Skip to content

Latest commit

 

History

History
64 lines (47 loc) · 2.21 KB

install.md

File metadata and controls

64 lines (47 loc) · 2.21 KB

Distilling Object Detectors with Feature Richness

1. Download MMdetection Framework and Dataset

Please first download mmdetection and MS COCO2017 datasets and make sure that you can run a baseline model successfully.

2. Download the Teacher Model

Before starting running the distillation codes, you need to download a pre-trained teacher model. We advise you to download the pretrained faster_rcnn_r101_fpn_2x_coco_bbox_mAP-0.398_20200504_210455-1d2dac9c.pth and retinanet_r101_fpn_1x_coco_20200130-7a93545f.pth for knowledge distillation on two-stage and one-stage students, respectively. Note that the downloading urls of the two models can be found in mmdetection/configs/faster_rcnn/README.md and mmdetection/configs/retinanet/README.md. Then, put them in the checkpoints folder as follows.

mmdetection
--model
----retinanet_r101_fpn_1x_coco_20200130-7a93545f.pth
----faster_rcnn_r101_fpn_2x_coco_bbox_mAP-0.398_20200504_210455-1d2dac9c.pth

3. Change the Codes of MMdetection

  1. move distill_frs_single.py & distill_frs_two.py in mmdetection/mmdet/models/detectors/ and change mmdetection/mmdet/models/detectors/__init__.py
from .distill_frs_single import Distilling_FRS_Single
from .distill_frs_two import Distilling_FRS_Two

# `__all__` add the follows:
__all__ = [
    'Distilling_FRS_Single', 'Distilling_FRS_Two'
]
  1. move adap.py in mmdetection/mmdet/models/necks/ and change mmdetection/mmdet/models/necks/__init__.py
from .adap import ADAP, ADAP_C, ADAP_Residule, ADAP_SINGLE
# `__all__` add the follows:
__all__ = [
    'ADAP', 'ADAP_C', 'ADAP_Residule','ADAP_SINGLE'
]

  1. move distill_frs/ into mmdetection/configs/

  2. move increase_hook.py in mmdetection/mmdet/core/utils/

4. Train model with FRS

export CONFIG_FILE="./configs/distill_frs/resnet50_resnet101_retinanet_frs.py"
export WORK_DIR="work_dirs/retinanet_r50"
export GPU_NUM=8
bash tools/dist_train.sh \
    ${CONFIG_FILE} \
    ${GPU_NUM} \
    --work-dir $WORK_DIR

5. Test model

python tools/test.py \
    configs/retinanet/retinanet_r50_fpn_2x_coco.py \
    work_dirs/retinanet_r50/epoch_12.pth \