- PyTorch 1.1
- PIL
- Numpy
- PyPlot
- Scipy
The goal of this project is to investigate the compression of deep learning neural networks, as discussed in Han, S., Mao, H. and Dally, W.J., 2015. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. This project focuses on AlexNet, but the method can be abstracted to any other network.
As the title suggests the project applies 3 compression methods to a model in an effort to reduce the size on disk. They are:
1. Pruning: The process of finding and removing redundant weights.
2. Quantisation: Limiting the number of effective weights.
3. Huffman Encoding: Algorithmically finding the most efficient way to store the weights.
Ensure that all dependencies are installed.
Simply run DeepCompression.py
Lion is provided as an example. To use your own image for testing, simply point to program to a different directory containing your photo(s). The program can test on more than one image, it will iterate over all images in the directory.
This can be adjusted from the Threshold method
Threshold(quality parameter for CONV layers, quality parameter for FC layers)
Similar to QP this can adjusted from the Quantise method
Quantise(thresholded_model, number of bits per weight)