zlszhonglongshen / centernet-darknet

Darknet implementation of CenterNet

CenterNet-darknet

A Darknet implementation of CenterNet, which support training and inference.

Usage

1. clone CenterNet-darknet repo

git clone https://github.com/CaoWGG/CenterNet-darknet.git

2. convert dla34v0(no deform conv) model to darknet.

cd torch2darknet/
python3 torch2darknet.py
cp dla34_ctdet.cfg ../darknet/train_config/
cp dla34_ctdet.weights ../darknet/weights/pretrain

Convtranspose2d is not supported. convert deconv layer to upsample layer.

3. prepare data

cd darknet/scripts/
bash get_coco_dataset.sh

and then update darknet/train_config/coco.data

4. train

cd darknet/
make -j4
./darknet detector train train_config/coco.data train_config/dla34_ctdet.cfg weights/pretrain/dla34_ctdet.weights -gpus 0,1

5. test

./darknet detect train_config/coco.data train_config/dla34_ctdet.cfg.test weights/backup/dla34_ctdet.backup 5k.txt -i 0

batch=1 and subdivisions=1 in dla34_ctdet.cfg.test

Notice

(xy and wh) loss is l1loss, hm loss is focalloss.

GPU is supported only.