.eval mode predicts [2, -2, 2] about 3d-coded

[ThibaultGROUEIX]

hum that's strange indeed. does the eval script work if you set the network in train mode ? ( and keep batch size of 32 ). that's a good starting point... i never had this issue.

[sonnyhu]

Yeah. train mode + 32 batch size works. Based on discussion here (pytorch/pytorch#4741), it seems training is unstable. Not sure what to do... I am going to close issue. If you have some suggestions, definitely let me know. Thanks!

[sonnyhu]

Hi again,

I just realized that I was using 0.4.1 pytorch.

For what you said here, "Please note that the Chamfer Distance code doesn't work on all versions of pytorch because of some weird error with the batch norm layers.", does chamfer distance use batch norm or is it in encoder-decoder?

[ThibaultGROUEIX]

hi, yes i strongly recomment using the pytorch 0.4 specific commit mentionned in the readme. i am currently porting atlasnet to pytorch 1. i will do 3d-coded after...

[sonnyhu]

I tried the commit you listed in readme. Same behavior. Val loss exploded.

[sonnyhu]

What should opt.nb_primitives be? There is no default value for it. https://github.com/ThibaultGROUEIX/3D-CODED/blob/master/training/train_sup.py#L142

I am looking for any possible way I am training it wrong. Thanks!

[ThibaultGROUEIX]

should be 1. but i think these variables are never used anyway.