UIUC CS521, Research Project Exploring Deep Compression for DNNs
Deep Compression is a pipeline for reducing the size of deep neural nets using a combination of pruning, quantization, and encoding, originally described by Han, Mao, and Dally.
Reproduce the paper's results, using the PyTorch framework
- For AlexNet
- For VGG-16
Explore the following pruning techniques
- L1 structured pruning
- TBD
Explore the following quantization techniques
- Incremental network quantization
- TBD
The alexnet.py script expects that the ImageNet dataset hosted on Kaggle is available locally. The annotations and data are in separate directories, which means they'll need to be zipped together for validating the model. The data loader expects the locations of those two directories are available in the environment. They can be populated in .env as IMAGENET_ANNOTATIONS_DIR and IMAGENET_DATA_DIR, respectively.