search

AntixK / PyTorch-VAE

A Collection of Variational Autoencoders (VAE) in PyTorch.
Apache License 2.0
6.44k stars 1.05k forks source link

Problem about KL loss when training CVAE model. #92

Open dayunyan opened 4 months ago

dayunyan commented 4 months ago

When I trained CVAE on my own dataset, I found that the results were very poor. Upon inspection, I found that the KL loss is too large at training and validation time, reaching around 10^27, and will not decrease. A closer look shows that the encoder output log_var is 50~60. I have been adjusting for a long time without effect. How to solve this situation?

sunny12345-bit commented 3 months ago

Have you solved this problem？

dayunyan commented 3 months ago

Have you solved this problem？

Not yet, cause I have switched to a different technical approach. The original technology might not be suitable for my dataset, and moreover, VAE itself tend to be less stable during training compared to other models.