search

yu4u / age-gender-estimation

Keras implementation of a CNN network for age and gender estimation
MIT License
1.46k stars 502 forks source link

mix up strategy decreased the trainning samples by 50% #106

Open gandad opened 4 years ago

gandad commented 4 years ago

Hello yu4u, It seems that the mixup strategy decreases the trainning smaple number by 50%. Have you tested the result of keeping the sample number? e.g., for each min-batch,we create a new min-batch using the following fumula: X=shuffled_X1beta+shuffled_X2(1-beta)

Another question, how about creating as many sample as we can, using different beta?

yu4u commented 4 years ago

X=shuffled_X1beta+shuffled_X2(1-beta)

Yes, such implementation seems to be more popular than that of this repo.

Another question, how about creating as many sample as we can, using different beta?

Beta? Beta here means beta distribution, and real mixing coefficients are all different as you want, while they are drawn from the same beta distribution.