Because detection-models (Yolo, Mask-RCNN, etc) are often developed by the different frameworks and evaluated the performance on different hardware, it's hard to evaluate the performance (i.e, fps) and fine-tune these on your customized datasets. This project builds all detection-models with pytorch and provides the general API for training, fine-tuning and detection in supervised and unsupervised settings.

1. Pytorch-1.0
2. python-3.6
3. CUDA-9.0
4. 2 TiTAN XP with 12 GPU Memory

1. Supervised Fine-tuning
2. Unsupervised Fine-tuning

1. Installation
2. Supporting Models
3. Loss visualization and analysis
4. Fine-tuning
5. Usage
   1. Mask-RCNN (in progress)
   2. RetinaNet (in progress)
   3. Yolov3
   4. M2Det (in progress)
6. References

$ git clone https://github.com/jacksonly/Detection-Fine-tuning-API.git
$ cd Detection-Fine-tuning-API/
$ sudo pip3 install -r requirements.txt

$ cd weights/
$ bash ./yolov3/weights/download_weights.sh

Pretrained pytorch models (.pt):

1. yolov3.pt, yolov3-tiny.pt, yolov3-spp.pt

• Two-stage Detection:
  1. Mask RCNN (CVPR'17) inspired by wkentaro
• One-stage Detection:
  1. RetinaNet (ICCV'17) inspired by yhenon
  2. Yolov3 (arXiv'18) inspired by ultralytics and TencentYoutuResearch
  3. M2Det (AAAI'19, the latest SSD) inspired by qijiezhao