Fixes to experiments
kudkudak opened this issue · comments
Stanislaw Jastrzebski commented
In second round of experiments (a bit annoying due to human labor) we should fix things I found:
- No l2 regularization in DNN on hidden layers, and in factorized as well (at least it is consistent)
- Sharing mode in factorized is by default
1
. This looks weird.
(List might grow)