AnchorFreeNets

get start

environment

You should prepare the TensorRT>=7.0 and cuda>=10.0 environment!

pip install -r requirement.txt
# or
# find torch version in https://pytorch.org/get-started/previous-versions/, e.g.
pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install pycocotools==2.0.2 timm==0.4.9 numpy fvcore Cython matplotlib opencv-python tqdm thop torchstat

then, you should run this commond to download the pretrained models!

python down_pretrained_model.py

---

Modify your own config

supported backbone

resnet, swin-transformer, regnet....

supported Net

FCOS, CenterNet, TTFNet

---

Train your own data and Test

Introduction based on CenterNet!

modify the config file

Firstly, modify your dataset file frame

COCO2017
|
|-----annotations----instances_train2017(or val).json
|                 
|-----train2017
|-----val2017

modify the data_path and other necessary params in the config file([config_centernet.py])

base_path = '/your coco frame data path/'

---

Train

cd ./train
./train.sh

---

Test

cd ./tools
python eval.py

---

Inference

cd ./tools
python inference.py

---

Json Analysis

show per class metrics, such as: ap precision recall f1

save FP FN images

cd ./tools
python json_analysis.py

---

Deploy

convert pytorch model to TRT model or RK model

tensorrt: trt_fp16.py pt2trt.py pt2onnx2trt.py

rk: export_trace_model.py convert_rknn.py