Ian Pointer

“you can also play with the LSTM. We’ve used a simple approach, but you can increase num_layers to create stacked LSTMs, increase or decrease the number of hidden features in the layer, or set bidirectional=true to create a biLSTM. Replacing the entire LSTM with a GRU layer would also be an interesting thing to try; does it train faster? Is it more accurate? Experiment and see what you find!”
― Ian Pointer, Programming PyTorch for Deep Learning: Creating and Deploying Deep Learning Applications