Hi everyone,
There is the following code which load “checkpoint_50.pth.tar”, but I didn't find it in the project, please help to get the file.
"state_dict = t.load('./checkpoint/checkpoint_50.pth.tar')"

Hi. i am very interested in your code. My problem is to know if it is possible to use it for language translation. If it is the case how?

Thank you for sharing the code.
According to the paper, Appendix A 2nd paragraph, dropout is not used for attention.

In line 205, the residual and result are concatenated, but I think they should be added elementwise and then passed through a layer_norm (Figure 8 ANP paper). I wonder if there is some reason for this modification.

Thanks,
Deep Pandey

In networ.py BCELoss has the default settings, which (looking both for pytorch1.1 and pytorch0.4) does the mean reduction. However, the KL divergence function (also in network.py) seems to be using the sum reduction. Intuitively, either both should be sum or both mean. Is this a bug or is this correct?

Thanks!

The paper specifically mentions that they do not use attention in the latent path. Was there a reason that you added it into the latent path in your repo?

Hi,
Thanks for your implementation!
I am a little confused about result = t.cat([residual, result], dim=-1) in line 205 as you mentioned very important. Why do you need to concatenate the original residual result ?