search

HolyBayes / pytorch_ard

Pytorch implementation of Variational Dropout Sparsifies Deep Neural Networks
MIT License
83 stars 16 forks source link

mnist_ard loss is nan #5

Closed N2606 closed 4 years ago

N2606 commented 4 years ago

Hi, thanks for the works. I ran the mnist_ard and the loss is always nan regardless how I initialize log_sigma2. Did anyone train successfully?

HolyBayes commented 4 years ago

Hi! Thanks for your comment. Fixed

lumosity4tpj commented 4 years ago

@N2606 how to slove the initialize problem? When I change the initial value of log_sigma2, the loss can become non-Nan. However, loss is almost constant, resulting in the failure of model convergence.