karolmajek/Axial-DeepLab-SWideRNet
0
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 preprocess_utils."""17import numpy as np18import tensorflow as tf19 20from deeplab2.data.preprocessing import preprocess_utils21 22 23class PreprocessUtilsTest(tf.test.TestCase):24 25 def testNoFlipWhenProbIsZero(self):26 numpy_image = np.dstack([[[5., 6.],27 [9., 0.]],28 [[4., 3.],29 [3., 5.]]])30 image = tf.convert_to_tensor(numpy_image)31 32 actual, is_flipped = preprocess_utils.flip_dim([image], prob=0, dim=0)33 self.assertAllEqual(numpy_image, actual)34 self.assertFalse(is_flipped)35 actual, is_flipped = preprocess_utils.flip_dim([image], prob=0, dim=1)36 self.assertAllEqual(numpy_image, actual)37 self.assertFalse(is_flipped)38 actual, is_flipped = preprocess_utils.flip_dim([image], prob=0, dim=2)39 self.assertAllEqual(numpy_image, actual)40 self.assertFalse(is_flipped)41 42 def testFlipWhenProbIsOne(self):43 numpy_image = np.dstack([[[5., 6.],44 [9., 0.]],45 [[4., 3.],46 [3., 5.]]])47 dim0_flipped = np.dstack([[[9., 0.],48 [5., 6.]],49 [[3., 5.],50 [4., 3.]]])51 dim1_flipped = np.dstack([[[6., 5.],52 [0., 9.]],53 [[3., 4.],54 [5., 3.]]])55 dim2_flipped = np.dstack([[[4., 3.],56 [3., 5.]],57 [[5., 6.],58 [9., 0.]]])59 image = tf.convert_to_tensor(numpy_image)60 61 actual, is_flipped = preprocess_utils.flip_dim([image], prob=1, dim=0)62 self.assertAllEqual(dim0_flipped, actual)63 self.assertTrue(is_flipped)64 actual, is_flipped = preprocess_utils.flip_dim([image], prob=1, dim=1)65 self.assertAllEqual(dim1_flipped, actual)66 self.assertTrue(is_flipped)67 actual, is_flipped = preprocess_utils.flip_dim([image], prob=1, dim=2)68 self.assertAllEqual(dim2_flipped, actual)69 self.assertTrue(is_flipped)70 71 def testFlipMultipleImagesConsistentlyWhenProbIsOne(self):72 numpy_image = np.dstack([[[5., 6.],73 [9., 0.]],74 [[4., 3.],75 [3., 5.]]])76 numpy_label = np.dstack([[[0., 1.],77 [2., 3.]]])78 image_dim1_flipped = np.dstack([[[6., 5.],79 [0., 9.]],80 [[3., 4.],81 [5., 3.]]])82 label_dim1_flipped = np.dstack([[[1., 0.],83 [3., 2.]]])84 image = tf.convert_to_tensor(numpy_image)85 label = tf.convert_to_tensor(numpy_label)86 87 image, label, is_flipped = preprocess_utils.flip_dim(88 [image, label], prob=1, dim=1)89 self.assertAllEqual(image_dim1_flipped, image)90 self.assertAllEqual(label_dim1_flipped, label)91 self.assertTrue(is_flipped)92 93 def testReturnRandomFlipsOnMultipleEvals(self):94 numpy_image = np.dstack([[[5., 6.],95 [9., 0.]],96 [[4., 3.],97 [3., 5.]]])98 dim1_flipped = np.dstack([[[6., 5.],99 [0., 9.]],100 [[3., 4.],101 [5., 3.]]])102 image = tf.convert_to_tensor(numpy_image)103 original_image, not_flipped = preprocess_utils.flip_dim(104 [image], prob=0, dim=1)105 flip_image, is_flipped = preprocess_utils.flip_dim(106 [image], prob=1.0, dim=1)107 self.assertAllEqual(numpy_image, original_image)108 self.assertFalse(not_flipped)109 self.assertAllEqual(dim1_flipped, flip_image)110 self.assertTrue(is_flipped)111 112 def testReturnCorrectCropOfSingleImage(self):113 np.random.seed(0)114 115 height, width = 10, 20116 image = np.random.randint(0, 256, size=(height, width, 3))117 118 crop_height, crop_width = 2, 4119 120 [cropped] = preprocess_utils.random_crop([tf.convert_to_tensor(image)],121 crop_height,122 crop_width)123 124 # Ensure we can find the cropped image in the original:125 is_found = False126 for x in range(0, width - crop_width + 1):127 for y in range(0, height - crop_height + 1):128 if np.isclose(image[y:y+crop_height, x:x+crop_width, :],129 cropped).all():130 is_found = True131 break132 133 self.assertTrue(is_found)134 135 def