neural network - Backpropagation through time in stateful RNNs -
if use stateful rnn in keras processing sequence of length n divided n parts (each time step processed individually),
- how backpropagation handled? affect last time step, or backpropagate through entire sequence?
- if not propagate through entire sequence, there way this?
the propagation horizon limited second dimension of input sequence. i.e. if data of type (num_sequences, num_time_steps_per_seq, data_dim)
prop done on time horizon of value num_time_steps_per_seq
take @
Comments
Post a Comment