bigbigdinosaur

i compile and train in the same conda env my torch version is 1.0

this one-shot method can find objects which are not in training set, so i think there is no need for so-called custom dataset