ops_test.py 60 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""Tests for object_detection.utils.ops."""
pkulzc's avatar
pkulzc committed
17
18
19
20
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

21
import numpy as np
pkulzc's avatar
pkulzc committed
22
23
import six
from six.moves import range
24
25
import tensorflow.compat.v1 as tf
import tf_slim as slim
26
27
from object_detection.core import standard_fields as fields
from object_detection.utils import ops
28
from object_detection.utils import test_case
29
30


31
class NormalizedToImageCoordinatesTest(test_case.TestCase):
32
33
34
35

  def test_normalized_to_image_coordinates(self):
    normalized_boxes_np = np.array([[[0.0, 0.0, 1.0, 1.0]],
                                    [[0.5, 0.5, 1.0, 1.0]]])
36
37
38
39
40
41

    def graph_fn(normalized_boxes):
      image_shape = tf.convert_to_tensor([1, 4, 4, 3], dtype=tf.int32)
      absolute_boxes = ops.normalized_to_image_coordinates(
          normalized_boxes, image_shape, parallel_iterations=2)
      return absolute_boxes
42
43
44
45

    expected_boxes = np.array([[[0, 0, 4, 4]],
                               [[2, 2, 4, 4]]])

46
    absolute_boxes = self.execute(graph_fn, [normalized_boxes_np])
47
48
49
    self.assertAllEqual(absolute_boxes, expected_boxes)


50
class ReduceSumTrailingDimensions(test_case.TestCase):
51
52

  def test_reduce_sum_trailing_dimensions(self):
53
54
55
56
57
58

    def graph_fn(input_tensor):
      reduced_tensor = ops.reduce_sum_trailing_dimensions(input_tensor, ndims=2)
      return reduced_tensor

    reduced_np = self.execute(graph_fn, [np.ones((2, 2, 2), np.float32)])
59
60
61
    self.assertAllClose(reduced_np, 2 * np.ones((2, 2), np.float32))


62
class MeshgridTest(test_case.TestCase):
63
64
65

  def test_meshgrid_numpy_comparison(self):
    """Tests meshgrid op with vectors, for which it should match numpy."""
66

67
68
    x = np.arange(4)
    y = np.arange(6)
69
70
71
72
73

    def graph_fn():
      xgrid, ygrid = ops.meshgrid(x, y)
      return xgrid, ygrid

74
    exp_xgrid, exp_ygrid = np.meshgrid(x, y)
75
76
77
    xgrid_output, ygrid_output = self.execute(graph_fn, [])
    self.assertAllEqual(xgrid_output, exp_xgrid)
    self.assertAllEqual(ygrid_output, exp_ygrid)
78
79
80
81
82
83
84

  def test_meshgrid_multidimensional(self):
    np.random.seed(18)
    x = np.random.rand(4, 1, 2).astype(np.float32)
    y = np.random.rand(2, 3).astype(np.float32)

    grid_shape = list(y.shape) + list(x.shape)
85
86
87
88
89
90
91
92

    def graph_fn():
      xgrid, ygrid = ops.meshgrid(x, y)
      self.assertEqual(xgrid.get_shape().as_list(), grid_shape)
      self.assertEqual(ygrid.get_shape().as_list(), grid_shape)
      return xgrid, ygrid

    xgrid_output, ygrid_output = self.execute(graph_fn, [])
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108

    # Check the shape of the output grids
    self.assertEqual(xgrid_output.shape, tuple(grid_shape))
    self.assertEqual(ygrid_output.shape, tuple(grid_shape))

    # Check a few elements
    test_elements = [((3, 0, 0), (1, 2)),
                     ((2, 0, 1), (0, 0)),
                     ((0, 0, 0), (1, 1))]
    for xind, yind in test_elements:
      # These are float equality tests, but the meshgrid op should not introduce
      # rounding.
      self.assertEqual(xgrid_output[yind + xind], x[xind])
      self.assertEqual(ygrid_output[yind + xind], y[yind])


109
class OpsTestFixedPadding(test_case.TestCase):
110
111

  def test_3x3_kernel(self):
112
113
114
115
116
117
118

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.fixed_padding(tensor, 3)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
119
120
121
    self.assertEqual((1, 4, 4, 1), padded_tensor_out.shape)

  def test_5x5_kernel(self):
122
123
124
125
126
127
128

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.fixed_padding(tensor, 5)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
129
130
131
    self.assertEqual((1, 6, 6, 1), padded_tensor_out.shape)

  def test_3x3_atrous_kernel(self):
132
133
134
135
136
137
138

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.fixed_padding(tensor, 3, 2)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
139
140
141
    self.assertEqual((1, 6, 6, 1), padded_tensor_out.shape)


142
class OpsTestPadToMultiple(test_case.TestCase):
143
144

  def test_zero_padding(self):
145
146
147
148
149
150
151

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.pad_to_multiple(tensor, 1)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
152
153
154
    self.assertEqual((1, 2, 2, 1), padded_tensor_out.shape)

  def test_no_padding(self):
155
156
157
158
159
160
161

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.pad_to_multiple(tensor, 2)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
162
163
    self.assertEqual((1, 2, 2, 1), padded_tensor_out.shape)

