- Install the packages required by backRazor
- Install packages
pip install visdom matplotlib
# install actnn
git clone [email protected]:ucbrise/actnn.git
cd actnn/actnn
pip install -v -e .
- Prepare datasets
- Standard Pascal VOC You can run train.py with "--download" option to download and extract pascal voc dataset. The defaut path is './datasets/data':
/datasets
/data
/VOCdevkit
/VOC2012
/SegmentationClass
/JPEGImages
...
...
/VOCtrainval_11-May-2012.tar
...
- Pascal VOC trainaug: ./datasets/data/train_aug.txt includes the file names of 10582 trainaug images (val images are excluded). Please to download their labels from Dropbox or Tencent Weiyun. Those labels come from DrSleep's repo.
Extract trainaug labels (SegmentationClassAug) to the VOC2012 directory.
/datasets
/data
/VOCdevkit
/VOC2012
/SegmentationClass
/SegmentationClassAug # <= the trainaug labels
/JPEGImages
...
...
/VOCtrainval_11-May-2012.tar
...
Baseline
CUDA_VISIBLE_DEVICES=1 python main.py --model deeplabv3_mobilenet --vis_port 23632 --gpu_id 0 --year 2012_aug --crop_val --lr 0.01 --crop_size 513 --batch_size 16 --output_stride 16
BackRazor
python main.py --model deeplabv3_mobilenet --vis_port 23632 --gpu_id 1 --year 2012_aug \
--crop_val --lr 0.01 --crop_size 513 --batch_size 16 --output_stride 16 --backRazorR 0.7
The partial code of this implement comes from DeepLabV3Plus-Pytorch
@inproceedings{
jiang2022back,
title={Back Razor: Memory-Efficient Transfer Learning by Self-Sparsified Backpropogation},
author={Jiang, Ziyu and Chen, Xuxi and Huang, Xueqin and Du, Xianzhi and Zhou, Denny and Wang, Zhangyang},
booktitle={Advances in Neural Information Processing Systems 36},
year={2022}
}