I’m going through the RNN notebooks (I have the kindle version of the second edition) and still can’t grasp how to choose the unit size of the LSTM layers.
In the univariate time series notebooks it’s says we’re using 20 hidden units, equivalent to one month, but I believe unit=10, which is from the book.
Then in the next notebook for the stacked LSTM we’re using 25, 10. No real idea how those were chosen.
Just curious about the rationale.
I’m new here and haven’t touched LSTMs for a bit but I figure if there’s a chance I can be helpful I should take a shot at it. The idea behind an LSTM is that you want the system to be influenced by data that came from previous steps and what that data would consist of and the rational behind it might change on a case by case basis, can you reference the notebooks in question?
Hi, I’m looking at chapter 19 on RNN and really any notebook in there with an LSTM, such as 02_stacked_lstm_with_feature_embeddings.