164
  def test_non_square_padding(self):
165
166
167
168
169
170
171

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]]]])
      padded_tensor = ops.pad_to_multiple(tensor, 2)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
172
173
    self.assertEqual((1, 2, 2, 1), padded_tensor_out.shape)

174
  def test_padding(self):
175
176
177
178
179
180
181

    def graph_fn():
      tensor = tf.constant([[[[0.], [0.]], [[0.], [0.]]]])
      padded_tensor = ops.pad_to_multiple(tensor, 4)
      return padded_tensor

    padded_tensor_out = self.execute(graph_fn, [])
182
183
184
    self.assertEqual((1, 4, 4, 1), padded_tensor_out.shape)


185
class OpsTestPaddedOneHotEncoding(test_case.TestCase):
186
187

  def test_correct_one_hot_tensor_with_no_pad(self):
188
189
190
191
192
193

    def graph_fn():
      indices = tf.constant([1, 2, 3, 5])
      one_hot_tensor = ops.padded_one_hot_encoding(indices, depth=6, left_pad=0)
      return one_hot_tensor

194
195
196
197
    expected_tensor = np.array([[0, 1, 0, 0, 0, 0],
                                [0, 0, 1, 0, 0, 0],
                                [0, 0, 0, 1, 0, 0],
                                [0, 0, 0, 0, 0, 1]], np.float32)
198

199
    out_one_hot_tensor = self.execute(graph_fn, [])
200
201
    self.assertAllClose(out_one_hot_tensor, expected_tensor, rtol=1e-10,
                        atol=1e-10)
202
203

  def test_correct_one_hot_tensor_with_pad_one(self):
204
205
206
207
208
209

    def graph_fn():
      indices = tf.constant([1, 2, 3, 5])
      one_hot_tensor = ops.padded_one_hot_encoding(indices, depth=6, left_pad=1)
      return one_hot_tensor

210
211
212
213
    expected_tensor = np.array([[0, 0, 1, 0, 0, 0, 0],
                                [0, 0, 0, 1, 0, 0, 0],
                                [0, 0, 0, 0, 1, 0, 0],
                                [0, 0, 0, 0, 0, 0, 1]], np.float32)
214
    out_one_hot_tensor = self.execute(graph_fn, [])
215
216
    self.assertAllClose(out_one_hot_tensor, expected_tensor, rtol=1e-10,
                        atol=1e-10)
217
218

  def test_correct_one_hot_tensor_with_pad_three(self):
219
220
221
222
223
224

    def graph_fn():
      indices = tf.constant([1, 2, 3, 5])
      one_hot_tensor = ops.padded_one_hot_encoding(indices, depth=6, left_pad=3)
      return one_hot_tensor

225
226
227
228
    expected_tensor = np.array([[0, 0, 0, 0, 1, 0, 0, 0, 0],
                                [0, 0, 0, 0, 0, 1, 0, 0, 0],
                                [0, 0, 0, 0, 0, 0, 1, 0, 0],
                                [0, 0, 0, 0, 0, 0, 0, 0, 1]], np.float32)
229

230
    out_one_hot_tensor = self.execute(graph_fn, [])
231
232
    self.assertAllClose(out_one_hot_tensor, expected_tensor, rtol=1e-10,
                        atol=1e-10)
233
234

  def test_correct_padded_one_hot_tensor_with_empty_indices(self):
235

236
237
    depth = 6
    pad = 2
238
239
240
241
242
243
244

    def graph_fn():
      indices = tf.constant([])
      one_hot_tensor = ops.padded_one_hot_encoding(
          indices, depth=depth, left_pad=pad)
      return one_hot_tensor

245
    expected_tensor = np.zeros((0, depth + pad))
246
    out_one_hot_tensor = self.execute(graph_fn, [])
247
248
    self.assertAllClose(out_one_hot_tensor, expected_tensor, rtol=1e-10,
                        atol=1e-10)
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275

  def test_return_none_on_zero_depth(self):
    indices = tf.constant([1, 2, 3, 4, 5])
    one_hot_tensor = ops.padded_one_hot_encoding(indices, depth=0, left_pad=2)
    self.assertEqual(one_hot_tensor, None)

  def test_raise_value_error_on_rank_two_input(self):
    indices = tf.constant(1.0, shape=(2, 3))
    with self.assertRaises(ValueError):
      ops.padded_one_hot_encoding(indices, depth=6, left_pad=2)

  def test_raise_value_error_on_negative_pad(self):
    indices = tf.constant(1.0, shape=(2, 3))
    with self.assertRaises(ValueError):
      ops.padded_one_hot_encoding(indices, depth=6, left_pad=-1)

  def test_raise_value_error_on_float_pad(self):
    indices = tf.constant(1.0, shape=(2, 3))
    with self.assertRaises(ValueError):
      ops.padded_one_hot_encoding(indices, depth=6, left_pad=0.1)

  def test_raise_value_error_on_float_depth(self):
    indices = tf.constant(1.0, shape=(2, 3))
    with self.assertRaises(ValueError):
      ops.padded_one_hot_encoding(indices, depth=0.1, left_pad=2)


