tohinz / pytorch-mac-network Goto Github PK

Hi,

Thanks for sharing this great repo! It's awesome.

I just have a few questions about the control unit:
(1) From

these lines seem to be commented out -- but I think these are used for masking padded tokens in the sequence. Is that intentional? Do I miss something here?
(2) It seems in this control function,
Previous control is never passed in as an argument, so the previous control unit is not passed to next in the reasoning step. I noticed that in the docstring an argument of control is mentioned but it seems to be removed from the function signature. Also the question_lengths is not used since the mask function call is commented out (previous question). Is that intentional as well?

Thanks a lot!

No critical, only the file is missing

Thanks for your work

Thank you for the great work.

I am trying to reproduce results on subset of 10% training questions of CLEVR but apparently, the pytorch implementation is not able to achieve the same numbers as in the paper. Have you ever tried this with your implementation?

Regards,
Thao

Hi!

First, thanks for your work, your code is very clean!

Second, I just trained the model with the default config file and it reaches maximum of 96.59% accuracy on the validation set.

Do I have to tune hyperparameters to reach your scores or is it an initialization issue? How many tries did you do before getting that accuracy?

Best regards and thanks again for your work!