This repo is implemented based on my dl_lib, some parts of code in my dl_lib is based on detectron2.
Objects as Points is one of my favorite paper in object detection area. However, its code is a little difficult to understand. I believe that CenterNet could get higher pts and implemented in a more elegant way, so I write this repo.
This repo use less training time to get a better performance, it nearly spend half training time and get 1~2 pts higher mAP compared with the old repo. Here is the table of performance.
| Backbone | mAP | FPS | trained model |
|---|---|---|---|
| ResNet-18 | 29.8 | 92 | google drive |
| ResNet-50 | 34.9 | 57 | google drive |
| ResNet-101 | 36.8 | 43 | google drive |
- treat config as a object. You could run your config file and check the config value, which is really helpful for debug.
- Common training / testing scripts in default. you just need to invoke
dl_train/test --num-gpus xin your playground and your projects only need to include all project-specific configs and network modules. - Performance report is dumped automaticly. After your training is over, we will evaluate your model automatically and generate a markdown file.
- Vectorize some operations. This improves the speed and efficiency.
- Support DLA backbone
- Support Hourglass backbone
- Support KeyPoints dataset
- Python >= 3.6
- PyTorch >= 1.3
- torchvision that matches the PyTorch installation.
- OpenCV
- pycocotools
pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'- GCC >= 4.9
gcc --versionMake sure that your get at least one gpu when compiled. Run:
pip install -e .For example, if you want to train CenterNet with resnet-18 backbone, run:
cd playground/centernet.res18.coco.512size
dl_train --num-gpus 8After training process, a README.md file will be generated automatically and this file will report your model's performance.
NOTE: For ResNet-18 and ResNet-50 backbone, we suppose your machine has over 150GB Memory for training. If your memory is not enough, please change NUM_WORKER (in config.py) to a smaller value.
dl_test --num-gpus 8 test downloaded model:
dl_test --num-gpus 8 MODEL.WEIGHTS path/to/your/save_dir/ckpt.pth please refer to google python coding style
I am an intern@megvii now, if you are interested in this company/want to get hired/want to be a intern, please send email to [email protected] or [email protected].