updateOutput called multiple times

[geert56]

Creating a graph did not work for me. I get errors. It turns out that apply_func in graphgen.lua runs the original updateOutput again (after the first forward() call) via base_func. This goes wrong for any layer that somehow creates a new tensor because then the new tensor pointer cannot be found in the nodes table. A simple fix is to do just "return self.output" in those cases.

[fmassa]

Hi, thanks for opening this issue.

The fix you propose is included in this pending PR https://github.com/fmassa/optimize-net/pull/10, but only that is not enough to handle all cases (like in nn.Parallel). I haven't merged that PR yet because I'm afraid that the other changes I did there (like overwriting __index) could have unexpected effects.

I can factorize the bit that removes the need of running forward twice and merge it to master if it solves your problem.

But usually, it's a good practice to keep the output tensors of a module unchanged during different forward calls, and only changing its data.

[geert56]

I understand that it is not easy to piggy-back on existing Torch routines and be able to handle the diverse suite of "layers" all correctly. I am glad to see my problem confirmed. I have my own work-around and just hope that other can benefit as well. Keep up the good work.