Skip to content

Files

Latest commit

00eec3d · Jun 22, 2022

History

History

detection

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
May 12, 2022
May 12, 2022
May 12, 2022
Jun 22, 2022
May 12, 2022
May 12, 2022
May 12, 2022
May 12, 2022
May 12, 2022
May 12, 2022
May 12, 2022

README.md

Detection codebase for FAN models

Our detection code is developed on top of MMDetection v2.11.0.

Dependencies

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

Data preparation

Prepare COCO according to the guidelines in MMDetection v2.11.0.

Training

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

COCO-C Dataset Generation

To generate COCO-C dataset:

python3 tools/gen_coco_c.py

Test robustness on COCO-C

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