Normalization of dataset

[f-amerehi]

Hi @GeorgeCazenavette

Hope that all is well. I'm trying to load images via images = torch.load("imagefruit/images_best.pt"). However, when I try to plot them, I get the error of

WARNING:matplotlib.image:Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

How can I normalize the data? I tried minmax normalization but the outputs are opaque and not similar to the ones that you have in the here.

Thanks

[GeorgeCazenavette]

Hello :)

For purely visualization purposes, the images are clipped to mean +- 2.5 * std before normalizing to [0, 1]. There is an example of this in the code where the images are logged.

If you naively normalize to [0,1], the extreme pixel values in the distilled images will result in low overall visual contrast.

Hope this helps! If not, could you share what your images look like?

[f-amerehi]

Thank you @GeorgeCazenavette very much for quick reply! Here you are these are the images

The code is also here:

images = torch.load("imagefruit/images_best.pt")
labels = torch.load("imagefruit/labels_best.pt")

v = images
v_min = torch.min(v)
v_max = torch.max(v)
new_max = 1
new_min=0

v_p = (v - v_min)/(v_max - v_min)*(new_max - new_min) + new_min

col = 5
rows = math.ceil(len(v_p)/col)
fig,axs = plt.subplots(rows,col,figsize=(col*2,rows*2))

for (i,ax) in enumerate(axs.flatten()):
  image = v_p[i]
  pic = image.numpy().transpose((1,2,0))

  ax.imshow(pic)
  ax.axis('off')

plt.tight_layout()
plt.show()

[GeorgeCazenavette]

Yeah your images look correct, they'll just be prettier if you clip them before normalizing.

Check the section at line 256 of distill.py (I would link it directly, but I can't figure out how to do that from my phone.)

[f-amerehi]

Thank you @GeorgeCazenavette so much!