Object Detection Knowledge Distillation(ODKD)

I’ve started Hand Mocap Distillation to continue exploring knowledge distillation for hand motion capture.

The function of this branch is not complete. For ssd and yolov5 distillation, checking other branches.

Release edition is coming Soon...

Update

1. The first edition is the refactor of branch mbv2-lite, which is an implementation of Chen, G. et al. (2017) ‘Learning efficient object detection models with knowledge distillation’ with SSD-lite structure.

2. Replace part of code with pytorch api which has same functionality.

3. Very friendly beginner guidance.

4. System Architecture

Useage

$ python setup.py install --user

$ odkd-train ./training_config.yml -t

$ odkd-train training_config.yml
or
$ python -m torch.distributed.launch --nproc_per_node=2 `which odkd-train` training_config.yml

$ odkd-eval ${CHECKPOINTS_PATH}/${RUN_INDEX}/config.yml

TODO

• Evaluation Module

• LOG Module

• Coco dataset support

• Yolov5 distillation