This code is for the paper http://researchers.lille.inria.fr/~valko/hp/publications/oyallon2018compressing.pdf
Please cite the following if you use this code in your paper @InProceedings{Oyallon_2018_ECCV, author = {Oyallon, Edouard and Belilovsky, Eugene and Zagoruyko, Sergey and Valko, Michal}, title = {Compressing the Input for CNNs with the First-Order Scattering Transform}, booktitle = {The European Conference on Computer Vision (ECCV)}, month = {September}, year = {2018} }
Installing pyscatlight:
pip install -r requirements.txt
python setup.py install
A typical call for classification looks like:
python main_scatMultiGPU.py PATHIMAGENET --arch scat50
It is also possible to employ main_scatDistributed.py which requires NCCL.
For detection, a pretrained model can be found here: https://s3.amazonaws.com/modelzoo-networks/scatresnet50.tar.gz
A typical call for detection looks like:
python trainval_net_scat.py --dataset pascal_voc --net scatnet --cuda --save_dir FOLDER --nw 4 --lr 1e-2 --lr_decay_step 6 --bs 3 --epochs 10 --s 3 --mGPUs
Our detection pipeline relies on https://github.com/jwyang/faster-rcnn.pytorch . It was successfully re-tested for pascal and coco using pytorch 0.4.0, with python 3.5 . For the sake of reproducibility, the files that were modified or created at time of submission were:
detection/lib/roi_data_layer/roibatchLoader.py
detection/trainval_net_scat.py
detection/test_net_scat.py
detection/lib/model/utils/blob.py - has to be modified for COCO
detection/cfgs/scatnet.yml
detection/cfgs/scatnet_ls.yml
detection/lib/model/faster_rcnn/scatnet.py