koz4k / pruning

Neural network pruning

This repo provides an investigation of two methods of pruning neural networks: weight pruning and neuron pruning. An MLP is trained on MNIST and pruned with various fractions of weights/neurons left in the network. Results in terms of prediction accuracy and inference time are plotted.

The code has been tested on Python 3.6 and TensorFlow 1.12. To install the dependencies:

pip install -r requirements.txt

To run:

python main.py

Note that it can take several minutes to run, as inference is done on CPU to get accurate measurement.