DSA: More Efficient Budgeted Pruning via Differentiable Sparsity Allocation (ECCV 2020)

[paper] [checkpoints]

Run test:

bash test.sh

• models/: all models. The codes of some basic models under this directory are copy and modified from this repo
• models/op.py: MaskedConvBNReLU with the differentiable pruning process
• graph_utils.py: topological grouping
• ckpt/: download checkpoints from this url, check the path in test.sh.

To also test the checkpoints on imagenet, specify the imagenet dataset path via the IMAGENET_PATH env. There should be two subdirs under the path: train, val.

IMAGENET_PATH=<imgnet_path> bash test.sh

You can cite the paper as

@article{ning2020dsa,
  title={DSA: More Efficient Budgeted Pruning via Differentiable Sparsity Allocation},
  author={Ning, Xuefei and Zhao, Tianchen and Li, Wenshuo and Lei, Peng and Wang, Yu and Yang, Huazhong},
  journal={arXiv preprint arXiv:2004.02164},
  year={2020}
}