Unexpected behavior of the "include_top" argument about zoo HOT 3 CLOSED

I agree, we should do the same as Keras.

If I recall correctly the reasoning behind not including relu if inlcude_top=False was that if you want to add more binary layers after the network it might not be desired to have the relu included. This is not a problem for full precision networks as there should always be non-linearity between layers.

from zoo.

Hey @nixingyang, thank you for your interest in Larq and Larq Zoo!

It does indeed look like our behaviour in this regard differs from that of keras-applications. @lgeiger / @koenhelwegen thoughts on moving the activation out of the if statement?

from zoo.

I see no reason to deviate from keras here, so would be happy to change it (but I think @lgeiger knows these models best, perhaps he knows why it's currently like this).

from zoo.