when fine-tune the pretrained model from kinetics on ucf101 or hmdb51, the accuracy is lower than it in your paper about 7, 8 percent?

[huangwei1995]

[huangwei1995]

The pretrained model is downloaded from yours, like 'resnet-18-kinetics.pth', 'resnext-101-kinetics.pth'. I set the scales as [1.0, 0.841, 0,7,7, 0,595, 0.5], the cropping methods as 'corner', sample duration as 16, batch size as 64

[huangwei1995]

the learning rate is from 0.001 to 0.000001

[Cong222]

i met the same problem with you

[slighting666]

@huangwei1995 @Cong222 Hi, I got the accuracy of val.log, but I want to calculate the accuracy of video level, where can I get the val.json file and how to use eval_ucf101.py?

[YTHmamba]

@huangwei1995 hello ,Have you solved it? How to obtain the accuracy in the paper