A seed for reproducibility?

[yug125lk]

Thank you again. I have another question about the seed. In your code, did you set a seed to reproduce the results? When I trained the model twice, I got different results.

[leondgarse]

Seldom do I set a seed, and most times the results are near. Did your results differ a lot?

[yug125lk]

I trained twice ResNet34 on the CASIA-WebFace dataset, using Swish activation. On LFW gave me 99.35 and 99.21. I thought it might be an issue related to the seed!!

[leondgarse]

Yes, that difference may be due to the seed. May try to save a randomly initialized model at first and use that for later other training tests.