Open maxstrobel opened 4 years ago
push
I am also facing the same issue.
Hi, Adding a beta term constraint over the KL divergence loss solves this issue.
Loss = NLL + KL
Adding a beta value multiplied with KL will solve the convergence issue.
Loss = NLL + ß * KL
where ß is a hyper-parameter.
We will update the repo soon with a good way to set ß.
Do you have a rough estimate, when you will update the repo, e.g. days / weeks / months?
Thanks!
In a week.
could you explain the parameter of train_ens and val_ens? Does it mean sample_number?
Yes, that's correct.
In a week.
Hello, with the latest version of the code still replicating the validation accuracy of over 80 percent on the cifar10 as stated in the paper (figure 3) is not successful. Would you be able to guide me a bit toward the configuration and initialization of the parameters?
@kumar-shridhar How to reproduce the validation accuracy as stated in the paper on the cifar10 dataset? The network cannot break through 64% in validation accuracy when I use the same setting in your configuration file. Thank you.
Hi,
thanks for this nice work, I really appreciate it! I tried to replicate the results from your paper with the repository, but I have not succeeded.
First, I downloaded your repo and the datasets. Then I adapted the configuration for the Bayesian Networks:
Finally I run the evaluation script with
main_bayesian.py --net_type alexnet --dataset CIFAR10
, but the network is not able to overcome a validation accuracy of around 58%:Can you explain, how to replicate the results from the paper?