Constraint-Aware Deep Neural Network Compression

Given a real-time operational constraints, this library can automatically compress the network to satisfy the constraint while preserving the accuracy. This library is built on top of SkimCaffe, which has implemented direct sparse convolution operations and has an effective speedup given an input sparse network. This framework can be applied to different constraint types(latency, memory size), different network(Alexnet, Resnet, Googlenet) and different datasets(ImagenNet, DTD).

For more technical details, please refer to the paper.

Usage

1. First follow the build-up instructions in SkimCaffe to build SkimCaffe.
2. Modify the cfp.config file to adapt to your own need(constraint values).
3. Run training in root directory:

python2 main.py

4. To visualize the training result, specify the output directory:

python2 pruning/visualize_cbo_results.py [OUTPUT_DIR]

Results

Visualization of sampled data points in the first exponential cooling step

Citation

If you use this code or ideas from the paper for your research, please cite our paper:

@InProceedings{Chen_2018_ECCV,
author = {Chen, Changan and Tung, Frederick and Vedula, Naveen and Mori, Greg},
title = {Constraint-Aware Deep Neural Network Compression},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}