karolmajek/Axial-DeepLab-SWideRNet

1# coding=utf-82# Copyright 2021 The Deeplab2 Authors.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16"""Tests for input_preprocessing."""17 18import numpy as np19import tensorflow as tf20 21from deeplab2.data.preprocessing import input_preprocessing22 23 24class InputPreprocessingTest(tf.test.TestCase):25 26  def setUp(self):27    super().setUp()28    self._image = tf.convert_to_tensor(np.random.randint(256, size=[33, 33, 3]))29    self._label = tf.convert_to_tensor(np.random.randint(19, size=[33, 33, 1]))30 31  def test_cropping(self):32    crop_height = np.random.randint(33)33    crop_width = np.random.randint(33)34 35    original_image, processed_image, processed_label, prev_image, prev_label = (36        input_preprocessing.preprocess_image_and_label(37            image=self._image,38            label=self._label,39            prev_image=tf.identity(self._image),40            prev_label=tf.identity(self._label),41            crop_height=crop_height,42            crop_width=crop_width,43            ignore_label=255))44 45    self.assertListEqual(original_image.shape.as_list(),46                         [33, 33, 3])47    self.assertListEqual(processed_image.shape.as_list(),48                         [crop_height, crop_width, 3])49    self.assertListEqual(processed_label.shape.as_list(),50                         [crop_height, crop_width, 1])51    np.testing.assert_equal(processed_image.numpy(), prev_image.numpy())52    np.testing.assert_equal(processed_label.numpy(), prev_label.numpy())53 54  def test_resizing(self):55    height, width = 65, 6556 57    original_image, processed_image, processed_label, prev_image, prev_label = (58        input_preprocessing.preprocess_image_and_label(59            image=self._image,60            label=self._label,61            prev_image=tf.identity(self._image),62            prev_label=tf.identity(self._label),63            crop_height=height,64            crop_width=width,65            min_resize_value=65,66            max_resize_value=65,67            resize_factor=32,68            ignore_label=255))69 70    self.assertListEqual(original_image.shape.as_list(),71                         [height, width, 3])72    self.assertListEqual(processed_image.shape.as_list(),73                         [height, width, 3])74    self.assertListEqual(processed_label.shape.as_list(),75                         [height, width, 1])76    np.testing.assert_equal(processed_image.numpy(), prev_image.numpy())77    np.testing.assert_equal(processed_label.numpy(), prev_label.numpy())78 79  def test_scaling(self):80    height, width = 65, 6581 82    original_image, processed_image, processed_label, prev_image, prev_label = (83        input_preprocessing.preprocess_image_and_label(84            image=self._image,85            label=self._label,86            prev_image=tf.identity(self._image),87            prev_label=tf.identity(self._label),88            crop_height=height,89            crop_width=width,90            min_scale_factor=0.5,91            max_scale_factor=2.0,92            ignore_label=255))93 94    self.assertListEqual(original_image.shape.as_list(),95                         [33, 33, 3])96    self.assertListEqual(processed_image.shape.as_list(),97                         [height, width, 3])98    self.assertListEqual(processed_label.shape.as_list(),99                         [height, width, 1])100    np.testing.assert_equal(processed_image.numpy(), prev_image.numpy())101    np.testing.assert_equal(processed_label.numpy(), prev_label.numpy())102 103  def test_return_padded_image_and_label(self):104    image = np.dstack([[[5, 6], [9, 0]], [[4, 3], [3, 5]], [[7, 8], [1, 2]]])105    image = tf.convert_to_tensor(image, dtype=tf.float32)106    label = np.array([[[1], [2]], [[3], [4]]])107    expected_image = np.dstack([[[127.5, 127.5, 127.5, 127.5, 127.5],108                                 [127.5, 127.5, 127.5, 127.5, 127.5],109                                 [127.5, 5, 6, 127.5, 127.5],110                                 [127.5, 9, 0, 127.5, 127.5],111                                 [127.5, 127.5, 127.5, 127.5, 127.5]],112                                [[127.5, 127.5, 127.5, 127.5, 127.5],113                                 [127.5, 127.5, 127.5, 127.5, 127.5],114                                 [127.5, 4, 3, 127.5, 127.5],115                                 [127.5, 3, 5, 127.5, 127.5],116                                 [127.5, 127.5, 127.5, 127.5, 127.5]],117                                [[127.5, 127.5, 127.5, 127.5, 127.5],118                                 [127.5, 127.5, 127.5, 127.5, 127.5],119                                 [127.5, 7, 8, 127.5, 127.5],120                                 [127.5, 1, 2, 127.5, 127.5],121                                 [127.5, 127.5, 127.5, 127.5, 127.5]]])122    expected_label = np.array([[[255], [255], [255], [255], [255]],123                               [[255], [255], [255], [255], [255]],124                               [[255], [1], [2], [255], [255]],125                               [[255], [3], [4], [255], [255]],126                               [[255], [255], [255], [255], [255]]])127 128    padded_image, padded_label = input_preprocessing._pad_image_and_label(129        image, label, 2, 1, 5, 5, 255)130    np.testing.assert_allclose(padded_image.numpy(), expected_image)131    np.testing.assert_allclose(padded_label.numpy(), expected_label)132 133  def test_return_original_image_when_target_size_is_equal_to_image_size(self):134    height, width, _ = tf.shape(self._image)135    padded_image, _ = input_preprocessing._pad_image_and_label(136        self._image, None, 0, 0, height, width)137    np.testing.assert_allclose(padded_image.numpy(), self._image)138 139  def test_die_on_target_size_greater_than_image_size(self):140    height, width, _ = tf.shape(self._image)141    with self.assertRaises(tf.errors.InvalidArgumentError):142      _ = input_preprocessing._pad_image_and_label(self._image, None, 0, 0,143                                                   height, width - 1)144 145    with self.assertRaises(tf.errors.InvalidArgumentError):146      _ = input_preprocessing._pad_image_and_label(self._image, None, 0, 0,147                                                   height - 1, width)148 149  def test_die_if_target_size_not_possible_with_given_offset(self):150    height, width, _ = tf.shape(self._image)151    with self.assertRaises(tf.errors.InvalidArgumentError):152      _ = input_preprocessing._pad_image_and_label(self._image, None, 3, 3,153                                                   height + 2, width + 2)154 155  def test_set_min_resize_value_only_during_training(self):156    crop_height = np.random.randint(33)157    crop_width = np.random.randint(33)158 159    _, processed_image, _, _, _ = (160        input_preprocessing.preprocess_image_and_label(161            image=self._image,162            label=self._label,163            crop_height=crop_height,164            crop_width=crop_width,165            min_resize_value=[10],166            max_resize_value=None,167            ignore_label=255))168 169    self.assertListEqual(processed_image.shape.as_list(),170                         [crop_height, crop_width, 3])171 172 173if __name__ == '__main__':174  tf.test.main()175