bad training results : large green area on face

[sunkyya]

Hi, thanks for your sharing! I used your stargan v2 to train celebA, gantype is "hinge",batchsize is "2",sn is "true". I trained about tow days , the training results are bad and having large green area on face. I doubt that should i continue training or change config?

[xx88xx]

could you upload your results? I am curious to see as I am going to start trying this out soon

[sunkyya]

@xx88xx :

I have fixed this by GHANGING

def inverse_transform(images): return ((images+1.) / 2) * 255.0

TO

def inverse_transform(images): return np.clip(((images+1.) / 2) * 255.0,0,255)

---

The code is in utils.py line 86. During my training process the fake images are out of the range[-1,1], so the saved images have larging strange color areas.

[Jessicall]

@sunkyya Hi~Excuse me.I'm a deep learning newbie and I wonder how to class the celeba into different folders according to the domains? Could you give me a help? Thanks a lot!

[sunkyya]

@Jessicall I'm a newbie too😀 . You can use python to do the class preprocess, read the file "CelebAMask-HQ-attribute-anno.txt" and class them according to the annotations , and then put some/all domains into ''./dataset/celebA/train"

[Jessicall]

@sunkyya Thank you very much!

[Jessicall]

@sunkyya Could you send me a link of downloading the CelebA-HQ.I used original celeba to train the model and the result was not well.And how many images are used for training?

[Jessicall]

@sunkyya Could you send me a link of downloading the CelebA-HQ.I used original celeba to train the model and the result was not well.And how many images are used for training?

I have found the CelebAMask-HQ .And there are 30000 images in total.Have you used 30000 or 202599 images for training?

[sunkyya]

@Jessicall I used 30000 images o(￣▽￣)ブ

[Jessicall]

@sunkyya how do your results look like using the parameters(gantype is "hinge",batchsize is "2",sn is "true") to train the model?I used the same parameters to run the code and "fake_0006750.jpg" as shown below

Should i continue to train?

[anshkapil]

I have the same problem as above. @Jessicall were you able to solve this?

[Johnson-yue]

@anshkapil what is your sty_loss and cyc_loss？？ My sty_loss is so small for all time。。。

[taki0112]

Hello ! I will re-upload the tensorflow2 version of stargan v2 to clova_ai

If you have any questions, use the above link ! Thank you.