search

moodoki / semantic_image_inpainting

Semantic Image Inpainting
MIT License
171 stars 55 forks source link

DCGAN Training code #1

Closed eldar closed 7 years ago

eldar commented 7 years ago

Hi,

Thanks for sharing your implementation! Can you please point to the DCGAN training code to use with your model?

Cheers, Eldar.

moodoki commented 7 years ago

Hi,

We reused a model trained previously using code from https://github.com/carpedm20/DCGAN-tensorflow. However that repository was written for tensorflow 0.12, and most probably would not work with tensorflow >= 1.0 without modifications.

Tensorflow now has GAN models in its contrib, see https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/gan. What we used was the vannila convolutional GAN with 5 conv layers as in carpedm20's code.

Do remember to change the variable names to match your trained GAN if you are using the newer TFGAN implementation.

If you need help freezing the checkpoints, I have a gist to do that for you here https://gist.github.com/moodoki/e37a85fb0258b045c005ca3db9cbc7f6