wjmaddox / drbayes Goto Github PK

Hello!

I was trying to run the visualisation jupyter notebook and had a couple of issues (I installed the repo as a package as mentioned):

• In setup.py one of the requirements is scikit_learn>=0.20.2. However , currently version 0.24 gets installed which actually throws some errors (I think some things have been refactored in the newest versions). I had to downgrade to 0.20.2 to get it to work, so I believe this should be changed to scikit_learn==0.20.2 (I actually just noticed that this is specified correctly in requirements.txt, just not in setup.py)
• It would maybe be helpful if the correct Python version needed was specified? setup.py says "3.6" but 3.6.0 and 3.6.1 don't actually work (sorry I can't remember exactly now what the issues were, but I ended up using 3.6.9 to get it to work)
• Additionally, things that I had to install manually, which were not included in setup.py: seaborn and pyro-ppl

And one more thing, in the curve subspaces section of the notebook it says "Note that for this to work you need to add the repo https://github.com/timgaripov/dnn-mode-connectivity to your Python path". However, I think the version of the curves.py file in that repo might not be the same as that used in the notebook. In fact, the notebook calls

model = curves.CurveNet(curve, architecture, 3, fix_end=True, fix_start=True, 
                        architecture_kwargs=model_cfg.kwargs)

but in the original repo that function is written as:

class CurveNet(Module):
    def __init__(self, num_classes, curve, architecture, num_bends, fix_start=True, fix_end=True,
                 architecture_kwargs={}):

so I believe the num_classes argument is missing in the function call in the notebook.

Thanks!

Hello,

The approach and results presented in the paper "Subspace Inference for Bayesian Deep Learning" impress me a lot and I would like to reproduce the experiments, specifically the UCI experiments, using this repository.

I followed the instructions and process described in the README and managed to get things run. However, I failed to obtain similar numbers for certain datasets.

To be more precise:

• versions of libraries match.

• the datasets are downloaded from Google Drive as suggested in the README under drbayes/experiments/uci_exps.

• all experiments on small UCI datasets are launched by the script run_ucismall.sh under drbayes/experiments/uci_exps/bayesian_benchmarks/tasks.

• the reproduced unnormalized test likelihoods for the yacht dataset are significantly different from those reported in the paper as shown below.

I would be extremely grateful if anyone could suggest what I could've done wrong in my reproduction experiments.

Thank you in advance!

Hi,

In subspaces.py you use two different SVD algorithms: TruncatedSVD which seems to be unused and then later randomized_svd. Which one is the correct one?

Regards,
Maciej

Hello @wjmaddox,

Thanks for sharing your code. I found your paper very inspiring.

I wonder if the sampling of SWAG is correct. In particular, it appears that the standard Gaussian sample is multiplied by the vector of variances for the SWAG-Diagonal part:

Whereas in the original SWAG paper and its implementation the standard Gaussian is multiplied by the vector of standard deviations to sample from N(0, diag(variances)):

https://github.com/wjmaddox/swa_gaussian/blob/b172d93278fdb92522c8fccb7c6bfdd6f710e4f0/swag/posteriors/swag.py#L121

Am I missing something? Thanks for your time.