is it possible to run this with a more recent version of tensorflow? the resources that I use don't have such an old version of CUDA for tensorflow 1.7

function eval in model.py
I can successfully generate targets, genes and reals by my saved checkpoint model.
However, I want to use only generator to generate the eye with target angle.
That is, input: x_test_r, angles_test_g, output: 64 x 64 x 3 eye image.
So, how to used the restored checkpoint and only use generator to generate image?

Thanks

hi，i am very appreciate for your excellent work，the method of solving minimal enclosed circle has confused me a lot，can your explain it with more detail。

Would you like to offer the pretrain model?

Hi, have you ever test your model with some in-the-wild images like celebA. I think this would enhance the experiment results in your paper as well as this repo.

hello, this project has a good performance handing with yellow-skinned face, but poor performance for whitening face.
As we all know, now many people like to use whitening technology of the phone or camera to process image to whiten the face skin. And I find that the Columbia Gaze Dataset has slightly white-skinned face which is still different the whitening face processed by whitening technology of the phone or camera. So there are some larger gaze dataset including whitening face processed by whitening technology? If no, how should I hand with this whitening face test image to get a good performance?
Anyone can give some advises? @HzDmS

Hello, i can not very clearly understand why should flip the right eye image horizontally as mentioned in the paper and source code.

And the meaning of the file name suffix('L' or 'R') in the dataset, what's the corresponding relation?

Thx~

thanks a lot，recently i train a model with your provided code and data，i find the color of generated eyes is not always consistent with the original input（test on other pic which is not in columbia dataset），when i cancel all instance normalization layer in generator， the color of generated eyes become approximate consitent with the original input？ did you pay attention on this phenomenon.

Hi,

I tried configuring the environment via Anaconda however, quite a few conflicts showed up.

I wonder could you provide a more detailed dependency list, including python version and tensorflow-gpu version as well?

Thanks a lot.

Could the model make the eyes closed?

the pretrained VGG-16 should accept the input image in range [0, 255] minus the mean value of RGB channels [123.6800, 116.7790, 103.9390]. if give the input image in range [-1,1], do that mean give an input image around the mean value ? so the pretrained network can not well extract the feature difference of source image and target image?

while setting the image size to 128 px, the generated images are mostly exactly similar to the input image with almost zero gaze redirection. I tried increasing the loss component for gaze redirection reg_g_loss by a factor of 200, still the actual redirections is almost zero. Any insights on how to get this to work?