This repository contains the PyTorch test implementation of: Channel-wise Distillation for Semantic Segmentation

All the codes are tested in the following environment:

• Linux (tested on Ubuntu 16.04 / CentOS 7.6)
• Python 3.6.2
• PyTorch 0.4.1
• Single TITAN Xp GPU

• Install PyTorch: conda install pytorch=0.4.1 cuda90 torchvision -c pytorch
• Install other dependences: pip install opencv-python scipy
• Install InPlace-ABN:

cd libs
sh build.sh
python build.py

The build.sh script assumes that the nvcc compiler is available in the current system search path. The CUDA kernels are compiled for sm_50, sm_52 and sm_61 by default. To change this (e.g. if you are using a Kepler GPU), please edit the CUDA_GENCODE variable in build.sh.

• Dataset: [Cityscapes]

• After distillation: PSPNet (ResNet-18) rn18-cityscape_singleAndWhole_val-75.05_test-73.86.pth [Google Drive] rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth [Google Drive]

Please create a new folder ckpt and move all downloaded models to it.

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.02_test-73.86.pth --gpu 0 --type val  --figsavepath 75.02val

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth --gpu 0 --type val  --figsavepath 75.90val

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.02_test-73.86.pth --gpu 0 --type test --figsavepath 73.86test

python valandtest.py --data-dir path/to/CITYSCAPES/data/cityscapes  --restore-from ckpt/new_rn18-cityscape_singleAndWhole_val-75.90_test-74.58.pth --gpu 0 --type test --figsavepath 74.58test

Please consider citing this work if it helps your research:

@inproceedings{shu2020cwd,
  title={Channel-wise Distillation for Semantic Segmentation},
  author={Shu, Changyong and Liu, Yifan and Gao, Jianfei and Xu, Lin and Shen, Chunhua},
}