276
class OpsDenseToSparseBoxesTest(test_case.TestCase):
277
278
279
280
281

  def test_return_all_boxes_when_all_input_boxes_are_valid(self):
    num_classes = 4
    num_valid_boxes = 3
    code_size = 4
282
283
284
285
286
287
288
289
290
291

    def graph_fn(dense_location, dense_num_boxes):
      box_locations, box_classes = ops.dense_to_sparse_boxes(
          dense_location, dense_num_boxes, num_classes)
      return box_locations, box_classes

    dense_location_np = np.random.uniform(size=[num_valid_boxes, code_size])
    dense_num_boxes_np = np.array([1, 0, 0, 2], dtype=np.int32)

    expected_box_locations = dense_location_np
292
    expected_box_classses = np.array([0, 3, 3])
293
294
295
296

    # Executing on CPU only since output shape is not constant.
    box_locations, box_classes = self.execute_cpu(
        graph_fn, [dense_location_np, dense_num_boxes_np])
297
298
299
300
301
302
303
304
305
306
307

    self.assertAllClose(box_locations, expected_box_locations, rtol=1e-6,
                        atol=1e-6)
    self.assertAllEqual(box_classes, expected_box_classses)

  def test_return_only_valid_boxes_when_input_contains_invalid_boxes(self):
    num_classes = 4
    num_valid_boxes = 3
    num_boxes = 10
    code_size = 4

308
309
310
311
312
313
314
315
316
    def graph_fn(dense_location, dense_num_boxes):
      box_locations, box_classes = ops.dense_to_sparse_boxes(
          dense_location, dense_num_boxes, num_classes)
      return box_locations, box_classes

    dense_location_np = np.random.uniform(size=[num_boxes, code_size])
    dense_num_boxes_np = np.array([1, 0, 0, 2], dtype=np.int32)

    expected_box_locations = dense_location_np[:num_valid_boxes]
317
    expected_box_classses = np.array([0, 3, 3])
318
319
320
321

    # Executing on CPU only since output shape is not constant.
    box_locations, box_classes = self.execute_cpu(
        graph_fn, [dense_location_np, dense_num_boxes_np])
322
323
324
325
326
327

    self.assertAllClose(box_locations, expected_box_locations, rtol=1e-6,
                        atol=1e-6)
    self.assertAllEqual(box_classes, expected_box_classses)


328
class OpsTestIndicesToDenseVector(test_case.TestCase):
329
330
331
332
333
334
335
336
337

  def test_indices_to_dense_vector(self):
    size = 10000
    num_indices = np.random.randint(size)
    rand_indices = np.random.permutation(np.arange(size))[0:num_indices]

    expected_output = np.zeros(size, dtype=np.float32)
    expected_output[rand_indices] = 1.

338
339
340
341
    def graph_fn():
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(tf_rand_indices, size)
      return indicator
342

343
344
345
    output = self.execute(graph_fn, [])
    self.assertAllEqual(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
346
347
348
349
350
351
352
353
354
355

  def test_indices_to_dense_vector_size_at_inference(self):
    size = 5000
    num_indices = 250
    all_indices = np.arange(size)
    rand_indices = np.random.permutation(all_indices)[0:num_indices]

    expected_output = np.zeros(size, dtype=np.float32)
    expected_output[rand_indices] = 1.

356
357
358
359
360
    def graph_fn(tf_all_indices):
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(tf_rand_indices,
                                              tf.shape(tf_all_indices)[0])
      return indicator
361

362
363
364
    output = self.execute(graph_fn, [all_indices])
    self.assertAllEqual(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
365
366
367
368
369
370
371
372
373

  def test_indices_to_dense_vector_int(self):
    size = 500
    num_indices = 25
    rand_indices = np.random.permutation(np.arange(size))[0:num_indices]

    expected_output = np.zeros(size, dtype=np.int64)
    expected_output[rand_indices] = 1

374
375
376
377
378
    def graph_fn():
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(
          tf_rand_indices, size, 1, dtype=tf.int64)
      return indicator
379

380
381
382
    output = self.execute(graph_fn, [])
    self.assertAllEqual(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
383
384
385
386
387
388
389
390
391
392
393

  def test_indices_to_dense_vector_custom_values(self):
    size = 100
    num_indices = 10
    rand_indices = np.random.permutation(np.arange(size))[0:num_indices]
    indices_value = np.random.rand(1)
    default_value = np.random.rand(1)

    expected_output = np.float32(np.ones(size) * default_value)
    expected_output[rand_indices] = indices_value

394
395
396
397
398
399
400
401
    def graph_fn():
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(
          tf_rand_indices,
          size,
          indices_value=indices_value,
          default_value=default_value)
      return indicator
402

403
404
405
    output = self.execute(graph_fn, [])
    self.assertAllClose(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
406
407
408
409
410
411
412
413

  def test_indices_to_dense_vector_all_indices_as_input(self):
    size = 500
    num_indices = 500
    rand_indices = np.random.permutation(np.arange(size))[0:num_indices]

    expected_output = np.ones(size, dtype=np.float32)

414
415
416
417
    def graph_fn():
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(tf_rand_indices, size)
      return indicator
418

