Do not apply activation for last MLP layer

[kevinddchen]

This fixes activation being applied on the last layer:

... -> linear -> `activation` -> `output_activation` -> (output tensor)

Also replacing tuple-lookup with set-lookup.