feats = feats[::-1]
msk = deconv(feats)
msk = functional.upsample(msk, scale_factor=4)
prior = functional.sigmoid(msk)
loss += criterion(msk, lbl)
deconv.zero_grad()
feature.zero_grad()
loss.backward()
optimizer_feature.step()
optimizer_deconv.step()
# visulize
image = make_image_grid(inputs.data[:, :3], mean, std)
writer.add_image('Image', torchvision.utils.make_grid(image), ib)
msk = functional.sigmoid(msk)
mask1 = msk.data # mskdata,分割出来的。
mask1 = mask1.repeat(1, 3, 1, 1)
acc = math.e**(0 - loss)
writer.add_image('Image2', torchvision.utils.make_grid(mask1), ib)
print('loss: %.4f, acc %.4f, (epoch: %d, step: %d)' %
(loss.data[0], acc, it, ib))
writer.add_scalar('M_global', loss.data[0], istep)
writer.add_scalar('acc', acc.data[0], istep)
istep += 1
del inputs, msk, lbl, loss, feats, mask1, image, acc
gc.collect()
if ib % 30 == 0:
optimizer_feat = torch.optim.Adam(res101.parameters(), lr=1e-4)
for t in range(10):
for i, (img, label) in enumerate(loader):
img = img.cuda()
label = label[0].cuda()
label = Variable(label)
input = Variable(img)
feats = res101(input)
output = seg(feats)
seg.zero_grad()
res101.zero_grad()
loss = criterion(output, label)
loss.backward()
optimizer_feat.step()
optimizer_seg.step()
## see
input = make_image_grid(img, mean, std)
label = make_label_grid(label.data)
label = Colorize()(label).type(torch.FloatTensor)
output = make_label_grid(torch.max(output, dim=1)[1].data)
output = Colorize()(output).type(torch.FloatTensor)
writer.add_image('image', input, i)
writer.add_image('label', label, i)
writer.add_image('pred', output, i)
writer.add_scalar('loss', loss.data[0], i)
print "epoch %d step %d, loss=%.4f" % (t, i, loss.data.cpu()[0])