JunlinHan
inference
I don't understand why inference (test of the model) need testA and testB?
Why can't we just do inference with the model using a single folder of image that we want to translate to B?
Thanks for your help.
Bests
Hello, remybonnav,
Yes, you can definitely test a single model only. But this requires some changes in code (either in DCL model or putting DCL checkpoints/models to the CUT code), which might not be as convenient as simply testing both A/B.
I understand, but what is the purpose of CUT model then if you cannot use your model on a dataset of images A only (or B only)? In my case, i just trained for few epochs, and the results of FakeA and fake B just seem to be fakeA =realB and fakeB =realA (in a really degraded color). My images in test A and testB have the same name (they are paired).
Am I missing something?
On Sun, 18 Dec 2022, 07:22 JunlinHan @.***> wrote:
Hello, remybonnav,
Yes, you can definitely test a single model only. But this requires some changes in code (either in DCL model or putting DCL model to the CUT code), which might not be not as convenient as simply testing both A/B.
— Reply to this email directly, view it on GitHub https://github.com/JunlinHan/DCLGAN/issues/17#issuecomment-1356685954, or unsubscribe https://github.com/notifications/unsubscribe-auth/AODP7RFMTM5NWTXSNZUTRUTWN2USJANCNFSM6AAAAAATCE53PA . You are receiving this because you authored the thread.Message ID: @.***>
Hello remybonnav, CUT model can support one direction translation (test A or B only). The architecture (ResNet-based Generator) of CUT and DCLGAN are identical. So if you want to test single direction translation you might use the CUT model but load your DCLGAN checkpoints.
For your case, try to train for longer epochs (>20) with sufficient images (>100), otherwise, the results might be degraded.