The environment for object detetction has been included in ../environment.yaml. Typically, You do not need to take care of it if you create the environment as specified in ../INSTALL.md. In case there are problems with mmcv or mmdetection, you may uninstall the package and then reinstall it mannually, e.g.
pip uninstall mmcv
pip install --no-cache-dir mmcv==1.7.0- STEP 0: prepare data
$ mkdir data && ln -s /your/path/to/coco data/coco # prepare data- STEP 1: run experiments
$ vim slurm_train.sh # change config file, slurm partition, etc.
$ bash slurm_train.shSee slurm_train.sh for details.
| name | Pretrained Model | Method | Lr Schd | mAP_box | mAP_mask | log | mAP_box* | mAP_mask* | tensorboard log* | config |
|---|---|---|---|---|---|---|---|---|---|---|
| BiFormer-S | IN1k | MaskRCNN | 1x | 47.8 | 43.2 | log | 48.1 | 43.6 | tensorboard.dev | config |
| BiFormer-B | IN1k | MaskRCNN | 1x | 48.6 | 43.7 | log | - | - | - | config |
| BiFormer-S | IN1k | RetinaNet | 1x | 45.9 | - | log | 47.3 | - | tensorboard.dev | config |
| BiFormer-B | IN1k | RetinaNet | 1x | 47.1 | - | log | - | - | - | config |
* : reproduced right before code release.
NOTE: This repository produces significantly better performance than the paper reports, possibly due to
- We fixed a "bug" of extra normalization layers.
- We used a different version of mmcv and mmdetetcion.
- We used native AMP provided by torch instead of Nvidia apex.
We do not know which factors actually work though.
This code is built using mmdetection, timm libraries, and UniFormer repository.