optimizer = torch.optim.SGD(net.parameters(),lr = Flags.lr, momentum=0.9, nesterov=True)
sensitivity_list = []
loss_list = []
epoch_valid = 0
for epoch in range(Flags.nepochs):
loss_val = 0
print("\n ...Train at epoch " +str(epoch))
cont_iter = 0
for batch_id, (img1, img2, label) in tqdm(enumerate(trainLoader, 1)):
net.train()
if Flags.cuda:
img1, img2, label = Variable(img1.cuda()), Variable(img2.cuda()), Variable(label.cuda())
else:
img1, img2, label = Variable(img1), Variable(img2), Variable(label)
optimizer.zero_grad()
output = net.forward(img1, img2)
loss = loss_MSE (output, label)
loss_val += loss.item ()
optimizer.zero_grad()
loss.backward()
optimizer.step()
cont_iter += 1
loss_epoch = loss_val/cont_iter
loss_list.append(loss_epoch)
plot_loss(loss_list,save_path)
if epoch % Flags.valid_every == 0:
net.eval()
print("\n ...Valid")
sensitivity_valid = []
for _, (valid1, valid2, label_valid) in tqdm(enumerate(validLoader, 1)):
if Flags.cuda:
net.cuda()
optimizer = torch.optim.Adam(net.parameters(),lr = Flags.lr )
optimizer.zero_grad()
loss_val = 0
valid_list = []
for batch_id, (img1, img2, label) in tqdm(enumerate(trainLoader, 1)):
print("\n ...Train at epoch " +str(batch_id))
net.train()
if batch_id > Flags.max_iter:
break
if Flags.cuda:
img1, img2, label = Variable(img1.cuda()), Variable(img2.cuda()), Variable(label.cuda())
else:
img1, img2, label = Variable(img1), Variable(img2), Variable(label)
optimizer.zero_grad()
output = net.forward(img1, img2)
loss = loss_BCE (output, label)
loss_val += loss.item ()
loss.backward()
optimizer.step()
if batch_id % Flags.valid_every == 0:
net.eval()
list_err = []
print("\n ...Valid")
r, e = 0, 0
for _, (valid1, valid2, label_valid) in tqdm(enumerate(validLoader, 1)):
if Flags.cuda:
test1, test2 = valid1.cuda(), valid2.cuda()
else:
test1, test2 = Variable(valid1), Variable(valid2)
output = net.forward(test1, test2).data.cpu().numpy()