I am unable to reproduce the results for ir-CSN-152 and ip-CSN-152 pretrained on Sports1M ft Kinetics and IG-65M ft Kinetics. When tested on Kinetics 400 val, the Acc@1 and Acc@5 are around 40% and 65% respectively. Table 5 in the paper indicates 82% and 95%. Are there issues with the Pytorch models?
I am unable to reproduce the results for
ir-CSN-152andip-CSN-152pretrained on Sports1M ft Kinetics and IG-65M ft Kinetics. When tested on Kinetics 400 val, theAcc@1andAcc@5are around 40% and 65% respectively. Table 5 in the paper indicates 82% and 95%. Are there issues with the Pytorch models?