Why do I get washedout images when displaying make_grid output?

1

Hello everyone.
Today I faced something strange. When using torchvision.utils.make_grid(), I noticed when ever I display the resulting image in matplotib, the image is washedout! but when I save it using save_img , it turns out just fine.
This is how it looks when I display the output of make_grid() in matplotlib :

make_grid_img1
make_grid_img12111×899 109 KB

and this is how it got saved to the disk :
imgs_0
and this is the snippet I wrote :

from` torchvision.utils` import save_image, make_grid

fig = plt.figure(figsize=(28, 28))
for i in range(5):
  grid_imgs = make_grid(torch.from_numpy(img_pairs[i]),
                        nrow=5,
                        normalize=True)
  save_image(grid_imgs,f'results/imgs_{i}.jpg')
  ax = fig.add_subplot(5, 1, i+1, xticks=[], yticks=[])
  ax.imshow(grid_imgs.numpy().transpose(1,2,0),cmap='Greys_r')

normalizing and not doesnt affect anything.
What am I missing here ?
Your kind help is greatly appreciated

1 Like
2

I think removing ax = … and using plt.imshow(…) instead of ax.imshow(…) could work.

3

Thanks thats not the case, since I have several images to display I needed as many axes so I need to use that.

Thanks to dear God, I found the cause!
This was caused by save_image() method! I moved the save_image() after showing the image and all was fine!
This seems like a bug to me. this should not happen at all! @smth

4

Thanks for raising this bug!
I’ve created an issue here to track and fix it.

1 Like
5

Fixed here. Should be available in the nightly builds.

1 Like