], Iter [%d/%d] C_loss : %.4f D_loss %.4f arc: %.4f trg: %.4f ,LR: %.5f'
%(epoch+1, num_epochs,index+1, len_loader,mtl_loss.item(), adv_loss.item(),D_loss_src.item(),D_loss_trg.item(),optimizer.param_groups[0]['lr']))
print("====> Domain Acc: %.4f %.4f Test: %.4f"%(D_src_acc,D_trg_acc,test_acc))
print('Validing: Acc %.4f ' %(tmp_valid_acc/len_loader))
if tmp_valid_acc/len_loader > best_acc :
best_acc = tmp_valid_acc/len_loader
print('Find best: Acc %.6f ' %(best_acc))
print('Best: Acc %.6f ' %(best_acc))
return 0