sammlapp
commented
1 year ago The reason I've stopped previous attempts to implement RNNs is the realization that they are fundamentally designed to train on and predict series-like data, not a single label of a sample. There are useful concepts from RNN and LSTM models like the ways that features are extracted from time-series data, but the models themselves are not directly applicable to the typical (classification) problem we are trying to solve.
This example might say otherwise though - they are performing classification of single labels (country of origin) based on samples (names) that have series-like structure (letters). This suggests that RNN could in fact be applied to our type of problem
Add module for using pytorch RNNs. It might be able to subclass the current cnn.PytorchModel base class. (If it does, this suggests that class should be moved out of the cnn module)