search

autonomousvision / occupancy_networks

This repository contains the code for the paper "Occupancy Networks - Learning 3D Reconstruction in Function Space"
https://avg.is.tuebingen.mpg.de/publications/occupancy-networks
MIT License
1.54k stars 294 forks source link

influence of diffrent normalization #62

Open hua-wu-que opened 4 years ago

hua-wu-que commented 4 years ago

Hi there,

I want to ask your advice regarding the normalization layers. I am implementing a network similar to yours, but I found if I use batch norm layer, during testing time model.eval() will give much worse results. Did you encounter this before? Thank you for your reply!

AlexsaseXie commented 4 years ago

I did several experiments about BN layers for OccNet. With BN layers, the training will be much stable and the final reconstruction results will be more 'average' and generally well. Without BN layers you cannot easily train the network(the loss fluctuated) and final results are statistically worse.

hua-wu-que commented 4 years ago

I did several experiments about BN layers for OccNet. With BN layers, the training will be much stable and the final reconstruction results will be more 'average' and generally well. Without BN layers you cannot easily train the network(the loss fluctuated) and final results are statistically worse.

Hi AlexsaseXie, did you call model.eval() before doing inference ? I get very bad results if I call model.eval()