Skip to content

Add check before int casting for PIL conversion #21969

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amyeroberts merged 3 commits into from
Mar 7, 2023
Merged

Add check before int casting for PIL conversion #21969

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4241539
Add check before int casting for PIL conversion
amyeroberts Mar 6, 2023
c33f9fd
Line length
amyeroberts Mar 6, 2023
2304eb6
Tidier logic
amyeroberts Mar 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 11 additions & 2 deletions src/transformers/image_transforms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -131,7 +131,8 @@ def to_pil_image(
The image to convert to the `PIL.Image` format.
do_rescale (`bool`, *optional*):
Whether or not to apply the scaling factor (to make pixel values integers between 0 and 255). Will default
to `True` if the image type is a floating type, `False` otherwise.
to `True` if the image type is a floating type and casting to `int` would result in a loss of precision,
and `False` otherwise.

Returns:
`PIL.Image.Image`: The converted image.
Expand All @@ -156,9 +157,17 @@ def to_pil_image(
image = np.squeeze(image, axis=-1) if image.shape[-1] == 1 else image

# PIL.Image can only store uint8 values, so we rescale the image to be between 0 and 255 if needed.
do_rescale = isinstance(image.flat[0], (float, np.float32, np.float64)) if do_rescale is None else do_rescale
if do_rescale is None:
do_rescale = isinstance(image.flat[0], (float, np.floating)) and not np.allclose(image, image.astype(int))

@ydshieh ydshieh Mar 7, 2023 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Personally, I would prefer to:

  • check if image in the range [0, 1]
    • yes: rescale
    • no: check if np.allclose(image, image.astype(int))
      • yes: not rescale
      • no: raise value error (before doing rescale)

But I am OK with this current version - I just feel it's a bit difficult to reason the logic (but not really from this PR)

@amyeroberts amyeroberts Mar 7, 2023

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed - that's tidier. I'll make this change 👍


if do_rescale:
image = rescale(image, 255)

if np.any(image < 0) or np.any(image > 255):
raise ValueError(
"The image to be converted to a PIL image contains values outside the range [0, 255], "
f"got [{image.min()}, {image.max()}] which cannot be converted to uint8."
)
image = image.astype(np.uint8)
return PIL.Image.fromarray(image)

Expand Down
5 changes: 5 additions & 0 deletions tests/test_image_transforms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,6 +96,11 @@ def test_to_pil_image_from_float(self, name, image_shape, dtype):
# make sure image is correctly rescaled
self.assertTrue(np.abs(np.asarray(pil_image)).sum() > 0)

# Make sure that an exception is raised if image is not in [0, 1]
image = np.random.randn(*image_shape).astype(dtype)
with self.assertRaises(ValueError):
to_pil_image(image)

@require_tf
def test_to_pil_image_from_tensorflow(self):
# channels_first
Expand Down