419
420
421
    output = self.execute(graph_fn, [])
    self.assertAllEqual(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
422
423
424
425
426
427
428

  def test_indices_to_dense_vector_empty_indices_as_input(self):
    size = 500
    rand_indices = []

    expected_output = np.zeros(size, dtype=np.float32)

429
430
431
432
    def graph_fn():
      tf_rand_indices = tf.constant(rand_indices)
      indicator = ops.indices_to_dense_vector(tf_rand_indices, size)
      return indicator
433

434
435
436
    output = self.execute(graph_fn, [])
    self.assertAllEqual(output, expected_output)
    self.assertEqual(output.dtype, expected_output.dtype)
437
438


439
class GroundtruthFilterTest(test_case.TestCase):
440
441

  def test_filter_groundtruth(self):
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478

    def graph_fn(input_image, input_boxes, input_classes, input_is_crowd,
                 input_area, input_difficult, input_label_types,
                 input_confidences, valid_indices):
      input_tensors = {
          fields.InputDataFields.image: input_image,
          fields.InputDataFields.groundtruth_boxes: input_boxes,
          fields.InputDataFields.groundtruth_classes: input_classes,
          fields.InputDataFields.groundtruth_is_crowd: input_is_crowd,
          fields.InputDataFields.groundtruth_area: input_area,
          fields.InputDataFields.groundtruth_difficult: input_difficult,
          fields.InputDataFields.groundtruth_label_types: input_label_types,
          fields.InputDataFields.groundtruth_confidences: input_confidences,
      }

      output_tensors = ops.retain_groundtruth(input_tensors, valid_indices)
      return output_tensors

    input_image = np.random.rand(224, 224, 3)
    input_boxes = np.array([[0.2, 0.4, 0.1, 0.8], [0.2, 0.4, 1.0, 0.8]],
                           dtype=np.float32)
    input_classes = np.array([1, 2], dtype=np.int32)
    input_is_crowd = np.array([False, True], dtype=np.bool)
    input_area = np.array([32, 48], dtype=np.float32)
    input_difficult = np.array([True, False], dtype=np.bool)
    input_label_types = np.array(['APPROPRIATE', 'INCORRECT'],
                                 dtype=np.string_)
    input_confidences = np.array([0.99, 0.5], dtype=np.float32)
    valid_indices = np.array([0], dtype=np.int32)

    # Strings are not supported on TPU.
    output_tensors = self.execute_cpu(
        graph_fn,
        [input_image, input_boxes, input_classes, input_is_crowd, input_area,
         input_difficult, input_label_types, input_confidences, valid_indices]
    )

479
    expected_tensors = {
480
        fields.InputDataFields.image: input_image,
481
482
483
484
485
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [1],
        fields.InputDataFields.groundtruth_is_crowd: [False],
        fields.InputDataFields.groundtruth_area: [32],
        fields.InputDataFields.groundtruth_difficult: [True],
pkulzc's avatar
pkulzc committed
486
        fields.InputDataFields.groundtruth_label_types: [six.b('APPROPRIATE')],
487
        fields.InputDataFields.groundtruth_confidences: [0.99],
488
    }
489
490
491
492
493
494
495
496
497
498
    for key in [fields.InputDataFields.image,
                fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])

    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd,
                fields.InputDataFields.groundtruth_label_types]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
499
500

  def test_filter_with_missing_fields(self):
501
502
503
504
505
506

    input_boxes = np.array([[0.2, 0.4, 0.1, 0.8], [0.2, 0.4, 1.0, 0.8]],
                           dtype=np.float)
    input_classes = np.array([1, 2], dtype=np.int32)
    valid_indices = np.array([0], dtype=np.int32)

507
508
509
510
511
512
513
    expected_tensors = {
        fields.InputDataFields.groundtruth_boxes:
        [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes:
        [1]
    }

514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
    def graph_fn(input_boxes, input_classes, valid_indices):
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes: input_boxes,
          fields.InputDataFields.groundtruth_classes: input_classes
      }
      output_tensors = ops.retain_groundtruth(input_tensors, valid_indices)
      return output_tensors

    output_tensors = self.execute(graph_fn, [input_boxes, input_classes,
                                             valid_indices])

    for key in [fields.InputDataFields.groundtruth_boxes]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
529
530

  def test_filter_with_empty_fields(self):
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553

    def graph_fn(input_boxes, input_classes, input_is_crowd, input_area,
                 input_difficult, input_confidences, valid_indices):
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes: input_boxes,
          fields.InputDataFields.groundtruth_classes: input_classes,
          fields.InputDataFields.groundtruth_is_crowd: input_is_crowd,
          fields.InputDataFields.groundtruth_area: input_area,
          fields.InputDataFields.groundtruth_difficult: input_difficult,
          fields.InputDataFields.groundtruth_confidences: input_confidences,
      }
      output_tensors = ops.retain_groundtruth(input_tensors, valid_indices)
      return output_tensors

    input_boxes = np.array([[0.2, 0.4, 0.1, 0.8], [0.2, 0.4, 1.0, 0.8]],
                           dtype=np.float)
    input_classes = np.array([1, 2], dtype=np.int32)
    input_is_crowd = np.array([False, True], dtype=np.bool)
    input_area = np.array([], dtype=np.float32)
    input_difficult = np.array([], dtype=np.float32)
    input_confidences = np.array([0.99, 0.5], dtype=np.float32)
    valid_indices = np.array([0], dtype=np.int32)

