SparseMask: Differentiable Connectivity Learning for Dense Image Prediction.
This project is maintained by wuhuikai
[Project] [Paper] [arXiv] [Home] [AWESOME NAS]
Official implementation of SparseMask: Differentiable Connectivity Learning for Dense Image Prediction. Automatically design the connectivity structure for dense image prediction tasks, achieving better fusion of multi-scale feature maps.
@inproceedings{wu2019sparsemask,
title = {SparseMask: Differentiable Connectivity Learning for Dense Image Prediction},
author = {Wu, Huikai and Zhang, Junge and Huang, Kaiqi},
booktitle = {arXiv preprint arXiv:1904.07642},
year = {2019}
}
Contact: Hui-Kai Wu (huikaiwu@icloud.com)
python==3.5
pytorch==1.0
cuda==9.0
scipy
scikit-image
tqdm
tensorboardX
tensorflow
ROOT
├── benchmark_RELEASE
└── VOCdevkit
python VOC12/convert_mat_to_png.py --sbd_path [ROOT]/benchmark_RELEASE
python VOC12/convert_labels.py \
[ROOT]/VOCdevkit/VOC2012/SegmentationClass \
[ROOT]/VOCdevkit/VOC2012/ImageSets/Segmentation/trainval.txt \
[ROOT]/VOCdevkit/VOC2012/SegmentationClass_1D
cd [ROOT]
mv VOCdevkit/VOC2012/SegmentationClass_1D/*.png benchmark_RELEASE/dataset/cls_png/
mv VOCdevkit/VOC2012/JPEGImages/*.jpg benchmark_RELEASE/dataset/img/
ln -s [ROOT]/benchmark_RELEASE/dataset/cls_png data/gt
ln -s [ROOT]/benchmark_RELEASE/dataset/img data/img
python train_sparse_mask.py --search
python prune.py --checkpoint search_checkpoint/checkpoint_33100.pth.tar
python train_sparse_mask.py --mask_path search_checkpoint/mask_thres_0.001.npy
python train_sparse_mask.py --mask_path search_checkpoint/mask_thres_0.001.npy --training_list VOC12/data/train.txt --lr 0.0005 --ft_model [MODEL_PATH]
python eval_sparse_mask.py --pretrained_model train_checkpoint/checkpoint_4600.pth.tar --mask_path search_checkpoint/mask_thres_0.001.npy
Part of the work was conducted while I was an intern in Preferred Networks.