testRandomCropMaintainsNumberOfChannels(self):136 np.random.seed(0)137 138 crop_height, crop_width = 10, 20139 image = np.random.randint(0, 256, size=(100, 200, 3))140 141 tf.random.set_seed(37)142 [cropped] = preprocess_utils.random_crop(143 [tf.convert_to_tensor(image)], crop_height, crop_width)144 145 self.assertListEqual(cropped.shape.as_list(), [crop_height, crop_width, 3])146 147 def testReturnDifferentCropAreasOnTwoEvals(self):148 tf.random.set_seed(0)149 150 crop_height, crop_width = 2, 3151 image = np.random.randint(0, 256, size=(100, 200, 3))152 [cropped0] = preprocess_utils.random_crop(153 [tf.convert_to_tensor(image)], crop_height, crop_width)154 [cropped1] = preprocess_utils.random_crop(155 [tf.convert_to_tensor(image)], crop_height, crop_width)156 157 self.assertFalse(np.isclose(cropped0.numpy(), cropped1.numpy()).all())158 159 def testReturnConsistenCropsOfImagesInTheList(self):160 tf.random.set_seed(0)161 162 height, width = 10, 20163 crop_height, crop_width = 2, 3164 labels = np.linspace(0, height * width-1, height * width)165 labels = labels.reshape((height, width, 1))166 image = np.tile(labels, (1, 1, 3))167 168 [cropped_image, cropped_label] = preprocess_utils.random_crop(169 [tf.convert_to_tensor(image), tf.convert_to_tensor(labels)],170 crop_height, crop_width)171 172 for i in range(3):173 self.assertAllEqual(cropped_image[:, :, i], tf.squeeze(cropped_label))174 175 def testDieOnRandomCropWhenImagesWithDifferentWidth(self):176 crop_height, crop_width = 2, 3177 image1 = tf.convert_to_tensor(np.random.rand(4, 5, 3))178 image2 = tf.convert_to_tensor(np.random.rand(4, 6, 1))179 180 with self.assertRaises(tf.errors.InvalidArgumentError):181 _ = preprocess_utils.random_crop([image1, image2], crop_height,182 crop_width)183 184 def testDieOnRandomCropWhenImagesWithDifferentHeight(self):185 crop_height, crop_width = 2, 3186 image1 = tf.convert_to_tensor(np.random.rand(4, 5, 3))187 image2 = tf.convert_to_tensor(np.random.rand(5, 5, 1))188 189 with self.assertRaises(tf.errors.InvalidArgumentError):190 _ = preprocess_utils.random_crop([image1, image2], crop_height,191 crop_width)192 193 def testDieOnRandomCropWhenCropSizeIsGreaterThanImage(self):194 crop_height, crop_width = 5, 9195 image1 = tf.convert_to_tensor(np.random.rand(4, 5, 3))196 image2 = tf.convert_to_tensor(np.random.rand(4, 5, 1))197 198 with self.assertRaises(tf.errors.InvalidArgumentError):199 _ = preprocess_utils.random_crop([image1, image2], crop_height,200 crop_width)201 202 def testRandomScaleFitsInRange(self):203 scale_value = preprocess_utils.get_random_scale(1., 2., 0.)204 self.assertGreaterEqual(scale_value, 1.)205 self.assertLessEqual(scale_value, 2.)206 207 def testDeterminedRandomScaleReturnsNumber(self):208 scale = preprocess_utils.get_random_scale(1., 1., 0.)209 self.assertEqual(scale, 1.)210 211 def testResizeTensorsToRange(self):212 test_shapes = [[60, 40],213 [15, 30],214 [15, 50]]215 min_size = 50216 max_size = 100217 factor = None218 expected_shape_list = [(75, 50, 3),219 (50, 100, 3),220 (30, 100, 3)]221 for i, test_shape in enumerate(test_shapes):222 image = tf.random.normal([test_shape[0], test_shape[1], 3])223 new_tensor_list = preprocess_utils.resize_to_range(224 image=image,225 label=None,226 min_size=min_size,227 max_size=max_size,228 factor=factor,229 align_corners=True)230 self.assertEqual(new_tensor_list[0].shape, expected_shape_list[i])231 232 def testResizeTensorsToRangeWithFactor(self):233 test_shapes = [[60, 40],234 [15, 30],235 [15, 50]]236 min_size = 50237 max_size = 98238 factor = 8239 expected_image_shape_list = [(81, 57, 3),240 (49, 97, 3),241 (33, 97, 3)]242 