554
    expected_tensors = {
555
556
557
558
559
560
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [1],
        fields.InputDataFields.groundtruth_is_crowd: [False],
        fields.InputDataFields.groundtruth_area: [],
        fields.InputDataFields.groundtruth_difficult: [],
        fields.InputDataFields.groundtruth_confidences: [0.99],
561
    }
562
563
564
565
566
567
568
569
570
571
572
    output_tensors = self.execute(graph_fn, [
        input_boxes, input_classes, input_is_crowd, input_area,
        input_difficult, input_confidences, valid_indices])

    for key in [fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
573
574
575

  def test_filter_with_empty_groundtruth_boxes(self):

576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
    def graph_fn(input_boxes, input_classes, input_is_crowd, input_area,
                 input_difficult, input_confidences, valid_indices):
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes: input_boxes,
          fields.InputDataFields.groundtruth_classes: input_classes,
          fields.InputDataFields.groundtruth_is_crowd: input_is_crowd,
          fields.InputDataFields.groundtruth_area: input_area,
          fields.InputDataFields.groundtruth_difficult: input_difficult,
          fields.InputDataFields.groundtruth_confidences: input_confidences,
      }
      output_tensors = ops.retain_groundtruth(input_tensors, valid_indices)
      return output_tensors

    input_boxes = np.array([], dtype=np.float).reshape(0, 4)
    input_classes = np.array([], dtype=np.int32)
    input_is_crowd = np.array([], dtype=np.bool)
    input_area = np.array([], dtype=np.float32)
    input_difficult = np.array([], dtype=np.float32)
    input_confidences = np.array([], dtype=np.float32)
    valid_indices = np.array([], dtype=np.int32)

    output_tensors = self.execute(graph_fn, [input_boxes, input_classes,
                                             input_is_crowd, input_area,
                                             input_difficult,
                                             input_confidences,
                                             valid_indices])
    for key in output_tensors:
      if key == fields.InputDataFields.groundtruth_boxes:
        self.assertAllEqual([0, 4], output_tensors[key].shape)
      else:
        self.assertAllEqual([0], output_tensors[key].shape)


class RetainGroundTruthWithPositiveClasses(test_case.TestCase):
610
611

  def test_filter_groundtruth_with_positive_classes(self):
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640

    def graph_fn(input_image, input_boxes, input_classes, input_is_crowd,
                 input_area, input_difficult, input_label_types,
                 input_confidences):
      input_tensors = {
          fields.InputDataFields.image: input_image,
          fields.InputDataFields.groundtruth_boxes: input_boxes,
          fields.InputDataFields.groundtruth_classes: input_classes,
          fields.InputDataFields.groundtruth_is_crowd: input_is_crowd,
          fields.InputDataFields.groundtruth_area: input_area,
          fields.InputDataFields.groundtruth_difficult: input_difficult,
          fields.InputDataFields.groundtruth_label_types: input_label_types,
          fields.InputDataFields.groundtruth_confidences: input_confidences,
      }
      output_tensors = ops.retain_groundtruth_with_positive_classes(
          input_tensors)
      return output_tensors

    input_image = np.random.rand(224, 224, 3)
    input_boxes = np.array([[0.2, 0.4, 0.1, 0.8], [0.2, 0.4, 1.0, 0.8]],
                           dtype=np.float)
    input_classes = np.array([1, 0], dtype=np.int32)
    input_is_crowd = np.array([False, True], dtype=np.bool)
    input_area = np.array([32, 48], dtype=np.float32)
    input_difficult = np.array([True, False], dtype=np.bool)
    input_label_types = np.array(['APPROPRIATE', 'INCORRECT'],
                                 dtype=np.string_)
    input_confidences = np.array([0.99, 0.5], dtype=np.float32)

641
    expected_tensors = {
642
        fields.InputDataFields.image: input_image,
643
644
645
646
647
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [1],
        fields.InputDataFields.groundtruth_is_crowd: [False],
        fields.InputDataFields.groundtruth_area: [32],
        fields.InputDataFields.groundtruth_difficult: [True],
pkulzc's avatar
pkulzc committed
648
        fields.InputDataFields.groundtruth_label_types: [six.b('APPROPRIATE')],
649
        fields.InputDataFields.groundtruth_confidences: [0.99],
650
651
    }

652
653
654
655
656
657
658
    # Executing on CPU because string types are not supported on TPU.
    output_tensors = self.execute_cpu(graph_fn,
                                      [input_image, input_boxes,
                                       input_classes, input_is_crowd,
                                       input_area,
                                       input_difficult, input_label_types,
                                       input_confidences])
659

