Open Alien84 opened 6 years ago
yu4u
commented
6 years ago I am wondering that this low accuracy is normal or there is something wrong?
The low accuracy is normal because it is essentially very difficult task to estimate real or apparent age as a classification task among 101 classes (age 0 - 100; I'm not sure whether the number of classes is 101 in your setting) even if for humans. Thus, you should evaluate your model in terms of MAE or other metrics referring to related work.
Alien84
commented
6 years ago Thank you very much for your email. The number of my classes is 100 (0-99). I haven't evaluated MAE becasue I thought there is an error in my training pipline. However, what is defference between age estimation and image classification in imagenet challenge? In imagenet, there are 1000 classes and the accuracy rate is very higher than the one I get.
So, what will the accuracy be at the final epoch 30? Based on the speed of inccreasing accuracy rate in my pipline, I think it will be something aroound 2, finally. Is that right? To be sure that everything is working fine, do you have any accuracy garph or could you please let me know what accuracy did you get at the last epoch, if possible.
yu4u
commented
6 years ago (copied from e-mail)
However, what is defference between age estimation and image classification in imagenet challenge?
Can you accurately guess the age of a person you met? For example 30; 29 and 31 is wrong. I think it is very difficult and somewhat impossible for real age. It is easier to tell the class in ImageNet as long as you know the class.
I think it will be something aroound 2
around 0.2? I think it is good!
To be sure that everything is working fine, do you have any accuracy garph or could you please let me know what accuracy did you get at the last epoch, if possible.
https://raw.githubusercontent.com/wiki/yu4u/age-gender-estimation/images/accuracy.png
Hi, Thank you very much for sharing this code. I have a simple question. I am training VGG16 for age estimation by using pre-trained model (imagenet). I would be so thankful if you could please kindly have a look at the attached figure. During training, I see the accuracy is very low and it starts from 0.04 and after 5 epoch it reachs around 0.1 but loss dcreases from 8 and after 5 opoch is around 2.5. I checked my code carefully. I am wondering that this low accuracy is normal or there is something wrong? Could you please help me? Thanks.