search

researchmm / AOT-GAN-for-Inpainting

[TVCG'2023] AOT-GAN for High-Resolution Image Inpainting (codebase for image inpainting)
https://arxiv.org/abs/2104.01431
Apache License 2.0
425 stars 68 forks source link

Bugs in Adversarial Training. #6

Closed praeclarumjj3 closed 3 years ago

praeclarumjj3 commented 3 years ago

Hi! I ran a few experiments on training the AOT-GAN on the places2 dataset. However, looking at the loss plots seems like the adversarial loss isn't affecting the learning. The magnitude remains the same during the whole course of training.

You can see the experiment results here: https://github.com/praeclarumjj3/AOT-GAN-Experiments#5-experiments.

praeclarumjj3 commented 3 years ago

Ohh never mind, the results come out nice if we use the testing pconv dataset anyways. So it's cool!