8 events
when toggle format what by license comment
Jun 26 '20 at 14:12 comment added harwiltz Ah, that's unfortunate. In PyTorch the implementation gives you the option to get the one hot representation. However I think you can actually get away with casting, I believe this is called "straight through sampling" or something.
Jun 26 '20 at 7:27 comment added Marco Garcia Macias Thank you for your reply harwiltz! I checked the Gumbel Softmax paper you mentioned, it is very interesting, I took a look into tensorflow implementation, tfp.RelaxedOneHotCategorical, the problem is that when you sample from the Gumbel distribution you get just the approximation of the one hot encoded vector, and my external software cannot receive any approximation, i'm forced to either Cast to int or modify the variables in such a way that I cannot do backward propagation
Jun 23 '20 at 14:07 comment added harwiltz This really doesn't sound like an RL problem to me. If the only issue you have is backpropagating through your thresholded softmax layer, I recommend you look into Gumbel Softmax. Maybe that can solve your problem?
S Jun 22 '20 at 12:03 history edited nbro CC BY-SA 4.0
Grammar improvements
S Jun 22 '20 at 12:03 history suggested BlueTurtle CC BY-SA 4.0
Grammar improvements
Jun 22 '20 at 10:22 review Suggested edits
S Jun 22 '20 at 12:03
Jun 22 '20 at 9:02 review First posts
Jun 22 '20 at 9:17
Jun 22 '20 at 9:02 history asked Marco Garcia Macias CC BY-SA 4.0