Skip to content

fix torch backend resize with pad_to_aspectio_ratio is set to True #20797

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
merged 9 commits into from
Jan 25, 2025
Merged

fix torch backend resize with pad_to_aspectio_ratio is set to True #20797

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
c767853
fix torch backend resize with `pad_to_aspectio_ratio` is set to `True`
sineeli Jan 22, 2025
c9a8f16
fix axis for single image
sineeli Jan 22, 2025
9921f1d
Fix issue for when running gpu
sineeli Jan 22, 2025
774f6d2
add missing device type
sineeli Jan 23, 2025
97414db
add unit test when pad_to_aspect_ratio set to True
sineeli Jan 24, 2025
ebcaa99
fix numpy backend
sineeli Jan 24, 2025
f7e1573
nit
sineeli Jan 24, 2025
f30e8df
fix api method
sineeli Jan 24, 2025
41ecb92
fix if condition for channels_first
sineeli Jan 24, 2025
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
185 changes: 125 additions & 60 deletions keras/src/backend/numpy/image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -231,72 +231,137 @@ def resize(
pad_width = max(width, pad_width)
img_box_hstart = int(float(pad_height - height) / 2)
img_box_wstart = int(float(pad_width - width) / 2)

if data_format == "channels_last":
if len(images.shape) == 4:
padded_img = (
np.ones(
(
batch_size,
pad_height + height,
pad_width + width,
channels,
),
dtype=images.dtype,
if img_box_hstart > 0:
if len(images.shape) == 4:
padded_img = np.concatenate(
[
np.ones(
(batch_size, img_box_hstart, width, channels),
dtype=images.dtype,
)
* fill_value,
images,
np.ones(
(batch_size, img_box_hstart, width, channels),
dtype=images.dtype,
)
* fill_value,
],
axis=1,
)
* fill_value
)
padded_img[
:,
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
:,
] = images
else:
padded_img = (
np.ones(
(pad_height + height, pad_width + width, channels),
dtype=images.dtype,
else:
padded_img = np.concatenate(
[
np.ones(
(img_box_hstart, width, channels),
dtype=images.dtype,
)
* fill_value,
images,
np.ones(
(img_box_hstart, width, channels),
dtype=images.dtype,
)
* fill_value,
],
axis=0,
)
* fill_value
)
padded_img[
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
:,
] = images
else:
if len(images.shape) == 4:
padded_img = (
np.ones(
(
batch_size,
channels,
pad_height + height,
pad_width + width,
),
dtype=images.dtype,
elif img_box_wstart > 0:
if len(images.shape) == 4:
padded_img = np.concatenate(
[
np.ones(
(batch_size, height, img_box_wstart, channels),
dtype=images.dtype,
)
* fill_value,
images,
np.ones(
(batch_size, height, img_box_wstart, channels),
dtype=images.dtype,
)
* fill_value,
],
axis=2,
)
else:
padded_img = np.concatenate(
[
np.ones(
(height, img_box_wstart, channels),
dtype=images.dtype,
)
* fill_value,
images,
np.ones(
(height, img_box_wstart, channels),
dtype=images.dtype,
)
* fill_value,
],
axis=1,
)
* fill_value
)
padded_img[
:,
:,
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
] = images
else:
padded_img = (
np.ones(
(channels, pad_height + height, pad_width + width),
dtype=images.dtype,
padded_img = images
else:
if img_box_hstart > 0:
if len(images.shape) == 4:
padded_img = np.concatenate(
[
np.ones(
(batch_size, channels, img_box_hstart, width)
)
* fill_value,
images,
np.ones(
(batch_size, channels, img_box_hstart, width)
)
* fill_value,
],
axis=2,
)
* fill_value
)
padded_img[
:,
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
] = images
else:
padded_img = np.concatenate(
[
np.ones((channels, img_box_hstart, width))
* fill_value,
images,
np.ones((channels, img_box_hstart, width))
* fill_value,
],
axis=1,
)
elif img_box_wstart > 0:
if len(images.shape) == 4:
padded_img = np.concatenate(
[
np.ones(
(batch_size, channels, height, img_box_wstart)
)
* fill_value,
images,
np.ones(
(batch_size, channels, height, img_box_wstart)
)
* fill_value,
],
axis=3,
)
else:
padded_img = np.concatenate(
[
np.ones((channels, height, img_box_wstart))
* fill_value,
images,
np.ones((channels, height, img_box_wstart))
* fill_value,
],
axis=2,
)
else:
padded_img = images
images = padded_img

