A Comprehensive Open-Source Toolkit for Neural Image Compression and Robustness Analysis. Please refer to the paper for more details:
- Docker 23+ with NVIDIA Container Toolkit
- 40 Gb of GPU RAM
- We provide Dockerfiles with complete list of requirements
Visit GitLab repository to find out more about GitLab CI/CD launch.
# Download source code
git pull https://github.com/msu-video-group/NIC-RobustBench
cd NIC-RobustBench
# build docker images
# if you change image names, please change DOCKER_IMAGE and DOCKER_IMAGE_JPEGAI variables accordingly in launch.sh script
sudo docker build -f main.Dockerfile -t codecs_main .
sudo docker build -f jpegai.Dockerfile -t codecs_jpegai .
# note that building can take up to 20-30 minutes
# run attack
chmod +x launch.sh
./launch.sh {attack_preset} {loss_name} {attack_name} {codec_name} {gpu_id}
# example of launching
./launch.sh 0 bpp_increase_loss random-noise jpegai-v51-hop-b05 0
# to launch a set of attacks/codecs you can edit launch_all.sh script
# configure launch_all.sh to run desired attacks/codecs
chmod +x launch.sh launch_all.sh
./launch_all.shBe aware that the first time script is launching it will download all necessary weights of about 10 Gb. To download them manually run in the root of directory:
wget --backups=1 -nv https://titan.gml-team.ru:5003/fsdownload/o3kmmUJdU/models.zip \
https://titan.gml-team.ru:5003/fsdownload/o3kmmUJdU/models.zip && rm models.zip.1We include one image from each of BSDS, NIPS2017, Kodak PhotoCD and CityScapes datasets in this repository. You can download these datasets completely via:
wget --backups=1 -nv https://titan.gml-team.ru:5003/fsdownload/Hy8LfJM6e/codec-datasets.zip https://titan.gml-team.ru:5003/fsdownload/Hy8LfJM6e/codec-datasets.zip && codec-datasets.zip.1 You can add any new dataset in the datasets/codec-datasets folder. If you do so, change TEST_DATASET_NAMES and TEST_DATASET_PATHS variables in scripts-docker/env_vars.sh accordingly. Make sure each dataset subfolder contains at least one image.
