Added pix2pix, CycleGAN, SRGAN

[shreyas-kowshik]

No description provided.

[shreyas-kowshik]

pix2pix and cycleGAN were partially trained (sort of overfitted on a small subset of the data) and srgan was not completely trained.

…

On Sun, Jan 19, 2020 at 9:02 PM Mateusz ***@***.***> wrote: Were these models trained? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#157?email_source=notifications&email_token=AHOTGLC5Z6YSQSRKGSQRQ6DQ6RXCPA5CNFSM4IMLOOO2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEJKVBQY#issuecomment-576016579>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHOTGLB7IXRHBSBRR76CTWTQ6RXCPANCNFSM4IMLOOOQ> .

-- *Shreyas Kowshik*Third Year UG | Department of Mathematics .............................................................................................. [image: photo] Email <shreyaskowshik@gmail.com> Github <https://github.com/shreyas-kowshik> Mobile : 9892425461 Indian Institute of Technology, Kharagpur ............................................................ ....................................

[gorskim]

Sorry, accidentally deleted my previous comment. Anyway: do you still plan to do something about it? If yes, I looked into the code of SRGAN and noticed that the padding was wrong - it will produce the output not as expected.

[shreyas-kowshik]

I plan to get these models trained soon. Just that I am not able to find enough time to debug them separately.
As for the dimension issues, what specific layer do you think is causing the issue?

[gorskim]

I think every _Conv and _ConvBN. The padding is 1, so after initial convolution in generator, the dimensions will be decreased and the final output will be not as big as expected. I compared it to some Python code and "they" usually use padding="same". Unfortunately, I don't think there is such an option in Julia (it needs to be calculated in the runtime, I guess).