ruqizhang / csgmcmc Goto Github PK

Hello! I'm having a bit of trouble reproducing the results for cyclic HMC with your code. Could you please help me with that?
I'm seeing parser.add_argument('--alpha', type=int, default=1, help='1: SGLD; <1: SGHMC') but shouldn't alpha here be float in [0,1]?
Also, if I allow it to be float and test the training on some value of alpha which seems reasonable (e.g. alpha = 0.05 resulting in momentum term = 0.95 which should be OK?), network weights seem to go to nan soon.

Hi! cifar100_ensemble.py does the ensembling in logit space while the textbook and commonsensical algorithm is to do the ensembling in prediction space. Is it OK?

I have some questions on the implementation of csghmc and how to test uncertainy
1 For csghmc, does the method in "Stochastic Gradient Hamiltonian Monte Carlo" is used at the sampling stage. I would like to know how do you implement it or corresponding function in the code.
2 How can i reproduce uncertainty estimation result shown in the paper.

Hi,

Thanks for an interesting paper and open-source code!

I'm struggling somewhat to match your SGLD/SGHMC implementation with the update equations in the paper. In https://github.com/ruqizhang/csgmcmc/blob/51e511478d607b2523fc803d82a26edd39b14b6d/experiments/cifar_csgmcmc.py (line 79), you have noise_std = 2*lr*alpha**0.5. We will take the gradient of this noise term and then multiply the result with lr, right? Will that actually match the update equation for SGLD then?

Here's how I would implement SGLD (with p(theta) = N(0, alpha*I)), N=datasize):

loss_fn = nn.CrossEntropyLoss()

loss_likelihood = loss_fn(logits, y)

loss_prior = 0.0
for param in network.parameters():
    loss_prior += (1.0/2.0)*(1.0/N)*(1.0/alpha)*torch.sum(torch.pow(param, 2))

loss_noise = 0.0
for param in network.parameters():
    loss_noise += (1.0/math.sqrt(N))*math.sqrt(2.0/lr)*torch.sum(param*Variable(torch.normal(torch.zeros(param.size()), std=1.0).cuda()))

loss = loss_likelihood + loss_prior + loss_noise

Hope you can help resolve my confusion.

Regards

//
Fredrik

Hello, Ruiqi Zhang,

I have read part of your paper and code, and it is a really interesting work! At the top of Page 3, in the second paragraph, when you wrote "use a minibatch to approximate U(\theta)", should the scale coefficient N'/N be N/N'? Since we are using the minibatch to approximate the full dataset.

Another question is in the code file "cifar_csghmc.py": the line 78, you wrote "d_p.add_(weight_decay, p.data)". My naive understanding is that you are computing the potential energy term, involving both the log-prior (from a normal distribution) and the log-likelihood. If that's the case, we seem to miss the coefficient "N" mentioned above...

I'm looking forward to your reply! Thank you very much!