Awesome

Deprecated as of March 11, 2022! Use https://github.com/neuroailab/convrnns instead (includes pretrained models).

Temporal Neural Networks

Run models in time.

Installation

git clone https://github.com/neuroailab/tnn.git
pip install -e tnn

(-e installs a developer version such that you can always update your code to the latest)

Note: networkx==1.11 is the latest version of the networkx package that works with this package (higher versions of networkx will not work).

Usage

Look at tutorials. tutorials/alexnet_example.py demonstrates the basic unrolling API with AlexNet. tutorials/customcell_example.py shows how to pass a custom cell to a model, and add edges.

tnn/convrnn.py contains examples of standard ConvRNN cells in the literature. tnn/resnetrnn.py contains the Reciprocal Gated Cell implementation (see https://arxiv.org/abs/1807.00053 for details). tnn/efficientgaternn.py contains the Efficient Gated Unit cell implementation used in https://arxiv.org/abs/2006.12373.

json contains a set of example graphs including 5 layer LSTM and Reciprocal Gated models. To use them with the customcell_example.py, set the global variables MODEL_JSON = 5L_imnet128_lstm345 and CUSTOM_CELL = tnn_ConvLSTMCell. You will also need to set the INPUT_LAYER and READOUT_LAYER to match the model JSON.

Contributors

• Jonas Kubilius (MIT)
• Daniel L.K. Yamins (Stanford)
• Maryann Rui (Berkeley)
• Harry Bleyan (MIT)
• Aran Nayebi (Stanford)
• Daniel Bear (Stanford)

License

MIT