660
661
662
663
664
665
666
667
668
669
670
671
    for key in [fields.InputDataFields.image,
                fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd,
                fields.InputDataFields.groundtruth_label_types]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])


class ReplaceNaNGroundtruthLabelScoresWithOnes(test_case.TestCase):
672
673

  def test_replace_nan_groundtruth_label_scores_with_ones(self):
674
675
676
677
678
679
680

    def graph_fn():
      label_scores = tf.constant([np.nan, 1.0, np.nan])
      output_tensor = ops.replace_nan_groundtruth_label_scores_with_ones(
          label_scores)
      return output_tensor

681
    expected_tensor = [1.0, 1.0, 1.0]
682
683
    output_tensor = self.execute(graph_fn, [])
    self.assertAllClose(expected_tensor, output_tensor)
684
685

  def test_input_equals_output_when_no_nans(self):
686

687
    input_label_scores = [0.5, 1.0, 1.0]
688
689
690
691
692
693
694
695
696
    def graph_fn():
      label_scores_tensor = tf.constant(input_label_scores)
      output_label_scores = ops.replace_nan_groundtruth_label_scores_with_ones(
          label_scores_tensor)
      return output_label_scores

    output_label_scores = self.execute(graph_fn, [])

    self.assertAllClose(input_label_scores, output_label_scores)
697
698


699
class GroundtruthFilterWithCrowdBoxesTest(test_case.TestCase):
700
701

  def test_filter_groundtruth_with_crowd_boxes(self):
702
703
704
705
706
707
708
709
710
711
712
713
714

    def graph_fn():
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes:
          [[0.1, 0.2, 0.6, 0.8], [0.2, 0.4, 0.1, 0.8]],
          fields.InputDataFields.groundtruth_classes: [1, 2],
          fields.InputDataFields.groundtruth_is_crowd: [True, False],
          fields.InputDataFields.groundtruth_area: [100.0, 238.7],
          fields.InputDataFields.groundtruth_confidences: [0.5, 0.99],
      }
      output_tensors = ops.filter_groundtruth_with_crowd_boxes(
          input_tensors)
      return output_tensors
715
716

    expected_tensors = {
717
718
719
720
721
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [2],
        fields.InputDataFields.groundtruth_is_crowd: [False],
        fields.InputDataFields.groundtruth_area: [238.7],
        fields.InputDataFields.groundtruth_confidences: [0.99],
722
723
    }

724
725
726
727
728
729
730
731
    output_tensors = self.execute(graph_fn, [])
    for key in [fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
732
733


734
class GroundtruthFilterWithNanBoxTest(test_case.TestCase):
735
736

  def test_filter_groundtruth_with_nan_box_coordinates(self):
737
738
739
740
741
742
743
744
745
746
747
748
749

    def graph_fn():
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes:
          [[np.nan, np.nan, np.nan, np.nan], [0.2, 0.4, 0.1, 0.8]],
          fields.InputDataFields.groundtruth_classes: [1, 2],
          fields.InputDataFields.groundtruth_is_crowd: [False, True],
          fields.InputDataFields.groundtruth_area: [100.0, 238.7],
          fields.InputDataFields.groundtruth_confidences: [0.5, 0.99],
      }
      output_tensors = ops.filter_groundtruth_with_nan_box_coordinates(
          input_tensors)
      return output_tensors
750
751

    expected_tensors = {
752
753
754
755
756
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [2],
        fields.InputDataFields.groundtruth_is_crowd: [True],
        fields.InputDataFields.groundtruth_area: [238.7],
        fields.InputDataFields.groundtruth_confidences: [0.99],
757
758
    }

759
760
761
762
763
764
765
766
    output_tensors = self.execute(graph_fn, [])
    for key in [fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
767
768


769
class GroundtruthFilterWithUnrecognizedClassesTest(test_case.TestCase):
770
771

  def test_filter_unrecognized_classes(self):
772
773
774
775
776
777
778
779
780
781
782
    def graph_fn():
      input_tensors = {
          fields.InputDataFields.groundtruth_boxes:
          [[.3, .3, .5, .7], [0.2, 0.4, 0.1, 0.8]],
          fields.InputDataFields.groundtruth_classes: [-1, 2],
          fields.InputDataFields.groundtruth_is_crowd: [False, True],
          fields.InputDataFields.groundtruth_area: [100.0, 238.7],
          fields.InputDataFields.groundtruth_confidences: [0.5, 0.99],
      }
      output_tensors = ops.filter_unrecognized_classes(input_tensors)
      return output_tensors
783
784
785
786
787
788
789
790
791

    expected_tensors = {
        fields.InputDataFields.groundtruth_boxes: [[0.2, 0.4, 0.1, 0.8]],
        fields.InputDataFields.groundtruth_classes: [2],
        fields.InputDataFields.groundtruth_is_crowd: [True],
        fields.InputDataFields.groundtruth_area: [238.7],
        fields.InputDataFields.groundtruth_confidences: [0.99],
    }

792
793
794
795
796
797
798
799
    output_tensors = self.execute(graph_fn, [])
    for key in [fields.InputDataFields.groundtruth_boxes,
                fields.InputDataFields.groundtruth_area,
                fields.InputDataFields.groundtruth_confidences]:
      self.assertAllClose(expected_tensors[key], output_tensors[key])
    for key in [fields.InputDataFields.groundtruth_classes,
                fields.InputDataFields.groundtruth_is_crowd]:
      self.assertAllEqual(expected_tensors[key], output_tensors[key])
800
801


802
class OpsTestNormalizeToTarget(test_case.TestCase):
803
804

  def test_create_normalize_to_target(self):
805
806
807
808

    if self.is_tf2():
      self.skipTest('Skipping as variable names not supported in eager mode.')

