A Python library for converting trained TensorFlow ANNs to GeNN spiking neural networks with low loss in performance using the algorithm proposed by Diehl et al. in "Fast-Classifying, High-Accuracy Spiking Deep Networks Through Weight and Threshold Balancing"
Below are the results of conversion experiments on MNIST and CIFAR-10 with and without data-based normalization.
ANN model:
Trained on 60000 examples for 5 epochs.
Train accuracy: 99.21% on 60000 examples
Test accuracy: 98.86% on 10000 examples
GeNN model:
Conversion parameters:
single_example_time = 350
dense_membrane_capacitance = 1.0
sparse_membrane_capacitance = 0.2
neuron_threshold_voltage = -57.0
neuron_resting_voltage = -60.0
On three trials: (all accuracies on test set)
Note: ANN test set contains 10000 example while the GeNN accuracies have been computed on 500 examples.
| ANN accuracy | GeNN (without normalization) | GeNN (with normalization) |
|---|---|---|
| 98.91% | 97.8% | 98.0% |
| 98.88% | 96.8% | 97.2% |
| 98.88% | 96.8% | 96.6% |
ANN model:
Trained on 50000 examples for 10 epochs.
Train accuracy: 90.99% on 50000 examples
Test accuracy: 77.29% on 10000 examples (overfitting since I haven't added dropout layers yet)
GeNN model:
Conversion parameters:
single_example_time = 3750 (though it takes only ~2500 with no weight normalization)
dense_membrane_capacitance = 0.1
sparse_membrane_capacitance = 0.01
neuron_threshold_voltage = -58.0
neuron_resting_voltage = -60.0
On one trials: (all accuracies on test set)
Note: ANN test set contains 10000 example while the GeNN accuracies have been computed on 500 examples.
| ANN accuracy | GeNN (without normalization) | GeNN (with normalization) |
|---|---|---|
| 77.29% | 75.0% | 77.0% |
- Python 3
- TensorFlow
- PyGeNN (see here for installation instructions)
- Matplotlib
Note: Since work is still in progress, we don't have a standard installation procedure. However, in the meantime, you can follow the steps below.
- Clone the repository:
git clone https://github.com/genn-team/tensor_genn.git - Change directory:
cd ~/tensor_genn - Create a new file and follow a procedure shown in
conversion_example.pyto convert your TensorFlow model into GeNN.
- Peter U. Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. 2015. Fast-Classifying, High-Accuracy Spiking Deep Networks Through Weight and Threshold Balancing. IJCNN (2015)