Sequence Models, Week 1, Jazz Assignment

In the Jazz assignment, we are training the LSTM and Dense layers and using these layers to generate a sequence of values. And the inference_model takes inputs x0,a0 and c0. When we are finally using the inference model to predict a sequence, we are initializing x0, a0 and c0 to zeros.

  1. Will the sequences that are generated always be the same? Because in the inference model we are alwyas taking argmax in the softmax output.
  2. In the inference_model we are using initial state of a0 and c0 for the LSTM cell. Doesn’t our LSTM_cell already learn all activations and memory cells?
  1. Yes.
  2. We don’t learn the initial state. We assume some initial values, and then learn the final values during training.
1 Like