809
810
811
812
813
814
    inputs = tf.random_uniform([5, 10, 12, 3])
    target_norm_value = 4.0
    dim = 3
    with self.test_session():
      output = ops.normalize_to_target(inputs, target_norm_value, dim)
      self.assertEqual(output.op.name, 'NormalizeToTarget/mul')
815
      var_name = slim.get_variables()[0].name
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
      self.assertEqual(var_name, 'NormalizeToTarget/weights:0')

  def test_invalid_dim(self):
    inputs = tf.random_uniform([5, 10, 12, 3])
    target_norm_value = 4.0
    dim = 10
    with self.assertRaisesRegexp(
        ValueError,
        'dim must be non-negative but smaller than the input rank.'):
      ops.normalize_to_target(inputs, target_norm_value, dim)

  def test_invalid_target_norm_values(self):
    inputs = tf.random_uniform([5, 10, 12, 3])
    target_norm_value = [4.0, 4.0]
    dim = 3
    with self.assertRaisesRegexp(
        ValueError, 'target_norm_value must be a float or a list of floats'):
      ops.normalize_to_target(inputs, target_norm_value, dim)

  def test_correct_output_shape(self):
836
837
838
839
840
841
842
843
844

    if self.is_tf2():
      self.skipTest('normalize_to_target not supported in eager mode because,'
                    ' it requires creating variables.')

    inputs = np.random.uniform(size=(5, 10, 12, 3)).astype(np.float32)
    def graph_fn(inputs):
      target_norm_value = 4.0
      dim = 3
845
      output = ops.normalize_to_target(inputs, target_norm_value, dim)
846
847
848
849
850
851
      return output

    # Executing on CPU since creating a variable inside a conditional is not
    # supported.
    outputs = self.execute_cpu(graph_fn, [inputs])
    self.assertEqual(outputs.shape, inputs.shape)
852
853

  def test_correct_initial_output_values(self):
854
855
856
857
858
859
860
861
862
863
864
865
866

    if self.is_tf2():
      self.skipTest('normalize_to_target not supported in eager mode because,'
                    ' it requires creating variables.')
    def graph_fn():
      inputs = tf.constant([[[[3, 4], [7, 24]],
                             [[5, -12], [-1, 0]]]], tf.float32)
      target_norm_value = 10.0
      dim = 3
      normalized_inputs = ops.normalize_to_target(inputs, target_norm_value,
                                                  dim)
      return normalized_inputs

867
868
    expected_output = [[[[30/5.0, 40/5.0], [70/25.0, 240/25.0]],
                        [[50/13.0, -120/13.0], [-10, 0]]]]
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
    # Executing on CPU since creating a variable inside a conditional is not
    # supported.
    output = self.execute_cpu(graph_fn, [])
    self.assertAllClose(output, expected_output)

  def test_multiple_target_norm_values(self):

    if self.is_tf2():
      self.skipTest('normalize_to_target not supported in eager mode because,'
                    ' it requires creating variables.')

    def graph_fn():
      inputs = tf.constant([[[[3, 4], [7, 24]],
                             [[5, -12], [-1, 0]]]], tf.float32)
      target_norm_value = [10.0, 20.0]
      dim = 3
885
886
      normalized_inputs = ops.normalize_to_target(inputs, target_norm_value,
                                                  dim)
887
      return normalized_inputs
888
889
890

    expected_output = [[[[30/5.0, 80/5.0], [70/25.0, 480/25.0]],
                        [[50/13.0, -240/13.0], [-10, 0]]]]
891
892
893
894
895

    # Executing on CPU since creating a variable inside a conditional is not
    # supported.
    output = self.execute_cpu(graph_fn, [])
    self.assertAllClose(output, expected_output)
896
897


898
class OpsTestPositionSensitiveCropRegions(test_case.TestCase):
899
900
901

  def test_position_sensitive(self):
    num_spatial_bins = [3, 2]
902
    image_shape = [3, 2, 6]
903
904
905
906
907
908
909
910

    # The result for both boxes should be [[1, 2], [3, 4], [5, 6]]
    # before averaging.
    expected_output = np.array([3.5, 3.5]).reshape([2, 1, 1, 1])

    for crop_size_mult in range(1, 3):
      crop_size = [3 * crop_size_mult, 2 * crop_size_mult]

911
912
913
914
915
916
917
918
919
920
921
922
923
      def graph_fn():
        # First channel is 1's, second channel is 2's, etc.
        image = tf.constant(
            list(range(1, 3 * 2 + 1)) * 6, dtype=tf.float32, shape=image_shape)
        boxes = tf.random_uniform((2, 4))

