RNN Sampling novel sequences with mixed types
I want to generate novel sequences with RNN similar to Char RNN in the programming assignment. The training examples would have shape (batch_size, Tx , 11).
The last dimension,11 = 10 (one hot representation representing 10 distinct actions) + 1(timestamp converted to minutes, an integer)
Therefore, on the output of the RNN, I would require 2 different parts, the first part being a softmax which predicts from the set of10 distinct actions, and a non-linear output which predicts the timestamp.
I have 2 questions:
1. Is it possible to implement the 2 types of output in Keras or there's other way or framework.
2. I am using Keras. How to deal with variable input sequence length, Tx. Do I have to change how cost is calculated based every times the timesteps varied?