search

fjxmlzn / InfoGAN-CR

[ICML 2020] InfoGAN-CR and ModelCentrality: Self-supervised Model Training and Selection for Disentangling GANs
https://arxiv.org/abs/1906.06034
41 stars 8 forks source link

testing in wild and result size #2

Closed ghost closed 5 years ago

ghost commented 5 years ago

can you provide how to test on your own images for the celeba pretrained model and the results sizes are really small and hard to see the difference currently

fjxmlzn commented 5 years ago

Sorry what do you mean by "test on your own images for the celeba pretrained model"? Do you want to run InfoGAN-CR on your own image dataset, or something else?

And what do you mean by "results sizes"?

ghost commented 5 years ago

@fjxmlzn I mean like testing on a single image like this one rdjtest2

I mean the current result grid is very small for each latent result like 32x32 for each one and you cant see it clearly sample23

fjxmlzn commented 5 years ago

This is because this pre-trained model was trained on 32x32 CelebA. If you want to generate images of higher resolution, just change the training data to whatever you want and retrain the model.

ghost commented 5 years ago

@fjxmlzn ok also you can test on single image instead of whole dataset?

fjxmlzn commented 5 years ago

This InfoGAN-CR is for generating images and is required to train on many images (same as other GANs). Sorry I don't know what you mean by "test on single image". Could you please clarify more on that?

Do you mean changing the factors of a single given image?