def main():
batch_size = 32
test_batch_size = 32
p = os.path.expanduser("./image-clef/p")
c = os.path.expanduser("./image-clef/c")
i = os.path.expanduser("./image-clef/i")
epochs = 200
lr = 0.0001
device = torch.device("cuda")
train_loader_source = DA_datasets.imageclef_train_loader(p, batch_size, 0)
train_loader_target = DA_datasets.imageclef_train_loader(c, batch_size, 0)
testloader_1_target = DA_datasets.imageclef_test_loader(
c, test_batch_size, 0)
model_dan = DAN_model.DANNet_ResNet(ResNet.resnet50, True).to(device)
#model_dan = DAN_model.DANNetVGG16(models.vgg16, True).to(device)
optimizer = torch.optim.SGD(model_dan.parameters(),
momentum=0.9,
lr=lr,
weight_decay=5e-4)
dan_train(epochs,
lr,
model_dan,
train_loader_source,
device,
train_loader_target,
testloader_1_target,
optimizer,
scheduler=None,
is_debug=False)
def main():
batch_size = 32
test_batch_size = 32
lr = 0.0005
momentum = 0.9
epochs = 200
epoch_step = 30
weight_decay = 5e-4
teacher_pretrained_path = "./da_resnet50_p_i.pth"
student_pretrained = False
device = torch.device("cuda")
p = os.path.expanduser("./image-clef/p")
c = os.path.expanduser("./image-clef/c")
i = os.path.expanduser("./image-clef/i")
train_loader_source = DA_datasets.imageclef_train_loader(p, batch_size, 0)
train_loader_target = DA_datasets.imageclef_train_loader(i, batch_size, 0)
testloader_target = DA_datasets.imageclef_test_loader(
i, test_batch_size, 0)
teacher_model = DAN_model.DANNet_ResNet(ResNet.resnet50, False)
student_model = DAN_model.DANNet_ResNet(ResNet.resnet34,
student_pretrained)
teacher_model = teacher_model.to(device)
student_model = student_model.to(device)
if teacher_pretrained_path != "":
teacher_model.load_state_dict(
torch.load(teacher_pretrained_path)['student_model'])
if torch.cuda.device_count() > 1:
teacher_model = torch.nn.DataParallel(teacher_model).to(device)
student_model = torch.nn.DataParallel(student_model).to(device)
distiller_model = distiller.Distiller(teacher_model,
student_model).to(device)
if torch.cuda.device_count() > 1:
distiller_model = torch.nn.DataParallel(distiller_model).to(device)
optimizer = torch.optim.SGD(list(distiller_model.s_net.parameters()) +
list(distiller_model.Connectors.parameters()),
lr,
momentum=momentum,
weight_decay=weight_decay)
od_kd_without_label(epochs, distiller_model, optimizer,
train_loader_target, testloader_target, device)
def train_normal():
batch_size = 32
test_batch_size = 32
lr = 0.001
momentum = 0.9
epochs = 200
weight_decay = 5e-4
device = torch.device("cuda")
p = os.path.expanduser("./image-clef/p")
c = os.path.expanduser("./image-clef/c")
i = os.path.expanduser("./image-clef/i")
train_dataset = DA_datasets.imageclef_train_loader(p, batch_size, 0)
test_dataset = DA_datasets.imageclef_test_loader(p, test_batch_size, 0)
model = DAN_model.DANNet_ResNet(ResNet.resnet50, True)
model = model.to(device)
optimizer = torch.optim.SGD(model.parameters(),
lr,
momentum=momentum,
weight_decay=weight_decay)
criterion = nn.CrossEntropyLoss()
best_acc = 0.0
for epoch in range(epochs):
running_loss = 0.0
for i, data in enumerate(dataset, 0):
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
acc = eval(model, device, test_dataset)
if (acc > best_acc):
best_acc = acc
torch.save(model, 'teacher_model_p.pth')
print(f'epoch : {epoch}, acc : {acc}')