accuracy on casia training dataset

[auroua]

I am using the sphereface code implemented by pytorch in the third-party re-implementation lists. After 20 epochs, I only get 93.3490% accuracy on training dataset. Does this correct? This model get 98.85% accuracy on lfw.

[jjennings955]

I have the same question.

@auroua is your 93.3490% on withheld validation data from CASIA?

[auroua]

@jjennings955 not on validate dataset. It's on training dataset.

[wy1iu]

I am not sure about the pytorch version, but my guess is that when you compute the accuracy, it uses the difficult decision rule (because m=4 makes the classification more difficult). You will need to use m=1 to compute the accuracy.

[auroua]

@wy1iu Thanks