search

junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch
Other
22.8k stars 6.29k forks source link

Input images are stretched when using test #1644

Open KreesoAuriga opened 5 months ago

KreesoAuriga commented 5 months ago

I trained a model on my own data set of 22,000 256x256 jpg image pairs. But when I use test to generate B images from A images that are also 256x256, the results are only the left 128x256 pixels of the image stretched out to 256x256. During training, the output in visdom looked correct, no stretching. I cannot figure out what would cause this. I've not added any extra switch values.

My command for test is python test.py --dataroot ./datasets/biosphereNormalMaps --name biosphereNormalMaps --model pix2pix --direction AtoB

I've tried making the test images jpg and png with the same result.

KreesoAuriga commented 5 months ago

Oh I see what it is now. The test script assumes that the images provided have an A domain image on the left and a B domain image on the right. So when I give it a 256x256 image that is just an A image, it stretches that image to 512x256 and treats the left half as A and the right half as B. Is there a command line option, or a dedicated script, for generating images from data that is just a single A or B image?