Distill_GID_detectron2

Paper Link: General Instance Distillation for Object Detection

GID Framework

1. Installation

Step1. Clone repository

https://github.com/daixinghome/Distill_GID_detectron2.git

Step2. Download COCO Dataset

The following code will download COCO2017 to path: dataset/coco.

cd dataset
./download_coco_2017.sh

Step3. Install Pytorch

Example with cuda-11.3 version:

pip install torch==1.11.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu113

Step4. Install Detectron2

Current code is only work for detectron 0.4.x version for compatibility reason. So you should build it from source with the following cmd:

git clone https://github.com/facebookresearch/detectron2.git
git checkout v0.4.1
python -m pip install -e detectron2

2. Train

Every pipeline is defined in the configs directory. You can use the scripts in tools and corresponding config to visualize.

Step1. Set Dataset Path

You should set the dataset directory path for train and evaluation. There are two ways (suppose you have downloaded coco2017 to path dataset/coco):

1. Set to OS ENV:
   export DETECTRON2_DATASETS=<GID_PATH>/dataset

2. Set to every bash file:
   Details in every bash file (4/5th line), such as:
   Distill_GID_detectron2/RetinaNet_Res50/train.sh
Distill_GID_detectron2/RetinaNet_Res50/eval.sh

Step2. Train Student and Teacher

The training cmd format is as following:

# cd RetinaNet_Res50 or RetinaNet_Res101
CUDA_VISIBLE_DEVICES=0,1,... ./train.sh <config_file> <GPU_NUM> [<resume_dir>]

Example with 4 GPUs.

# Training ...
CUDA_VISIBLE_DEVICES=0,1,2,3 ./train.sh config/RetinaNet_2x_multi-scale.yaml 4

# Resume training ...
CUDA_VISIBLE_DEVICES=0,1,2,3 ./train.sh config/RetinaNet_2x_multi-scale.yaml 4 output_2x_multi-scale

Step2. Distillation

The training cmd format is as following:

# cd GID
CUDA_VISIBLE_DEVICES=0,1,... ./train.sh <config_file> <GPU_NUM>

3. Eval

Every pipeline is defined in the configs directory. You can use the scripts in tools and corresponding config to visualize.

Step1. Set Dataset Path

The same with section Train.

Step2. Running

The evaluation cmd format is as following:

CUDA_VISIBLE_DEVICES=0 ./eval.sh <config_file> <model_weight>

---

This code is a reproduce version on detectron2, for the original version is realized in a private internal training framework. Due to some internal equipment resource reasons, we are unable to continue the accuracy alignment work and the code is reference only.