agadetsky / pytorch-pl-variance-reduction Goto Github PK

Can you add license information?

I am confused about the comments in estimators.py:
"in all computations (including those made by the user later), we don't
"want to backpropagate past "logits" into the model. We make a detached
"copy of logits and rebuild the graph from the detached copy to z

logits = logits.detach().require_grad(True)

if logits are parametrized by a network theta, then if you detach here then the network won't get updated right?

Hello,

Thanks a lot for your code. A really interesting work. Will the code for experiments on DAGs be released? It would be very nice to see them on some real problems like causal inference as given in the paper.

I tried to manually reproduce the results for 10, 20, 50 nodes but unfortunately failed to do so. If the code is not being released for it, can I kindly know some details regarding the implementation of accelerated proximal method (hyperparmeters, optimizer used, number of iterations run, learning rate etc.).

Thanks a lot for your help.