Our detection code is developed on top of MMDetection v2.11.0.
Install according to the guidelines in MMDetection v2.11.0.
or
pip install mmdet==2.13.0 --user
Below is the environment configuration for our codebase:
CUDA Version: 11.1
Torchvision: 0.8.1
Pytorch: 1.7.1
mmcv-full: 1.3.0
timm: 0.5.4
Prepare COCO according to the guidelines in MMDetection v2.11.0.
To train FAN-T-Hybrid + Cascade MRCNN on COCO train2017 on 4 nodes with 32 gpus, run following commands on each node separately:
bash mn_train.sh configs/cascade_mask_fan_tiny_fpn_3x_mstrain_fp16.py 4 $local_rank "master node address"
To train FAN-T-Hybrid on a single node:
bash dist_train.sh configs/cascade_mask_fan_tiny_fpn_3x_mstrain_fp16.py 8
To generate COCO-C dataset:
python3 tools/gen_coco_c.py
To test robustness on a single corruption type:
bash dist_test_coco_c.sh $CONFIG $CKP 8 gaussian_noise 1 --work-dir $WORKDIR
We also provide a script to test robustness over all corruption categories:
bash tools/coco_c_test_all.sh