Jingwei Zhang

Thanks for the quick reply! I've been digging the pointer nets paper and the tsp drl paper for the past few hours, and now this is what I think: 1....

And why I think that for the `critic` the loop is not over `city_t` but rather `num_process_steps` is that: if you still loop through the sequence and the `dec_input` is...

Hi, thanks a lot for the reply! Your repo helps me a lot in understanding the technical details, I'll try to experiment a bit more and turn for your help...

So I actually find the solution to this problem. Just have to run ```"th -ldisplay.start 8000 0.0.0.0"``` before you call ```display.image``` Before in the lua version, the package will work...

Oh thanks for the quick reply! I understand that e.g. when integrating the plots of multiple runs the standard error should be calculated and plotted, however is there a particular...

@AjayTalati You said you also have an implementation so I guessed you also used the ```cumprod``` for computing the ```allocation weights```, for me the backward simply fails with ```cumprod```, do...

Ah I just found that you have a custom ```cumprod``` function. But if I am reading it correctly, this is treating the input only as a tensor but not as...

@ypxie Hey, thanks a lot for the reply! Can you explain further why it is a leaf node? When we input a sequence row by row, doesn't the current usage_t...

@AjayTalati Hey, it's not an error from this repo, but when I use the current ```cumprod``` from pytorch, and is due to the autograd has not been implemented for this...

@ypxie I mean the allocation_weight depend on the usage, then the write_weight depend on the allocation weight, then the memory is updated with the write_weight, and the usage_{t} is also...