Questions about the results about visual-feature-attribution-using-wasserstein-gans-pytorch HOT 2 CLOSED

For what concerns the train with ADNI data you can check the official repo https://github.com/baumgach/vagan-code/ and write a pr, as it hasn't been implemented yet.

About the images instead, I don't understand what are you asking for, (you also missed to post the original image), but that's how image are saved:

As you said, the sum samples are the fake images (anomaly map, or G(x)) + orig image (real samples, x), the discriminator distinguishes that sum from original, non-anomalous images (noise without white sqaures). One of the trouble you can get is that the fake image that get saved is the inverted map plt.imsave(path, -img, cmap='gray') so what you see white is in reality black. But I can assure you there is not such thing as images added w/o certain areas.

Finally, I suggest you to pull/reclone the repo and re-run the code, these are the images I get after 600 epochs:

Thank you, I'm closing the issue, feel free to reopen it and/or add comments.

from visual-feature-attribution-using-wasserstein-gans-pytorch.

my fake image

I have trained fake image to generate non-rectangular part, is it wrong?

What distinguishes good training standards?

thank you very much!

from visual-feature-attribution-using-wasserstein-gans-pytorch.