Questions about the experiments shown in Table 4

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)

MIT License

3.89k stars 931 forks source link

Questions about the experiments shown in Table 4 #182

Open jszgz opened 4 years ago

jszgz commented 4 years ago

“Top-1 accuracies on UCF-101 and HMDB-51. Allaccuracies are averaged over three splits.” Did you train on three splits or only train on one split but test on three split? And is it fair to use ReduceLRonPlateau based on losses calculated from testlist01.txt(valition subset in ucf101_1.json) and then test acc on testlist01.txt?

cotton-ahn commented 3 years ago

I am also concerned about this issue. Is it fair to use ReduceLRonPlateau based on the validation loss? actually the "validation" in HMDB51 in here stands for the test dataset.