wondervictor / retinanet

Standard RetinaNet implemented with Pure PyTorch (Work in progress)

RetinaNet

Work in progress

Standard RetinaNet implemented with Pure PyTorch

Usage

• Compile NMS

cd lib/nms
make

• Modify config.py in cfgs

• Train

python train.py -d 2 -b 4 -e voc_baseline -ds VOC

• Test

python test.py -m log/voc_baseline/model_dump/epochxxx.pth -ds VOC -e voc_baseline

will output the detection results in results.det file.

Output File '*.det'

each line is a json string.

{"im_id":'xxxx', "results":[{"class":xxx, "bbox":[xx,xx,xx,xx], "class":xxxx}]}

Benchmarks

Notice: This part is under construction. Evaluation results and models on COCO,PASCAL VOC will be released soon.

Visualization Results

• PASCAL VOC

Licence

This project is under the MIT Licence

retinanet's Issues

version of pytorch

No matter whether I change 0.3.1 or 0.4.0, there is incompatibility. So which version of the pytorch you are used?

what's the performance on VOC dataset?