Why the model trained by this version is stabler than caffe2 version?what's different between this version and caffe2?

[Rosesor]

I had trained masktext spotter model with my own dataset(1000 pics) by pytorch and caffe2 version. I found pytorch model is stabler, acc is higher, and it has better perform on avoiding overfitting. You have done a good job! Im interesting in what you have modify on this version, thx~

[MhLiao]

@Rosesor The caffe2 version is an implementation of the ECCV paper while this repo is an implementation of the TPAMI paper.