Can I get the 92.5% accuracy by directly run your code?

[MubarkLa]

Hello.
I downloaded your code for the paper 'Two-Stream Convolutional Networks for Action Recognition in Videos', and run it directly. I tried many times but I can only get a result of 91.5%.
Should I change somewhere in your code to get the 92.5% in your paper?

[feichtenhofer]

Hi,
In the paper a different testing strategy is employed. "For testing we average
20 temporal predictions from each network by densely sampling the input-frame-stacks and their horizontal flips". Consider changing the opts.train.numValFrames variable.

[MubarkLa]

OK! Understand!
Thank you very very much.

[lvyali]

@MubarkLa
Sorry to bother you,but can you tell me how to test this code to get the accuray?