Skip to content

Round float values in to_pil_image uint8 conversion #9646

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

Open
YeonwooSung wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Round float values in to_pil_image uint8 conversion #9646

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
31 changes: 22 additions & 9 deletions test/test_transforms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -283,9 +283,9 @@ def test_pil_to_tensor(self, channels):
torch.testing.assert_close(output.numpy(), expected_output)

input_data = torch.as_tensor(np_rng.rand(channels, height, width).astype(np.float32))
img = transforms.ToPILImage()(input_data) # CHW -> HWC and (* 255).byte()
img = transforms.ToPILImage()(input_data) # CHW -> HWC and rint(* 255) to uint8
output = trans(img) # HWC -> CHW
expected_output = (input_data * 255).byte()
expected_output = (input_data * 255).round().clamp(0, 255).to(torch.uint8)
torch.testing.assert_close(output, expected_output)

# separate test for mode '1' PIL images
Expand Down Expand Up @@ -607,7 +607,7 @@ def test_autocontrast_equal_minmax():
class TestToPil:
def _get_1_channel_tensor_various_types():
img_data_float = torch.Tensor(1, 4, 4).uniform_()
expected_output = img_data_float.mul(255).int().float().div(255).numpy()
expected_output = img_data_float.mul(255).round().clamp(0, 255).div(255).numpy()
yield img_data_float, expected_output, "L"

img_data_byte = torch.ByteTensor(1, 4, 4).random_(0, 255)
Expand All @@ -624,7 +624,7 @@ def _get_1_channel_tensor_various_types():

def _get_2d_tensor_various_types():
img_data_float = torch.Tensor(4, 4).uniform_()
expected_output = img_data_float.mul(255).int().float().div(255).numpy()
expected_output = img_data_float.mul(255).round().clamp(0, 255).div(255).numpy()
yield img_data_float, expected_output, "L"

img_data_byte = torch.ByteTensor(4, 4).random_(0, 255)
Expand Down Expand Up @@ -673,7 +673,7 @@ def test_1_channel_ndarray_to_pil_image(self, with_mode, img_data, expected_mode
img = transform(img_data)
assert img.mode == expected_mode
if np.issubdtype(img_data.dtype, np.floating):
img_data = (img_data * 255).astype(np.uint8)
img_data = np.clip(np.rint(img_data * 255.0), 0, 255).astype(np.uint8)
# note: we explicitly convert img's dtype because pytorch doesn't support uint16
# and otherwise assert_close wouldn't be able to construct a tensor from the uint16 array
torch.testing.assert_close(img_data[:, :, 0], np.asarray(img).astype(img_data.dtype))
Expand Down Expand Up @@ -707,7 +707,7 @@ def test_2_channel_ndarray_to_pil_image_error(self):
@pytest.mark.parametrize("expected_mode", [None, "LA"])
def test_2_channel_tensor_to_pil_image(self, expected_mode):
img_data = torch.Tensor(2, 4, 4).uniform_()
expected_output = img_data.mul(255).int().float().div(255)
expected_output = img_data.mul(255).round().clamp(0, 255).div(255)
if expected_mode is None:
img = transforms.ToPILImage()(img_data)
assert img.mode == "LA" # default should assume LA
Expand Down Expand Up @@ -755,13 +755,13 @@ def test_2d_ndarray_to_pil_image(self, with_mode, img_data, expected_mode):
img = transform(img_data)
assert img.mode == expected_mode
if np.issubdtype(img_data.dtype, np.floating):
img_data = (img_data * 255).astype(np.uint8)
img_data = np.clip(np.rint(img_data * 255.0), 0, 255).astype(np.uint8)
np.testing.assert_allclose(img_data, img)

@pytest.mark.parametrize("expected_mode", [None, "RGB", "HSV", "YCbCr"])
def test_3_channel_tensor_to_pil_image(self, expected_mode):
img_data = torch.Tensor(3, 4, 4).uniform_()
expected_output = img_data.mul(255).int().float().div(255)
expected_output = img_data.mul(255).round().clamp(0, 255).div(255)

if expected_mode is None:
img = transforms.ToPILImage()(img_data)
Expand Down Expand Up @@ -819,7 +819,7 @@ def test_3_channel_ndarray_to_pil_image_error(self):
@pytest.mark.parametrize("expected_mode", [None, "RGBA", "CMYK", "RGBX"])
def test_4_channel_tensor_to_pil_image(self, expected_mode):
img_data = torch.Tensor(4, 4, 4).uniform_()
expected_output = img_data.mul(255).int().float().div(255)
expected_output = img_data.mul(255).round().clamp(0, 255).div(255)

if expected_mode is None:
img = transforms.ToPILImage()(img_data)
Expand Down Expand Up @@ -891,6 +891,19 @@ def test_tensor_bad_types_to_pil_image(self):
with pytest.raises(ValueError, match=r"pic should not have > 4 channels. Got \d+ channels."):
transforms.ToPILImage()(torch.ones(6, 4, 4))

@pytest.mark.parametrize("as_tensor", [True, False])
def test_to_pil_image_float_rounds_to_nearest(self, as_tensor):
# Regression for https://github.com/pytorch/vision/issues/8141:
# float values must be rounded to nearest uint8, not truncated.
values = np.array([0.0, 0.001, 0.9999, 1.0], dtype=np.float32)
expected = np.array([[0, 0, 255, 255]], dtype=np.uint8)
if as_tensor:
pic = torch.from_numpy(values).reshape(1, 1, -1)
else:
pic = values.reshape(1, -1, 1)
img = F.to_pil_image(pic)
np.testing.assert_array_equal(np.asarray(img), expected)


def test_adjust_brightness():
x_shape = [2, 2, 3]
Expand Down
2 changes: 1 addition & 1 deletion torchvision/transforms/functional.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,7 +279,7 @@ def to_pil_image(pic, mode=None):
npimg = pic

if np.issubdtype(npimg.dtype, np.floating) and mode != "F":
npimg = (npimg * 255).astype(np.uint8)
npimg = np.clip(np.rint(npimg * 255.0), 0, 255).astype(np.uint8)

if npimg.shape[2] == 1:
expected_mode = None
Expand Down