hejinIzhanghao

@fsx950223 i did not find squash button for this repo.

I make r1.14 support the model structure of r.1.13 instead of using r1.13, because r1.13 does NOT work well in my multi-gpu machine.

I did not try on voc. it works better than the one of r1.14 on my own dataset.

OK. Once my machines have spare time. And the model is exactly same as the one in r1.13. And the weights file trained by r1.13 can be loaded.

@fsx950223 BTW: backbone of mobilenetv2 + loss of GIOU/MSE worked on my dataset. backbone of darknet53 + loss of MSE worked, but the loss of GIOU did not work with...