def _parse_eval_image(self, decoded_tensors):
"""Parses image data for evaluation."""
image_bytes = decoded_tensors[self._image_field_key]
if self._decode_jpeg_only:
image_shape = tf.image.extract_jpeg_shape(image_bytes)
# Center crops.
image = preprocess_ops.center_crop_image_v2(
image_bytes, image_shape)
else:
# Decodes image.
image = tf.io.decode_image(image_bytes, channels=3)
image.set_shape([None, None, 3])
# Center crops.
image = preprocess_ops.center_crop_image(image)
image = tf.image.resize(image,
self._output_size,
method=tf.image.ResizeMethod.BILINEAR)
image.set_shape([self._output_size[0], self._output_size[1], 3])
# Convert image to self._dtype.
image = tf.image.convert_image_dtype(image, self._dtype)
image = image / 255.0
return image
def _parse_eval_image(self, decoded_tensors):
"""Parses image data for evaluation."""
image_bytes = decoded_tensors[self._image_field_key]
if self._decode_jpeg_only and self._aug_crop:
image_shape = tf.image.extract_jpeg_shape(image_bytes)
# Center crops.
image = preprocess_ops.center_crop_image_v2(
image_bytes, image_shape)
else:
# Decodes image.
image = tf.io.decode_image(image_bytes, channels=3)
image.set_shape([None, None, 3])
# Center crops.
if self._aug_crop:
image = preprocess_ops.center_crop_image(image)
image = tf.image.resize(image,
self._output_size,
method=tf.image.ResizeMethod.BILINEAR)
image.set_shape([self._output_size[0], self._output_size[1], 3])
# Normalizes image with mean and std pixel values.
image = preprocess_ops.normalize_image(image,
offset=MEAN_RGB,
scale=STDDEV_RGB)
# Convert image to self._dtype.
image = tf.image.convert_image_dtype(image, self._dtype)
return image
def _parse_train_image(self, decoded_tensors):
"""Parses image data for training."""
image_bytes = decoded_tensors[self._image_field_key]
if self._decode_jpeg_only:
image_shape = tf.image.extract_jpeg_shape(image_bytes)
# Crops image.
cropped_image = preprocess_ops.random_crop_image_v2(
image_bytes, image_shape)
image = tf.cond(
tf.reduce_all(tf.equal(tf.shape(cropped_image), image_shape)),
lambda: preprocess_ops.center_crop_image_v2(image_bytes, image_shape),
lambda: cropped_image)
else:
# Decodes image.
image = tf.io.decode_image(image_bytes, channels=3)
image.set_shape([None, None, 3])
# Crops image.
cropped_image = preprocess_ops.random_crop_image(image)
image = tf.cond(
tf.reduce_all(tf.equal(tf.shape(cropped_image), tf.shape(image))),
lambda: preprocess_ops.center_crop_image(image),
lambda: cropped_image)
if self._aug_rand_hflip:
image = tf.image.random_flip_left_right(image)
# Color jitter.
if self._color_jitter > 0:
image = preprocess_ops.color_jitter(image, self._color_jitter,
self._color_jitter,
self._color_jitter)
# Resizes image.
image = tf.image.resize(
image, self._output_size, method=tf.image.ResizeMethod.BILINEAR)
image.set_shape([self._output_size[0], self._output_size[1], 3])
# Apply autoaug or randaug.
if self._augmenter is not None:
image = self._augmenter.distort(image)
# Normalizes image with mean and std pixel values.
image = preprocess_ops.normalize_image(image,
offset=MEAN_RGB,
scale=STDDEV_RGB)
# Random erasing after the image has been normalized
if self._random_erasing is not None:
image = self._random_erasing.distort(image)
# Convert image to self._dtype.
image = tf.image.convert_image_dtype(image, self._dtype)
return image
def _build_inputs(self, image):
"""Builds classification model inputs for serving."""
# Center crops and resizes image.
image = preprocess_ops.center_crop_image(image)
image = tf.image.resize(
image, self._input_image_size, method=tf.image.ResizeMethod.BILINEAR)
image = tf.reshape(
image, [self._input_image_size[0], self._input_image_size[1], 3])
# Normalizes image with mean and std pixel values.
image = preprocess_ops.normalize_image(image,
offset=MEAN_RGB,
scale=STDDEV_RGB)
return image
def inference_fn(cls,
image: tf.Tensor,
input_image_size: List[int],
num_channels: int = 3) -> tf.Tensor:
"""Builds image model inputs for serving."""
image = tf.cast(image, dtype=tf.float32)
image = preprocess_ops.center_crop_image(image)
image = tf.image.resize(
image, input_image_size, method=tf.image.ResizeMethod.BILINEAR)
# Normalizes image with mean and std pixel values.
image = preprocess_ops.normalize_image(
image, offset=MEAN_RGB, scale=STDDEV_RGB)
image.set_shape(input_image_size + [num_channels])
return image
def _parse_train_image(self, decoded_tensors):
"""Parses image data for training."""
image_bytes = decoded_tensors[self._image_field_key]
if self._decode_jpeg_only:
image_shape = tf.image.extract_jpeg_shape(image_bytes)
# Crops image.
cropped_image = preprocess_ops.random_crop_image_v2(
image_bytes, image_shape)
image = tf.cond(
tf.reduce_all(tf.equal(tf.shape(cropped_image), image_shape)),
lambda: preprocess_ops.center_crop_image_v2(
image_bytes, image_shape), lambda: cropped_image)
else:
# Decodes image.
image = tf.io.decode_image(image_bytes, channels=3)
image.set_shape([None, None, 3])
# Crops image.
cropped_image = preprocess_ops.random_crop_image(image)
image = tf.cond(
tf.reduce_all(
tf.equal(tf.shape(cropped_image), tf.shape(image))),
lambda: preprocess_ops.center_crop_image(image),
lambda: cropped_image)
if self._aug_rand_hflip:
image = tf.image.random_flip_left_right(image)
# Resizes image.
image = tf.image.resize(image,
self._output_size,
method=tf.image.ResizeMethod.BILINEAR)
image.set_shape([self._output_size[0], self._output_size[1], 3])
# Apply autoaug or randaug.
if self._augmenter is not None:
image = self._augmenter.distort(image)
# Convert image to self._dtype.
image = tf.image.convert_image_dtype(image, self._dtype)
image = image / 255.0
return image
def test_center_crop_image(self, input_height, input_width):
image = tf.convert_to_tensor(
np.random.rand(input_height, input_width, 3))
cropped_image = preprocess_ops.center_crop_image(image)
cropped_image_shape = tf.shape(cropped_image)
self.assertAllEqual([350, 350, 3], cropped_image_shape.numpy())