def test_predicting_with_mock_longer_width(self):
np.random.seed(1337)
height, width = 4, 6
inp = np.random.random((12, 8, 16, 3))
with tf_test_util.use_gpu():
layer = image_preprocessing.RandomCrop(height, width)
actual_output = layer(inp, training=0)
resized_inp = image_ops.resize_images_v2(inp, size=[4, 8])
expected_output = resized_inp[:, :, 1:7, :]
self.assertAllClose(expected_output, actual_output)
def test_predicting_with_mock_longer_height(self):
np.random.seed(1337)
height, width = 3, 3
inp = np.random.random((12, 10, 6, 3))
with self.cached_session(use_gpu=True):
layer = image_preprocessing.RandomCrop(height, width)
actual_output = layer(inp, training=0)
resized_inp = image_ops.resize_images_v2(inp, size=[5, 3])
expected_output = resized_inp[:, 1:4, :, :]
self.assertAllClose(expected_output, actual_output)
def image_to_input_size(version, images, antialias=False):
"""Resizes images to size `version_input_shape`. This will convert
grayscale to rgb.
Note: Should be run on CPU.
Preprocessing as in: https://github.com/tensorflow/models/blob/1af55e018eebce03fb61bba9959a04672536107d/research/slim/preprocessing/vgg_preprocessing.py#L319
Note: Does not preserve aspect ratio if resize is necessary.
Args:
version: A supported inception version. See `versions()`.
images: 4-D Tensor of shape `[batch, height, width, channels]` or 3-D
Tensor of shape `[height, width, channels]`. Where channels can
be `1` or `3`.
antialias: Whether to use an anti-aliasing filter when downsampling
an image.
Returns:
A tensor of shape `[batch] + version_input_shape` or
`version_input_shape`.
"""
channels = images.shape[-1]
assert channels == 1 or channels == 3
info = version_info(version)
batch_input_shape = info["input_shape"]
input_shape = info["input_shape"][1:]
if not (
images.shape.is_compatible_with(batch_input_shape)
or images.shape.is_compatible_with(input_shape)
):
images = img_ops.resize_images_v2(
images,
size=input_shape[0:2],
preserve_aspect_ratio=False,
antialias=antialias,
)
if channels == 1:
rank = array_ops.rank(images) - 1
tile_shape = array_ops.concat(
[array_ops.ones([rank], dtype=dtypes.int32), [3]], 0
)
images = gen_array_ops.tile(images, tile_shape)
images = math_ops.cast(images, dtype=dtypes.float32)
images -= array_ops.reshape(
constant_op.constant([_R_MEAN, _G_MEAN, _B_MEAN]), [1, 1, 3]
)
return images
def image_to_input_size(version, images, antialias=False):
"""Resizes images to size `version_input_shape`. This will convert
grayscale to rgb.
Note: Should be run on CPU.
Preprocessing as in: https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L253
Args:
version: A supported inception version. See `versions()`.
images: 4-D Tensor of shape `[batch, height, width, channels]` or 3-D
Tensor of shape `[height, width, channels]`. Where channels can
be `1` or `3`.
antialias: Whether to use an anti-aliasing filter when downsampling
an image.
Returns:
A tensor of shape `[batch] + version_input_shape` or
`version_input_shape`.
"""
channels = images.shape[-1]
assert channels == 1 or channels == 3
info = version_info(version)
batch_input_shape = info["input_shape"]
input_shape = info["input_shape"][1:]
images = img_ops.convert_image_dtype(images, dtype=dtypes.float32)
if not (images.shape.is_compatible_with(batch_input_shape)
or images.shape.is_compatible_with(input_shape)):
images = img_ops.resize_images_v2(
images,
size=input_shape[0:2],
preserve_aspect_ratio=False,
antialias=antialias,
)
if channels == 1:
rank = array_ops.rank(images) - 1
tile_shape = array_ops.concat(
[array_ops.ones([rank], dtype=dtypes.int32), [3]], 0)
images = gen_array_ops.tile(images, tile_shape)
images -= 0.5
images *= 2.0
return images
def call(self, inputs):
outputs = image_ops.resize_images_v2(
images=inputs,
size=[self.target_height, self.target_width],
method=self._interpolation_method)
return outputs
