Skip to content

Latest commit

 

History

History
80 lines (62 loc) · 3.38 KB

README.md

File metadata and controls

80 lines (62 loc) · 3.38 KB
Raw

WaveletCompressedConvolution

Official implementation for Wavelet Feature Maps Compression for Image-to-Image CNNs, NeurIPS 2022.

Use as a Drop-in Replacement

Code example is available at WCC/transform_model.py.
Note that it is a common practice to avoid quantizing/compressing the first and last layers of the network.

Recreating Paper Experiments

For best results it's recommended to reduce the number of bits and the compression rate gradually.
E.g., load quantized 8/8 checkpoint to train quantized 8/6, load 4/8 checkpoint to train wcc 4/8 50%, and load wcc 25% to train wcc 12.5%.

In all the experiments we used a popular implementation, and changed only the main file to include the model transform after creation. For example, from the Deeplab implementation:

WCC.wavelet_deeplabmobilev2(model, opts.wt_levels, opts.wt_compression, opts.bit_w, opts.bit_a)

1. Object Detection

We used the following implementation:
https://github.com/rwightman/efficientdet-pytorch

Precision Wavelet Shrinkage BOPs(B) mAP ↑
FP32 --- 6,144 40.08
4/8 --- 280.4 31.44
4/8 50% 198.5 31.15
4/8 25% 155.4 27.49

2. Semantic Segmentation

We used the following implementation:
https://github.com/VainF/DeepLabV3Plus-Pytorch

We trained the model with the optional flag --separable_conv.

Cityscapes results: (see paper for Pascal VOC as well as more configurations)

Precision Wavelet Shrinkage BOPs(B) mIoU ↑
FP32 --- 36,377 0.717
8/8 --- 2,273 0.701
8/6 --- 1,705 0.683
8/4 --- 1,136 0.173
8/8 50% 1,213 0.681
8/8 25% 673 0.620
8/8 12.5% 403 0.552

3. Depth Prediction

We used the following implementation:
https://github.com/nianticlabs/monodepth2

Precision Wavelet Shrinkage BOPs(B) AbsRel ↓ RMSE ↓
FP32 --- 1,163.6 0.093 4.022
8/8 --- 133.6 0.092 4.018
8/4 --- 99.26 0.097 4.166
8/2 --- 82.1 0.268 8.223
8/8 50% 103.9 0.098 4.217
8/8 25% 88.5 0.112 4.663
8/8 12.5% 80.8 0.131 5.046

4. Super-resolution

We used the following implementation:
https://github.com/sanghyun-son/EDSR-PyTorch

Citation

@inproceedings{finder2022wavelet,
  title={Wavelet Feature Maps Compression for Image-to-Image CNNs},
  author={Finder, Shahaf E and Zohav, Yair and Ashkenazi, Maor and Treister, Eran},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}