diff --git a/python/paddle/tests/test_transforms.py b/python/paddle/tests/test_transforms.py index 978200fd531c5b82df64ca3c1cf3d70741567700..47977bdf5352bb867b69ed648492fb9a060a13c9 100644 --- a/python/paddle/tests/test_transforms.py +++ b/python/paddle/tests/test_transforms.py @@ -444,6 +444,16 @@ class TestFunctional(unittest.TestCase): os.remove(path) + def test_rotate(self): + np_img = (np.random.rand(28, 28, 3) * 255).astype('uint8') + pil_img = Image.fromarray(np_img).convert('RGB') + + rotated_np_img = F.rotate(np_img, 80, expand=True) + rotated_pil_img = F.rotate(pil_img, 80, expand=True) + + np.testing.assert_equal(rotated_np_img.shape, + np.array(rotated_pil_img).shape) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/vision/transforms/functional.py b/python/paddle/vision/transforms/functional.py index 576415d54302b595254b6c82533194080ee93fd1..da90e4907e410a1a8587f33812c515f5106526fd 100644 --- a/python/paddle/vision/transforms/functional.py +++ b/python/paddle/vision/transforms/functional.py @@ -512,14 +512,19 @@ def adjust_hue(img, hue_factor): return F_cv2.adjust_hue(img, hue_factor) -def rotate(img, angle, resample=False, expand=False, center=None, fill=0): +def rotate(img, + angle, + interpolation="nearest", + expand=False, + center=None, + fill=0): """Rotates the image by angle. Args: img (PIL.Image|np.array): Image to be rotated. angle (float or int): In degrees degrees counter clockwise order. - resample (int|str, optional): An optional resampling filter. If omitted, or if the + interpolation (str, optional): Interpolation method. If omitted, or if the image has only one channel, it is set to PIL.Image.NEAREST or cv2.INTER_NEAREST according the backend. when use pil backend, support method are as following: - "nearest": Image.NEAREST, @@ -564,9 +569,9 @@ def rotate(img, angle, resample=False, expand=False, center=None, fill=0): format(type(img))) if _is_pil_image(img): - return F_pil.rotate(img, angle, resample, expand, center, fill) + return F_pil.rotate(img, angle, interpolation, expand, center, fill) else: - return F_cv2.rotate(img, angle, resample, expand, center, fill) + return F_cv2.rotate(img, angle, interpolation, expand, center, fill) def to_grayscale(img, num_output_channels=1): diff --git a/python/paddle/vision/transforms/functional_cv2.py b/python/paddle/vision/transforms/functional_cv2.py index 65884f4ee5fe15c2ed71375e0cd4a2cbdd9e1369..d50ba7b23c74a501a65a4004dc745c0f4845954e 100644 --- a/python/paddle/vision/transforms/functional_cv2.py +++ b/python/paddle/vision/transforms/functional_cv2.py @@ -15,6 +15,7 @@ from __future__ import division import sys +import math import numbers import warnings import collections @@ -407,13 +408,18 @@ def adjust_hue(img, hue_factor): return cv2.cvtColor(hsv_img, cv2.COLOR_HSV2BGR_FULL).astype(dtype) -def rotate(img, angle, resample=False, expand=False, center=None, fill=0): +def rotate(img, + angle, + interpolation='nearest', + expand=False, + center=None, + fill=0): """Rotates the image by angle. Args: img (np.array): Image to be rotated. angle (float or int): In degrees degrees counter clockwise order. - resample (int|str, optional): An optional resampling filter. If omitted, or if the + interpolation (int|str, optional): Interpolation method. If omitted, or if the image has only one channel, it is set to cv2.INTER_NEAREST. when use cv2 backend, support method are as following: - "nearest": cv2.INTER_NEAREST, @@ -434,15 +440,70 @@ def rotate(img, angle, resample=False, expand=False, center=None, fill=0): """ cv2 = try_import('cv2') + _cv2_interp_from_str = { + 'nearest': cv2.INTER_NEAREST, + 'bilinear': cv2.INTER_LINEAR, + 'area': cv2.INTER_AREA, + 'bicubic': cv2.INTER_CUBIC, + 'lanczos': cv2.INTER_LANCZOS4 + } - rows, cols = img.shape[0:2] + h, w = img.shape[0:2] if center is None: - center = (cols / 2, rows / 2) + center = (w / 2.0, h / 2.0) M = cv2.getRotationMatrix2D(center, angle, 1) + + if expand: + + def transform(x, y, matrix): + (a, b, c, d, e, f) = matrix + return a * x + b * y + c, d * x + e * y + f + + # calculate output size + xx = [] + yy = [] + + angle = -math.radians(angle) + expand_matrix = [ + round(math.cos(angle), 