save_image behavior

[rtqichen]

https://github.com/pytorch/vision/blob/master/torchvision/utils.py#L52

The .mul(0.5).add(0.5) in this line seems to be assuming that the input values are in [-1,1]? Isn't the Torch convention to use [0,1] for images?

[soumith]

you're right. i first did that as a hack and transferred it as-is into vision.
I'll do this:

1. Introduce kwargs called: normalize, range, which if normalize=True -- would directly re-range the input into [0, 255] by per-instance renormalization, and then save it.
2. If normalize=True and range=(min, max), it will use the min, max values to renormalize before saving.
3. if normalize=False, then will save the input image as-is, i.e. with the exact current range.

[rtqichen]

What about something similar to the Lua Torch's image.display() behavior:

• range=None (default) uses the max and min of the input images to renormalize into [0,255].
• range=(min,max) (specified by the user) uses the provided min/max values to clamp then renormalize.

(The first option can be per-instance, but I think using the min/max across all provided images makes more sense.)

The difference here is that by default it's better to be able to handle all ranges instead of assuming a certain range. (I think normalize=False is the intended default value? Otherwise, using normalize=False and range=(0,255) is equivalent anyway.)

[stepelu]

Personally, I vote for having a default behavior that assumes a range [0, 1] and doesn't normalize the image.
Doing otherwise would result in issues like the one mentioned in pull 41 and potentially introduces bugs difficult to spot (as the normalization is "silent").

[soumith]

This is fixed in master by removing the offset, but definitely having a range parameter would be quite useful!