pathak22
AlexNet results
As I understand it all images in the paper are using the small (not AlexNet) architecture, since the adversarial loss makes the images look much nicer. I was hoping it is possible to share some of the images from the AlexNet implementation so that I can compare with the results of my own implementation.
(If there are other smart ways of performing intermediate validation of the AlexNet implementation before recreating the Table 2 results they are also welcome)
When I was training context encoder (~2yrs ago), I found it difficult for the training to converge when the generator is an AlexNet and discriminator is same as in the current code. Hence, I used a smaller architecture for the generator too. Maybe you can try AlexNet generator now with all the latest GAN tricks (200+ papers since then) and it might work better. :-)
In the paper, while using AlexNet in the generator, I discarded adversarial loss and only used L1.
