Jordy Van Landeghem
Jordy Van Landeghem
Your question is rather unspecified ... I believe you should look up the **bias-variance tradeoff**, a standard concept in Machine Learning. There you will get more feeling for how to...
As I read it, they report the mean, std, and max over 25 seeds (=runs). So they make the claim that lower std over seeds == more stability. Additionally, reporting...
Looking forward to that recording!
@dustinvtran @znado Is there any concrete plan to make SG-MCMC and SG-HMC implementations in TensorFlow? I have only found some implementations in Pytorch, but they might help in conversion: https://github.com/reml-lab/URSABench/blob/master/URSABench/inference/sghmc.py...
> Thank you for the quick reply! > > > 1. After some experimentation, we replaced the Laplace-approximated posterior variance with that under Gaussian likelihood. So that one matrix is...
It seems the issue resides in how the sampling pairs are generated (in full :O): See `shuffle-_combinations`, which considers the WHOLE dataset, it should be a generator in full, rather...
I am not providing a solution, just pointing to where the issue resides. A solution would require being smarter about how samples are created for a large dataset. For example,...
Highly anticipating this release! :) Keep up the great work
> Hi @NielsRogge, are you planning to do one of your wonderful notebook tutorials once this PR is closed? I'm rather curios on how can we approach a token-classification task...