AOT-GAN Experiments
I conducted various experiments with AOT-GAN proposed in the paper: Aggregated Contextual Transformations for High-Resolution Image Inpainting) on the Places2 dataset using the PConv Free Form Masks for Inpainting.
In particular, I focused on finding the effectivity of different losses while training the framework. I made several observations:
-
The adversarial loss doesn’t seem to be contributing to the learning of the model as it stays almost the same throughout the training.
-
Training for longer than
1e4
iterations doesn’t add much improvement to the results. -
Training without style loss produces blurry results. Therefore, style loss is an important component for texture related synthesis of images.
-
Training without adversarial loss also produces good quality results!
For more info, please see the Code on GitHub
.