Planned Extensions on Shift-Net about shift-net_pytorch HOT 3 CLOSED

#22 This PR adds a lot of new features, thanks to @tchaton !

from shift-net_pytorch.

I agree with the roadmap so far.

I have added densenet based discriminator, and the game seems more equilibrated, as the generator is not easily winning.

I will focus on finishing U-Net Inception model, to give more power to the generator. However, I think we should also inspire ourself with DeepLabV3. Check this medium link from segmentation methods. We want the generator to be able to pull global features to fill up the gap, with shift layer help as much as possible.
(https://medium.com/@arthur_ouaknine/review-of-deep-learning-algorithms-for-image-semantic-segmentation-509a600f7b57)

from shift-net_pytorch.

It looks good.
I will train acc_unet_shift_triple and soft_unet_triple to see the results.
By the way, your PR doesn't work well in my computer(python2, pytorch0.4). Some extra fixes are also included. I will change the code a little bit later in this day.

from shift-net_pytorch.