Full-Stack Filters

Code for Paper: Full-Stack Filters to Build Minimum Viable CNNs (MVNet)

In this paper, we propose a novel convolution operation with full-stack filters and auxiliary masks to reduce redundancy in convolution filters for building minimum viable CNNs, i.e. MVNets. A series of full-stack filters with additional binary masks which requiring negligible memory resource in the proposed convolution method can generate much more inherited sub-filters and achieve approximate performance to the original convolution.

Description

• MVNet with Full-Stack Filters inference code on CIFAR-10 dataset.
• We provide the trained model in models/cifar10-s64.th which occupies 15 MB, including 1.0 MB float32 weights and 14.0 MB binary masks stored in int8 format (so it's 1.75 MB in bool format which is not supported in PyTorch yet).

Files description

Requirements: Python 3.6, PyTorch 0.4

• MaskConv2d in mask_conv.py is the new convolution class with full-stack filters used for replacing nn.Conv2d.
• mask_vgg.py is the VGG-16 using the proposed full-stack filters.
• test.py is the test script on CIFAR-10 dataset.

Usage

1. Download CIFAR-10 python version from cifar-10-python.tar.gz, put it in anywhere you like such as $data/cifar-10-python.tar.gz.
2. Run python test.py --resume=models/cifar10-s64.th --data-dir=$data in default setting, you will get 93.1% accuracy.

Performance

Performance on CIFAR-10 with VGG-16 as architecture:

For other experimental results, please refer to the original paper.

Citation

If you use these models in your research, please cite:

@article{han2019full,
  title={Full-Stack Filters to Build Minimum Viable CNNs},
  author={Han, Kai and Wang, Yunhe and Xu, Yixing and Xu, Chunjing and Tao, Dacheng and Xu, Chang},
  journal={arXiv preprint arXiv:1908.02023},
  year={2019}
}