###### Training ######
with open('./output_' + opt.output_str + '/loss.csv', 'w',
newline='') as csvfile:
writer = csv.writer(csvfile, delimiter=' ')
writer.writerow([
'D_real_loss', 'D_fake_loss', 'G_loss', 'recon_loss', 'l2', 'l1',
'preceptual'
])
for epoch in range(opt.n_epochs):
print("epoch = " + str(epoch))
model.train(mode=True)
discriminator.train(mode=True)
G_loss_e = 0
D_real_loss_e = 0
D_fake_loss_e = 0
recon_loss_e = 0
seq_l2_loss_e = 0
seq_l1_loss_e = 0
seq_preceptual_e = 0
for i, batch in enumerate(trainloader):
curr_image, curr_depth = batch['image'], batch['depth']
next_image, next_depth = batch['next_image'], batch['next_depth']
label = torch.FloatTensor(opt.batchSize)
real_A = Variable(real_A)
real_B = Variable(real_B)
label = Variable(label)
if (opt.cuda):
real_A = real_A.cuda()
real_B = real_B.cuda()
label = label.cuda()
real_label = 1
fake_label = 0
########### Training ###########
netD.train()
netG.train()
for epoch in range(1, opt.niter + 1):
nowtime = time.time()
for i, image in enumerate(train_loader):
########### fDx ###########
netD.zero_grad()
imgA = image[0]
imgB = image[1]
# train with real data
real_A.data.copy_(imgA)
real_B.data.copy_(imgB)
real_AB = torch.cat((real_A, real_B), 1)
output = netD(real_AB)
########### Testing ###########
def test(niter):
loaderA, loaderB = iter(loader_A), iter(loader_B)
imgA = loaderA.next()
imgB = loaderB.next()
real_A.data.resize_(imgA.size()).copy_(imgA)
real_B.data.resize_(imgB.size()).copy_(imgB)
AB = G_AB(real_A)
BA = G_BA(real_B)
vutils.save_image(AB.data, 'AB_niter_%03d.png' % (niter), normalize=True)
vutils.save_image(BA.data, 'BA_niter_%03d.png' % (niter), normalize=True)
########### Training ###########
D_A.train()
D_B.train()
G_AB.train()
G_BA.train()
for iteration in range(1, opt.niter + 1):
########### data ###########
try:
imgA = loaderA.next()
imgB = loaderB.next()
except StopIteration:
loaderA, loaderB = iter(loader_A), iter(loader_B)
imgA = loaderA.next()
imgB = loaderB.next()
real_A.data.resize_(imgA.size()).copy_(imgA)
real_B.data.resize_(imgB.size()).copy_(imgB)