The accuracy is not as good as said in the paper

[ifp-uiuc]

The accuracy is not as good as said in the paper

[pkhorrami4]

Hello. I am sorry the code is not running as it should. I believe I was able to reproduce the problem you are experiencing. I think it has to do with the input images. When the images are resized to 96x96, they are returned as data type float64 instead of uint8 resulting in a range mismatch when they are saved as uint8 to the numpy (.npy) file.

I think if you display an image from the X.npy file, it will probably look like noise rather than a face. I made a correction to the code and posted it to a branch called fix_image_range. Hopefully, this should fix it.

[wmonica]

I have tried as you said. The accuracy is a little higher than before. But the accuracy is still lower then 55%, and I plot the train/test accuracy figure, as shown. I found the training process is overfitting. Could tell me how to solver the problem? Or how I can get the result just as you said. Thanks!

[pkhorrami4]

Could you try and display a few of the images for me? The code should be saving the images to a file named X.npy. Do the images look like faces or like noise?

[pkhorrami4]

I received an email in which you showed that the images were just noise (strangely, I don't see your message here). As I mentioned before, I believe the issue is because the images are being saved out to the .npy files incorrectly.

Just to be clear, did you rerun the make_ck_dataset.py file in the fix_image_range branch to reconstruct the .npy data files before retraining the model? If not, could you please rerun the script, save the .npy files to a temp directory and confirm that the images look like faces. Thanks

[wmonica]

Thank you for your help. I checked the .npy file, and the images are just noise. Then I reconstruct the .npy data, now the images are just like faces. Just as shown below.




And the training process is very smoothly. The test accuracy is very high.

Thank you very much.

[pkhorrami4]

No problem. Thanks for bringing the issue to my attention.