[bug report (serious)] All types of autoencoder and VAE have linear activation.

[ncble]

Describe the bug
A normal autoencoder or VAE for images should have either sigmoid or tanh activation, but in the current master version, we have linear activation (i.e. no activation). This bug seriously affects all the previous work (results and the conclusion). (especially the paper: https://arxiv.org/abs/1901.08651 and maybe the other works built on the toolbox, etc. ) By fixing this issue, I have noticed substantial improvements and more coherent results. Some other issues may be related to this bug, e.g. #24

Code example
For both autoencoder and VAE (CNN version):
https://github.com/araffin/srl-zoo/blob/438a05ab625a2c5ada573b47f73469d92de82132/models/models.py#L65-L83

For MLP version:
https://github.com/araffin/srl-zoo/blob/438a05ab625a2c5ada573b47f73469d92de82132/models/vae.py#L22-L28

The above examples are not exhaustive.

P.S. I have confirmed this issue with @kalifou. My solution has already committed and pushed (c.f. my pull request)

[araffin]

As discussed, that's ok when using "image net" normalisation, however, you should get better results with activation when using "tf" normalisation.