Cannot reproduce experiment result

kumar-shridhar / PyTorch-BayesianCNN

Bayesian Convolutional Neural Network with Variational Inference based on Bayes by Backprop in PyTorch.

MIT License

1.42k stars 323 forks source link

Cannot reproduce experiment result #35

Closed Cliu2 closed 4 years ago

Cliu2 commented 4 years ago

Hi,

I tried your demo code at

Image Recognition/Bayesian_CNN_Detailed.ipynb

But I cannot reproduce your training results, the accuracy stays around 10% and it seems the model is learning nothing.

Would you please advise what could be the problem? Thanks

kumar-shridhar commented 4 years ago

Hey @Cliu2 ,

What dataset did you use?

Cliu2 commented 4 years ago

@kumar-shridhar Thank you for your reply.

I used the CIFAR10 dataset retrived remotely, following your code in the jupyter notebook. train_data = datasets.CIFAR10('data', train=True, download=True, transform=transform)

I simply copy all the codes in the notebook to a python script file and run it in terminal. So all the settings should be same.

arcini commented 4 years ago

I am experiencing the exact same issue. It seems that the network is unable to improve on its prediction accuracy due to some bug. I am going to try to look into it as much as I can, but @kumar-shridhar you should check this out since others have posted similar issues.

kumar-shridhar commented 4 years ago

Hi, Please check with the current version. Things should work fine.