yjxiong / tsn-pytorch

Temporal Segment Networks (TSN) in PyTorch
BSD 2-Clause "Simplified" License
1.07k stars 308 forks source link

test_accuracy on hmdb split1 is low #70

Closed Zhuysheng closed 6 years ago

Zhuysheng commented 6 years ago

when train on hmdb split1,the val accuracy is about 70%, but test accuracy is only about 60%。i use 2 segments and each segment sample 16 frames sequentially in training.when in testing,i sample 25 frames at equal intervals。 Is there a problem with the way I sample?

yjxiong commented 6 years ago

each segment sample 16 frames I don't think this is the way we use TSN in the paper. For your own experiments, you are responsible for tuning the learning parameters.