search

feidfoe / learning-not-to-learn

[CVPR2019]Learning Not to Learn : An adversarial method to train deep neural networks with biased data
112 stars 13 forks source link

Number of classes for age classification #7

Open erobic opened 4 years ago

erobic commented 4 years ago

How many output units do you use for age classification? Is each age a separate class? Or is it a binary problem 0-29 vs 40+?

zwvews commented 4 years ago

Hi, I think they should use the binary label since I found that the number of testing data stated by their paper is 13129. This is exactly the sum of data with age <=29 and >40

erobic commented 4 years ago

Hi,

I asked the authors about this and they replied that they use 12 nodes for age classification with intervals from https://arxiv.org/abs/1809.02169. Unfortunately, I wasn't able to replicate the numbers with that setting though. Even the baseline numbers were off, so I asked for hyperparameters/other settings, but haven't heard back since then.

Also, treating this as binary classification problem would lose the distinction between age classification vs gender classification, so I think # of classes has to be different than 2.

carloalbertobarbano commented 4 years ago

I agree with @erobic. @feidfoe could you provide us with more info? thanks