search

Harrypotterrrr / DVD-GAN

A pytorch implementation of Efficient Video Generation on Complex Datasets
62 stars 20 forks source link

Reason to use tanh instead of ReLU in ConvGRUCell? #2

Open cvcore opened 4 years ago

cvcore commented 4 years ago

Hi @Harrypotterrrr ,

thanks for the nice implementation. I was wondering if there is a specific reason for you to choose tanh instead of ReLU for the activation of out_gate, as in the paper https://arxiv.org/pdf/1907.06571v2.pdf?

To be specific, it's this line I'm referring to: https://github.com/Harrypotterrrr/DVD-GAN/blob/d8c4b6a4adb6217eb959eed0ea041a67589869cf/Module/ConvGRU.py#L51

Harrypotterrrr commented 4 years ago

Hi, I haven't done this work for a long time since the result is not plausible, which I think is caused by limited computing resources, comparing Nvidia used 100 x V100 to train the DVD-GAN. Even so, I think there are still some bugs and omissions of this implementation. Anyway, thanks for your pushing this issue. Could you please try this out and find if the result improves? Moreover, I think you could try to contact @vominhduc for more information? @vominhduc Do you have any progress on this implementation? :confounded:

vominhduc commented 4 years ago

Hi @cvcore @Harrypotterrrr , I have tried to improve the code to some extent but the results are not improved much. Currently, we can observe the shape of the people but the quality is too bad. I am still looking for a better PC to train the code. Btw, I will keep you updated if any progress.

Harrypotterrrr commented 4 years ago

@vominhduc Thanks for your effort sincerely.

cvcore commented 4 years ago

Thanks for the reply @Harrypotterrrr @vominhduc ! I look forward to your further update.