hmlstm

This package implements the Hierarchical Multiscale LSTM network described by Chung et al. in https://arxiv.org/abs/1609.01704

The network operates much like a normal multi-layerd RNN, with the addition of boundary detection neturons. These are neurons in each layer that, ideally, learn to fire when there is a 'boundary' at the scale in the original signal corresponding to that layer of the network.

Installation

pip install git+https://github.com/n-s-f/hierarchical-rnn.git

Or, if you're interested in changing the code:

git clone https://github.com/n-s-f/hierarchical-rnn.git
cd hierarchical
python setup.py develop

Character Classification

In this example, we'll consider the text8 data set, which contains only lower case english characters, and spaces. We'll train on all batches but the last, and test on just the last batch.

from hmlstm import HMLSTMNetwork, prepare_inputs, get_text

batches_in, batches_out = prepare_inputs(batch_size=10, truncate_len=5000, 
                                         step_size=2500, text_path='text8.txt')
                                         
network = HMLSTMNetwork(output_size=27, input_size=27, embed_size=2048, 
                        out_hidden_size=1024, hidden_state_sizes=1024, 
                        task='classification')

network.train(batches_in[:-1], batches_out[:-1], save_vars_to_disk=True, 
              load_vars_from_disk=False, variable_path='./text8')

predictions = network.predict(batches_in[-1], variable_path='./text8')
boundaries = network.predict_boundaries(batches_in[-1], variable_path='./text8')

# visualize boundaries
viz_char_boundaries(get_text(batches_out[-1][0]), get_text(predictions[0]), boundaries[0])

Time series prediction

In this example, we'll do three-step-ahead prediction on a noisy set of signals with sinusoidal activity at two scales.

from hmlstm import HMLSTMNetwork, convert_to_batches, plot_indicators

network = HMLSTMNetwork(input_size=1, task='regression', hidden_state_sizes=30,
                       embed_size=50, out_hidden_size=30, num_layers=2)
                       
# generate signals
num_signals = 300
signal_length = 400
x = np.linspace(0, 50 * np.pi, signal_length)
signals = [np.random.normal(0, .5, size=signal_length) +
           (2 * np.sin(.6 * x + np.random.random() * 10)) +
           (5 * np.sin(.1* x + np.random.random() * 10))
    for _ in range(num_signals)] 
    
batches_in, batches_out = convert_to_batches(signals, batch_size=10, steps_ahead=3)


network.train(batches_in[:-1], batches_out[:-1], save_vars_to_disk=True, 
              load_vars_from_disk=False, variable_path='./sinusoidal')

predictions = network.predict(batches_in[-1], variable_path='./sinusoidal')
boundaries = network.predict_boundaries(batches_in[-1], variable_path='./sinusoidal')

# visualize boundaries
plot_indicators(batches_out[-1][0], predictions[0], indicators=boundaries[0])

Further information

Please see the doc strings in the code for more detailed documentation, and the demo notebook for more thorough examples.

Pull requests or open github issues for improvements are very welcome.