return np.array(
Expand Down
107 changes: 79 additions & 28 deletions keras/src/backend/torch/image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,38 +224,89 @@ def resize(
if len(images.shape) == 4:
batch_size = images.shape[0]
channels = images.shape[1]
padded_img = (
torch.ones(
(
batch_size,
channels,
pad_height + height,
pad_width + width,
),
dtype=images.dtype,
if img_box_hstart > 0:
padded_img = torch.cat(
[
torch.ones(
(batch_size, channels, img_box_hstart, width),
dtype=images.dtype,
device=images.device,
)
* fill_value,
images,
torch.ones(
(batch_size, channels, img_box_hstart, width),
dtype=images.dtype,
device=images.device,
)
* fill_value,
],
axis=2,
)
* fill_value
)
padded_img[
:,
:,
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
] = images
else:
padded_img = images

if img_box_wstart > 0:
padded_img = torch.cat(
[
torch.ones(
(batch_size, channels, height, img_box_wstart),
dtype=images.dtype,
device=images.device,
),
padded_img,
torch.ones(
(batch_size, channels, height, img_box_wstart),
dtype=images.dtype,
device=images.device,
)
* fill_value,
],
axis=3,
)

else:
channels = images.shape[0]
padded_img = (
torch.ones(
(channels, pad_height + height, pad_width + width),
dtype=images.dtype,
if img_box_wstart > 0:
padded_img = torch.cat(
[
torch.ones(
(channels, img_box_hstart, width),
dtype=images.dtype,
device=images.device,
)
* fill_value,
images,
torch.ones(
(channels, img_box_hstart, width),
dtype=images.dtype,
device=images.device,
)
* fill_value,
],
axis=1,
)
else:
padded_img = images
if img_box_wstart > 0:
torch.cat(
[
torch.ones(
(channels, height, img_box_wstart),
dtype=images.dtype,
device=images.device,
)
* fill_value,
padded_img,
torch.ones(
(channels, height, img_box_wstart),
dtype=images.dtype,
device=images.device,
)
* fill_value,
],
axis=2,
)
* fill_value
)
padded_img[
:,
img_box_hstart : img_box_hstart + height,
img_box_wstart : img_box_wstart + width,
] = images
images = padded_img

resized = torchvision.transforms.functional.resize(
Expand Down
25 changes: 25 additions & 0 deletions keras/src/ops/image_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -791,6 +791,31 @@ def test_resize_with_pad(self, fill_value):
)
self.assertEqual(out.shape, (2, 3, 25, 25))

x = np.ones((2, 3, 10, 10)) * 128
out = kimage.resize(
x, size=(4, 4), pad_to_aspect_ratio=True, fill_value=fill_value
)
self.assertEqual(out.shape, (2, 3, 4, 4))
self.assertAllClose(out[:, 0, :, :], np.ones((2, 4, 4)) * 128)

x = np.ones((2, 3, 10, 8)) * 128
out = kimage.resize(
x, size=(4, 4), pad_to_aspect_ratio=True, fill_value=fill_value
)
self.assertEqual(out.shape, (2, 3, 4, 4))
self.assertAllClose(
out,
np.concatenate(
[
np.ones((2, 3, 4, 1)) * 96.25,
np.ones((2, 3, 4, 2)) * 128.0,
np.ones((2, 3, 4, 1)) * 96.25,
],
axis=3,
),
atol=1.0,
)

@parameterized.named_parameters(
named_product(
interpolation=["bilinear", "nearest"],
Expand Down
Loading