def testResizeUp(self, dtype, image_shape, target_shape, method):
data = [64, 32, 32, 64, 50, 100]
expected_data = {}
expected_data["linear"] = [
64.0, 56.0, 40.0, 32.0, 56.0, 52.0, 44.0, 40.0, 40.0, 44.0, 52.0,
56.0, 36.5, 45.625, 63.875, 73.0, 45.5, 56.875, 79.625, 91.0, 50.0,
62.5, 87.5, 100.0
]
expected_data["lanczos3"] = [
75.8294, 59.6281, 38.4313, 22.23, 60.6851, 52.0037, 40.6454,
31.964, 35.8344, 41.0779, 47.9383, 53.1818, 24.6968, 43.0769,
67.1244, 85.5045, 35.7939, 56.4713, 83.5243, 104.2017, 44.8138,
65.1949, 91.8603, 112.2413
]
expected_data["lanczos5"] = [
77.5699, 60.0223, 40.6694, 23.1219, 61.8253, 51.2369, 39.5593,
28.9709, 35.7438, 40.8875, 46.5604, 51.7041, 21.5942, 43.5299,
67.7223, 89.658, 32.1213, 56.784, 83.984, 108.6467, 44.5802,
66.183, 90.0082, 111.6109
]
expected_data["cubic"] = [
70.1453, 59.0252, 36.9748, 25.8547, 59.3195, 53.3386, 41.4789,
35.4981, 36.383, 41.285, 51.0051, 55.9071, 30.2232, 42.151,
65.8032, 77.731, 41.6492, 55.823, 83.9288, 98.1026, 47.0363,
62.2744, 92.4903, 107.7284
]
x = np.array(data, dtype=dtype).reshape(image_shape)
output = image.resize(x, target_shape, method)
expected = np.array(expected_data[method],
dtype=dtype).reshape(target_shape)
self.assertAllClose(output, expected, atol=1e-04)
def onnx_upsample(x, scales: float, mode='nearest'):
sizes = jnp.asarray(x.shape) * scales
sizes = sizes.astype(jnp.int32)
if mode == 'nearest':
method = ResizeMethod.NEAREST
elif mode == 'linear':
method = ResizeMethod.LINEAR
elif mode == 'cubic':
method = ResizeMethod.CUBIC
else:
raise NotImplemented(f"Resize mode: {mode}")
return resize(image=x, shape=sizes, method=method, antialias=True)
def upsample(image):
b, h, w, c = image.shape
upsampled_image = ji.resize(image, [b, 2 * h, 2 * w, c], 'nearest')
return upsampled_image
def resize(img, shape):
n, _, _, c = img.shape
return image.resize(img, (n, ) + shape + (c, ), 'nearest')
def __call__(self, img):
return image.resize(img, self.size, "bilinear").reshape(-1)
