search

1Konny / Beta-VAE

Pytorch implementation of β-VAE
MIT License
525 stars 125 forks source link

Tensor dimension bug in visualization code #1

Closed tonymetger closed 6 years ago

tonymetger commented 6 years ago

Hi I just started using your code to play around with beta-VAE and it's great. Unfortunately, there is one bug when I try to turn on visualization:

Traceback (most recent call last): File "main.py", line 65, in main(args) File "main.py", line 24, in main net.train() File "/home/tony/Github/Beta-VAE/solver.py", line 178, in train self.viz_lines() File "/home/tony/Github/Beta-VAE/solver.py", line 230, in viz_lines klds = torch.cat([dim_wise_klds, mean_klds, total_klds], 1).cpu() RuntimeError: invalid argument 0: Tensors must have same number of dimensions: got 2 and 1 at /pytorch/aten/src/TH/generic/THTensorMath.c:3577

Steps to reproduce: python -m visdom.server sh run_dsprites_B_gamma100_z10.sh

I guess this is quite easy to fix and it doesn't appear in an earlier version (commit 66fcd41996d50a3647d952ad7a923615cf081ecd), but I'm very new to Pytorch, so it's hard for me to find a fix.

Thanks again for the code and I hope this can be fixed easily, Tony

P.S.: Would it be possible for you to add a license so that we can use your code to benchmark other models in our research (which might get published at some point)? That would be great!

1Konny commented 6 years ago

Hi. sorry for the late reply, I just checked this issue.

I found that this issue occurs on the latest version of pytorch(0.4), not on the previous versions(0.3.1, 0.3.0, 0.2.0). anyway the bug has been fixed and now the codes will work.(though some warnings may still appear on the latest version 0.4.0 but it seems not a big problem.)

for the license, thank you for noticing me, and I will update regarding this soon.

if you have any questions and find any issues in the near future, feel free to make an issue or contact me via email. thanks.

tonymetger commented 6 years ago

Hi. Thanks for the fix, it works well now. Odd that the pytorch update caused this...

There is one other bug I came across, which occurs when one sets "--viz-on False" in the parameters. Running the same sh script as above with that option gives me:

Traceback (most recent call last): File "main.py", line 65, in main(args) File "main.py", line 24, in main net.train() File "/home/tony/Github/Beta-VAE/solver.py", line 176, in train self.viz_reconstruction() File "/home/tony/Github/Beta-VAE/solver.py", line 212, in viz_reconstruction self.viz.images(images, env=self.viz_name+'_reconstruction', AttributeError: 'Solver' object has no attribute 'viz'

I tried fixing it myself (see pull request), I would be happy if you could have a look at it and merge it if you like it.

Thanks Tony

1Konny commented 6 years ago

Thank you for your interest and contribution to my humble(?) codes!

I merged your PR after modifying some minor things(e.g. maybe printing and checking the values of "C" on the terminal is necessary because it represents the current capacity of the bottlenecks, though it can also be checked in visdom.)

Your pull request is very helpful and much appreciated.

Thanks Konny