search

facebookresearch / NeuralCompression

A collection of tools for neural compression enthusiasts.
MIT License
496 stars 43 forks source link

Can't reproduce the result of torch_vct #209

Closed cccleon closed 1 year ago

cccleon commented 1 year ago

Thanks for your excellent implementation!

I tried to train VCT on 2 RTX 4090, a total of 100k step used. Initially I follow the training config mentioned in README but fail to reproduce.Then I adjust the learning rate and distortion lambda for several groups, PSNR/bpp is finally 35/0.11

Would you share the checkpoint?

mmuckley commented 1 year ago

Hello @cccleon, I'm going to CC @desi-ivanova for running the code while I check in on the status of releasing the checkpoint for you.

desi-ivanova commented 1 year ago

Hi @cccleon -- I don't think 2x4090s and 100K steps will get you there, unfortunately. The checkpoint @mmuckley is referring to was trained for 400K steps on 32 V100 (32GB) cards (as described in the readme).

cccleon commented 1 year ago

@desi-ivanova @mmuckley Thanks for your reply! Looks like I should lower my learning rate and increase my total steps.I'll keep trying~