training using celeba, testing using custom dataset

[maliho0803]

if I using the model trained by celeba to test my custom images, will it get good results?

[CastellanLiu]

I should remind that this model is trained by the aligned CelebA dataset. The results on your custom images may look weird on attributes such as Mouth_Slightly_Open, Mustache, etc., but should be okay on attributes like Hair_Color, Pale_Skin if the face is the main part of your image. Maybe your can try training the model on the unaligned CelebA dataset.

[maliho0803]

@CastellanLiuthank you for your reply, if I aligned my custom dataset, may it will get better results? and more, if the methods of gan can not work well out of training data?

[LynnHo]

@maliho0803 First, you should align your dataset as the CelebA manner. Second, you should provide accurate attributes of your dataset. It is inconvenient for users to do these and therefore a drawback of this project. I will improve it when I have free time. Besides, I test the model on LFW and it works like on CelebA.

[maliho0803]

@LynnHo when testing, needed images labeled attributes?

[LynnHo]

@LynnHo when testing, needed images labeled attributes?

Yes.

[maliho0803]

@LynnHo how do you labeled LFW datasets?

[LynnHo]

@LynnHo how do you labeled LFW datasets?

I use an attribute predictor trained on CelebA.

[huxianer]

@LynnHo Could you share the predictor trained on CelebA?Thank you for any help!

[CastellanLiu]

Maybe you can find what you want in #11

[LynnHo]

@huxianer Yes, as @CastellanLiu said, you may refer to #11. Or here is an attribute classifier from STGAN which has the same protocol of our AttGAN.

[huxianer]

Thanks!

[Varshul]

hey @maliho0803 , did you try this on your custom dataset? If yes, then were the results satisfying? Also, can you share your the modified files?