15), + round(math.sin(angle), 15), + 0.0, + round(-math.sin(angle), 15), + round(math.cos(angle), 15), + 0.0, + ] + + post_trans = (0, 0) + expand_matrix[2], expand_matrix[5] = transform( + -center[0] - post_trans[0], -center[1] - post_trans[1], + expand_matrix) + expand_matrix[2] += center[0] + expand_matrix[5] += center[1] + + for x, y in ((0, 0), (w, 0), (w, h), (0, h)): + x, y = transform(x, y, expand_matrix) + xx.append(x) + yy.append(y) + nw = math.ceil(max(xx)) - math.floor(min(xx)) + nh = math.ceil(max(yy)) - math.floor(min(yy)) + + M[0, 2] += (nw - w) * 0.5 + M[1, 2] += (nh - h) * 0.5 + + w, h = int(nw), int(nh) + if len(img.shape) == 3 and img.shape[2] == 1: - return cv2.warpAffine(img, M, (cols, rows))[:, :, np.newaxis] + return cv2.warpAffine( + img, + M, (w, h), + flags=_cv2_interp_from_str[interpolation], + borderValue=fill)[:, :, np.newaxis] else: - return cv2.warpAffine(img, M, (cols, rows)) + return cv2.warpAffine( + img, + M, (w, h), + flags=_cv2_interp_from_str[interpolation], + borderValue=fill) def to_grayscale(img, num_output_channels=1): diff --git a/python/paddle/vision/transforms/functional_pil.py b/python/paddle/vision/transforms/functional_pil.py index 1f06600b999ae5cd6b5800a61ef836c91fb69b02..516c28f849915c3546a40bc4a7e962968ce56b23 100644 --- a/python/paddle/vision/transforms/functional_pil.py +++ b/python/paddle/vision/transforms/functional_pil.py @@ -396,13 +396,18 @@ def adjust_hue(img, hue_factor): return img -def rotate(img, angle, resample=False, expand=False, center=None, fill=0): +def rotate(img, + angle, + interpolation="nearest", + expand=False, + center=None, + fill=0): """Rotates the image by angle. Args: img (PIL.Image): Image to be rotated. angle (float or int): In degrees degrees counter clockwise order. - resample (int|str, optional): An optional resampling filter. If omitted, or if the + interpolation (str, optional): Interpolation method. If omitted, or if the image has only one channel, it is set to PIL.Image.NEAREST . when use pil backend, support method are as following: - "nearest": Image.NEAREST, @@ -426,7 +431,12 @@ def rotate(img, angle, resample=False, expand=False, center=None, fill=0): if isinstance(fill, int): fill = tuple([fill] * 3) - return img.rotate(angle, resample, expand, center, fillcolor=fill) + return img.rotate( + angle, + _pil_interp_from_str[interpolation], + expand, + center, + fillcolor=fill) def to_grayscale(img, num_output_channels=1): diff --git a/python/paddle/vision/transforms/transforms.py b/python/paddle/vision/transforms/transforms.py index 55790d977f131453a5451f60ad30e6148ff901cd..4101c41f2aa30291bb414eeaaf895c5fdedd4562 100644 --- a/python/paddle/vision/transforms/transforms.py +++ b/python/paddle/vision/transforms/transforms.py @@ -1093,8 +1093,7 @@ class RandomRotation(BaseTransform): degrees (sequence or float or int): Range of degrees to select from. If degrees is a number instead of sequence like (min, max), the range of degrees will be (-degrees, +degrees) clockwise order. - interpolation (int|str, optional): Interpolation method. Default: 'bilinear'. - resample (int|str, optional): An optional resampling filter. If omitted, or if the + interpolation (str, optional): Interpolation method. If omitted, or if the image has only one channel, it is set to PIL.Image.NEAREST or cv2.INTER_NEAREST according the backend. when use pil backend, support method are as following: - "nearest": Image.NEAREST, @@ -1131,7 +1130,7 @@ class RandomRotation(BaseTransform): def __init__(self, degrees, - resample=False, + interpolation='nearest', expand=False, center=None, fill=0, @@ -1148,7 +1147,7 @@ class RandomRotation(BaseTransform): self.degrees = degrees super(RandomRotation, self).__init__(keys) - self.resample = resample + self.interpolation = interpolation self.expand = expand self.center = center self.fill = fill @@ -1169,8 +1168,8 @@ class RandomRotation(BaseTransform): angle = self._get_param(self.degrees) - return F.rotate(img, angle, self.resample, self.expand, self.center, - self.fill) + return F.rotate(img, angle, self.interpolation, self.expand, + self.center, self.fill) class Grayscale(BaseTransform):