emtiyaz / vadam Goto Github PK

According to Yarin Gal's repo DropoutUncertaintyExps, there was test set contamination due to the bayesian optmisation of hypers:

Update (2018) We replaced the Bayesian optimisation implementation (which was used to find hypers) with a grid-search over the hypers. This is following feedback from @capybaralet who spotted test-set contamination (some train-set points, used to tune hypers which were shared across all splits, were used as test-set points in later splits). The new implementation iterates over the 20 splits, and for each train-test split it creates a new train-val split to tune hypers. These hypers are discarded between different train-test splits.

Yet, it seems the UCI regression experiments here use the same splits as in the DropoutUncertaintyExps repo. Was the bug fixed or is it still there and the reason for doing so here was the reviewers' request?

This should work for any model, right? Not just for RL? I'm a bit confused for why the ActorNetworkEpsilon in the Tensorflow implementation contains code to optimise the policy network.

I'm trying to use it for StyleGAN-based image reconstruction, i.e. find a distribution over low-dimensional latents z such that G(z) match a given input image. The network is already trained and fixed, it's just the latent that I want to optimise.

Also, why is the design such that the critic network is also exposed to the outside DDPG.py? That is also something RL-specific, right?

Thanks,
Raz

In your UCI experiments, you set your objective function to a log loss (and likewise the BO objective). Surely one wants to perform VI with VADAM and should optimize an ELBO?

See also:

Make sure that the same mini-batch should be used here.
One way to achieve that is to use state['step'] as the seed of the random number generator for generating mini-batches. I did that in my Autograd implementation