I am adopting the approach used in this repository to train two discriminators simultaneously.
https://github.com/KangBK0120/D2GAN

D2GAN paper
https://arxiv.org/pdf/1709.03831.pdf

Other related papers
https://arxiv.org/pdf/1811.03850.pdf

Using normalization in custom transforms required as below:

normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                 std=[0.229, 0.224, 0.225])

or

normalize = transforms.Normalize(mean= [0.5, 0.5, 0.5],
                                 std=[0.5, 0.5, 0.5])

Because of implementation overview, I've just considered inputs in arbitrary sizes. After all other parts being fully implemented, we can make sure of the size.

Which hidden layer should be chosen as a low-spatial latent vector to pass to loss function?
In the DetailsNet, it seems layers 2 and 5 have been chosen on VGG19BN.

Note: the last layer of VGGs is classifier so we cannot use them because of 1-D vector representation of the last layer to match with a class of object.

In the layers in line 70, we have to use Tanh to map output between [-1, 1], but because we did not apply any normalization at first to map our images from [0, 1] to [-1, 1], we use Sigmoid for testing purposes, then after applying normalization with respect to issue #2, we can proceed to use Tanh as an activation function.

And the activation:

Set inplace=False in all of the submodules of DetailsNet and Discriminators.

https://discuss.pytorch.org/t/whats-the-difference-between-nn-relu-and-nn-relu-inplace-true/948/2?u=nikronic