Results about atlasnet HOT 12 CLOSED

hi ! indeed this is not right. can you include your training and testing loss plots ?

from atlasnet.

Here is the visdom visualization

from atlasnet.

I just saw the issue about the pytorch version. Maybe that is my problem. I am using 0.3.1 version. I will try 0.4.0 and get back to you. Thanks.

from atlasnet.

I am actually using python 3.6. Judging by the table you provided, the Chamfer loss doesn't work with 3.6 and pytorch 0.4, is that correct?

from atlasnet.

definitly the pytorch version. your curves should look as in the readme. good luck !

from atlasnet.

apparently not (i don'r remember it exactly). it boils down to the chamfer distance code which is not compatible to every pytorch version. why not install a conda env with python 2.7?

from atlasnet.

I tried conda python 2.7 and pytorch 1.12 and training was stable. Thanks!

from atlasnet.

@ThibaultGROUEIX It seems that removing the batch normalization layers solves the instability issues during training due to pytorch versions. I was able to train the model without BN in python 3.6, pytorch 0.3.1 and achieve 1.81 loss for the autoencoder after 200 epochs.

from atlasnet.

unfortunatly, i can't provide support if you change the architecture but thanks for linking Chamfer error to batch norm!
regards, Thibault

from atlasnet.

actually i am not clear on if 1.81 is an issue or a success in your case. what error are you reporting ?

from atlasnet.

It’s not an error. I’m just pointing out a potential workaround for those who may want to use the architecture without reverting back to an old pytorch version.

from atlasnet.

Great ! Thanks for clarifying !

from atlasnet.