sigmoid + special bias initialization version training code has been
released.
End2end testing: mAP(0.6792)
CUDA_VISIBLE_DEVICES=0 python ./tools/test_net.py --gpu 0 --weights output/sigmoid_RetinaNet_end2end/voc_0712_trainval/sigmoid_RetinaNet_iter_230000.ckpt --imdb voc_0712_test --cfg ./experiments/cfgs/sigmoid_RetinaNet_end2end.yml --network sigmoid_RetinaNet_train_test
end2end training:
nohup ./experiments/scripts/sigmoid_RetinaNet_end2end.sh 0 sigmoid_RetinaNet pascal_voc0712 --set RNG_SEED 42 TRAIN.SCALES "[600]" > sigmoid_RetinaNet.log 2>&1 &
softmax + gradient clipping version
end2end testing:
mAP(0.6813)
python ./tools/test_net.py --gpu 0 --weights output/RetinaNet_end2end/voc_0712_trainval/FPN_iter_140000.ckpt --imdb voc_0712_test --cfg ./experiments/cfgs/RetinaNet_end2end.yml --network RetinaNet_train_test
end2end training:
nohup ./experiments/scripts/RetinaNet_end2end.sh 0 RetinaNet pascal_voc0712 --set RNG_SEED 42 TRAIN.SCALES "[600]" > RetinaNet.log 2>&1 &
tail -f RetinaNet.log
TODO:
top-down and lateral connections from P7 to P5 through P6
which the paper has not mentioned.