I am working on training siamfc on pytorch 1.3.1.
The results in paper were reproducing using pytorch 1.0.1 with AUC 0.57, however, the AUC in pytorch 1.3.1 is about 0.52.
I also trained model in pytorch 1.0.1 and load it in pytorch 1.3.1, the results is still bad. I found there is slightly difference in the outputs of backbone on some images.
I have no idea what cause this difference because it is a very simple network，is there any suggestions?