A hidden size of 70 and 2 layers yield good results:
python train_hannds.py --hidden_size 70 --layers 2 --length 100 --cuda --cv_partition 1 --network 88 --rnn_type LSTM
Provide the --cuda flag to use the GPU.
python kalman_mapper.py
python hannds-server.py
Then open simple-text.maxpat in Max/MSP.