expected_label_shape_list = [(81, 57, 1),243 (49, 97, 1),244 (33, 97, 1)]245 for i, test_shape in enumerate(test_shapes):246 image = tf.random.normal([test_shape[0], test_shape[1], 3])247 label = tf.random.normal([test_shape[0], test_shape[1], 1])248 new_tensor_list = preprocess_utils.resize_to_range(249 image=image,250 label=label,251 min_size=min_size,252 max_size=max_size,253 factor=factor,254 align_corners=True)255 self.assertEqual(new_tensor_list[0].shape, expected_image_shape_list[i])256 self.assertEqual(new_tensor_list[1].shape, expected_label_shape_list[i])257 258 def testResizeTensorsToRangeWithSimilarMinMaxSizes(self):259 test_shapes = [[60, 40],260 [15, 30],261 [15, 50]]262 # Values set so that one of the side = 97.263 min_size = 96264 max_size = 98265 factor = 8266 expected_image_shape_list = [(97, 65, 3),267 (49, 97, 3),268 (33, 97, 3)]269 expected_label_shape_list = [(97, 65, 1),270 (49, 97, 1),271 (33, 97, 1)]272 for i, test_shape in enumerate(test_shapes):273 image = tf.random.normal([test_shape[0], test_shape[1], 3])274 label = tf.random.normal([test_shape[0], test_shape[1], 1])275 new_tensor_list = preprocess_utils.resize_to_range(276 image=image,277 label=label,278 min_size=min_size,279 max_size=max_size,280 factor=factor,281 align_corners=True)282 self.assertEqual(new_tensor_list[0].shape, expected_image_shape_list[i])283 self.assertEqual(new_tensor_list[1].shape, expected_label_shape_list[i])284 285 def testResizeTensorsToRangeWithEqualMaxSize(self):286 test_shapes = [[97, 38],287 [96, 97]]288 # Make max_size equal to the larger value of test_shapes.289 min_size = 97290 max_size = 97291 factor = 8292 expected_image_shape_list = [(97, 41, 3),293 (97, 97, 3)]294 expected_label_shape_list = [(97, 41, 1),295 (97, 97, 1)]296 for i, test_shape in enumerate(test_shapes):297 image = tf.random.normal([test_shape[0], test_shape[1], 3])298 label = tf.random.normal([test_shape[0], test_shape[1], 1])299 new_tensor_list = preprocess_utils.resize_to_range(300 image=image,301 label=label,302 min_size=min_size,303 max_size=max_size,304 factor=factor,305 align_corners=True)306 self.assertEqual(new_tensor_list[0].shape, expected_image_shape_list[i])307 self.assertEqual(new_tensor_list[1].shape, expected_label_shape_list[i])308 309 def testResizeTensorsToRangeWithPotentialErrorInTFCeil(self):310 test_shape = [3936, 5248]311 # Make max_size equal to the larger value of test_shapes.312 min_size = 1441313 max_size = 1441314 factor = 16315 expected_image_shape = (1089, 1441, 3)316 expected_label_shape = (1089, 1441, 1)317 image = tf.random.normal([test_shape[0], test_shape[1], 3])318 label = tf.random.normal([test_shape[0], test_shape[1], 1])319 new_tensor_list = preprocess_utils.resize_to_range(320 image=image,321 label=label,322 min_size=min_size,323 max_size=max_size,324 factor=factor,325 align_corners=True)326 self.assertEqual(new_tensor_list[0].shape, expected_image_shape)327 self.assertEqual(new_tensor_list[1].shape, expected_label_shape)328 329 def testResizeTensorWithOnlyMaxSize(self):330 test_shapes = [[97, 38],331 [96, 18]]332 333 max_size = (97, 28)334 # Since the second test shape already fits max size, do nothing.335 expected_image_shape_list = [(71, 28, 3),336 (96, 18, 3)]337 for i, test_shape in enumerate(test_shapes):338 image = tf.random.normal([test_shape[0], test_shape[1], 3])339 new_tensor_list = preprocess_utils.resize_to_range(340 image=image,341 label=None,342 min_size=None,343 max_size=max_size,344 align_corners=True)345 self.assertEqual(new_tensor_list[0].shape, expected_image_shape_list[i])346 347 348if __name__ == '__main__':349 tf.test.main()350