search

aimagelab / VATr

MIT License
74 stars 4 forks source link

Cycle loss #5

Closed Fran-zis-ka closed 1 year ago

Fran-zis-ka commented 1 year ago

Hi, I've read your paper and your code and it's a great work:) However, I've got a question: in the code you provided, it seems like that the cycle loss you mentioned in the paper is not used to reinforce the model. In models/model.py, I found "Lcycle1" and "Lcycle2" two variables are initialized zero and not updated ever since. May I know the reason why? Thanks a lot!

vittoriopippi commented 1 year ago

Hi @Fran-zis-ka thank you for pointing this out. We double-checked in these days, and we apologize because we discovered that the Cycle Loss was disabled. In particular, the --is_cycle flag was detached and had no effect on the behavior of the network.

In any case, the --is_cycle flag is currently working, to use it you have run the training in this way:

python train.py --is_cycle
Fran-zis-ka commented 1 year ago

ok, thank you for your quick response:)