MappingToLatent has no activation

[eitanrich]

Is it intentional that the D module (MappingToLatent) consists of three F.Linear layers w/o any activations (e.g. no ReLU / Leaky ReLU)?

ALAE/net.py

Line 894 in 5d8362f

block = ln.Linear(inputs, outputs, lrmul=0.1)

[podgorskiy]

That's an error. But I won't change it, to be consistent with the published results.
Most likely, the effect won't be significant, but I'm curious to see how the result will differ.

[eitanrich]

Maybe the behavior is related to Implicit Rank-Minimizing Autoencoder :)