More modeling options

[ndiamant]

More modeling options

[ndiamant]

Updating with more options would be easier if _conv_block1d and _conv_block2d, etc. were the same instead of separate functions. @lucidtronix why are they different?

[lucidtronix]

@ndiamant, I agree it would be great to unify those functions. They are different because we call several different functions based on the shape of the input x: Conv1D, MaxPooling1D, and SpatialDropout1D, for example. We could make it a second order function and pass these functions in as arguments. The tradeoff is that the fxn signature gets much more complicated, but we get rid of 3x code duplication. Optimization scheme and VAE could be implemented outside of these fxns though.

[lucidtronix]

Also adding todos for our model builder:

• More activation fxns (PreLUs, Selus, ...)
• More initialization methods (at least enable all keras strings glorot_uniform xavier etc
• More flexible autoencoder topologies (at least backdoor connection toggle)
• TensorMap parents should be specified by TensorMap not by strings
• Allow multiple input TensorMaps of the same shape (I think this doesn't work yet)

[lucidtronix]

we basically did this stuff.