I’ve build a network for super-resolution which block of layers are stored in a nn.ModuleList() and trained incrementally. When training the last block, the output start to become uniform and after some iterations, all the previous trained layers do the same.
I tried to optimize the last block alone after training the other blocks and it worked, meaning that the last block in nn.ModuleList() isn’t optimized correctly.
(1) Does anyone had the same problem before?
(2) Is there a way to get the entire computational graph and the gradients in order to monitor efficiently all the variables trained by the optimizer at each time step?
Thank you in advance!