2

I want to compare three different types of RNNs to decide which architecture can handle my data best.

To do that, I want them to have the same complexity. Can I simply define the complexity by the number of trainable parameters? If not, why?

Suppose I can compute the network's complexity like this, how do I correctly compute the complexity of a vanilla RNN cell, a GRU cell and an LSTM cell?

Gilfoyle
  • 141
  • 4
  • Just for clarity, you would like to compare space and time complexity across rnn layer classes in terms of scalability? – hH1sG0n3 Nov 10 '22 at 15:33

1 Answers1

1

One proxy for complexity is training time. You can benchmark different architectures to see how running time differs.

Brian Spiering
  • 21,136
  • 2
  • 26
  • 109