RetinaNet

Implementation of RetinaNet (this is converted from maskrcnn-benchmark)

Environment

• Python 3.6
• PyTorch 1.1
• OpenCV

Performance

Model	maskrcnn	converted	train(TODO)
RetinaNet_R-50-FPN_1x	36.3	36.3
RetinaNet_R-101-FPN_1x	38.5	38.5

Demo

$ cd scripts
$ python demo.py [--config-file ../configs/retina_resnet50_v1b_coco.yaml] [--images ../png/biking.jpg]

Eval

$ cd scripts
$ python eval.py [--config-file ../configs/retina_resnet50_v1b_coco.yaml]

Train

$ cd scripts
$ export NGPUS=4
$ python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py [--config-file ../configs/retina_resnet50_v1b_coco.yaml] [--skip-test false]

You can modify training and test setting in configs/xxx.yaml and configs/defaults.py

TODO

• Add freeze backbone
• Add training results