        # pylint:disable=cell-var-from-loop
        ps_crop_and_pool = ops.position_sensitive_crop_regions(
            image, boxes, crop_size, num_spatial_bins, global_pool=True)
        return ps_crop_and_pool

      output = self.execute(graph_fn, [])
      self.assertAllClose(output, expected_output)
924
925
926

  def test_position_sensitive_with_equal_channels(self):
    num_spatial_bins = [2, 2]
927
    image_shape = [3, 3, 4]
928
929
    crop_size = [2, 2]

930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
    def graph_fn():
      image = tf.constant(
          list(range(1, 3 * 3 + 1)), dtype=tf.float32, shape=[3, 3, 1])
      tiled_image = tf.tile(image, [1, 1, image_shape[2]])
      boxes = tf.random_uniform((3, 4))
      box_ind = tf.constant([0, 0, 0], dtype=tf.int32)

      # All channels are equal so position-sensitive crop and resize should
      # work as the usual crop and resize for just one channel.
      crop = tf.image.crop_and_resize(tf.expand_dims(image, axis=0), boxes,
                                      box_ind, crop_size)
      crop_and_pool = tf.reduce_mean(crop, [1, 2], keepdims=True)

      ps_crop_and_pool = ops.position_sensitive_crop_regions(
          tiled_image,
          boxes,
          crop_size,
          num_spatial_bins,
          global_pool=True)

      return crop_and_pool, ps_crop_and_pool

    # Crop and resize op is not supported in TPUs.
    expected_output, output = self.execute_cpu(graph_fn, [])
    self.assertAllClose(output, expected_output)
955
956
957

  def test_raise_value_error_on_num_bins_less_than_one(self):
    num_spatial_bins = [1, -1]
958
    image_shape = [1, 1, 2]
959
960
961
962
963
964
965
    crop_size = [2, 2]

    image = tf.constant(1, dtype=tf.float32, shape=image_shape)
    boxes = tf.constant([[0, 0, 1, 1]], dtype=tf.float32)

    with self.assertRaisesRegexp(ValueError, 'num_spatial_bins should be >= 1'):
      ops.position_sensitive_crop_regions(
966
          image, boxes, crop_size, num_spatial_bins, global_pool=True)
967
968
969

  def test_raise_value_error_on_non_divisible_crop_size(self):
    num_spatial_bins = [2, 3]
970
    image_shape = [1, 1, 6]
971
972
973
974
975
976
977
978
    crop_size = [3, 2]

    image = tf.constant(1, dtype=tf.float32, shape=image_shape)
    boxes = tf.constant([[0, 0, 1, 1]], dtype=tf.float32)

    with self.assertRaisesRegexp(
        ValueError, 'crop_size should be divisible by num_spatial_bins'):
      ops.position_sensitive_crop_regions(
979
          image, boxes, crop_size, num_spatial_bins, global_pool=True)
980
981
982

  def test_raise_value_error_on_non_divisible_num_channels(self):
    num_spatial_bins = [2, 2]
983
    image_shape = [1, 1, 5]
984
985
    crop_size = [2, 2]

986
987
988
989
990
991
    def graph_fn():
      image = tf.constant(1, dtype=tf.float32, shape=image_shape)
      boxes = tf.constant([[0, 0, 1, 1]], dtype=tf.float32)

      return ops.position_sensitive_crop_regions(
          image, boxes, crop_size, num_spatial_bins, global_pool=True)
992
993
994

    with self.assertRaisesRegexp(
        ValueError, 'Dimension size must be evenly divisible by 4 but is 5'):
995
      self.execute(graph_fn, [])
996
997
998

  def test_position_sensitive_with_global_pool_false(self):
    num_spatial_bins = [3, 2]
999
    image_shape = [3, 2, 6]
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    num_boxes = 2

    expected_output = []

    # Expected output, when crop_size = [3, 2].
    expected_output.append(np.expand_dims(
        np.tile(np.array([[1, 2],
                          [3, 4],
                          [5, 6]]), (num_boxes, 1, 1)),
        axis=-1))

    # Expected output, when crop_size = [6, 4].
    expected_output.append(np.expand_dims(
        np.tile(np.array([[1, 1, 2, 2],
                          [1, 1, 2, 2],
                          [3, 3, 4, 4],
                          [3, 3, 4, 4],
                          [5, 5, 6, 6],
                          [5, 5, 6, 6]]), (num_boxes, 1, 1)),
        axis=-1))

    for crop_size_mult in range(1, 3):
      crop_size = [3 * crop_size_mult, 2 * crop_size_mult]
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
      # First channel is 1's, second channel is 2's, etc.

      def graph_fn():
        # pylint:disable=cell-var-from-loop
        image = tf.constant(
            list(range(1, 3 * 2 + 1)) * 6, dtype=tf.float32, shape=image_shape)
        boxes = tf.random_uniform((num_boxes, 4))

        ps_crop = ops.position_sensitive_crop_regions(
            image, boxes, crop_size, num_spatial_bins, global_pool=False)
        return ps_crop

      output = self.execute(graph_fn, [])
1036
      self.assertAllClose(output, expected_output[crop_size_mult - 1])
1037
1038
1039

  def test_position_sensitive_with_global_pool_false_and_do_global_pool(self):
    num_spatial_bins = [3, 2]
1040
    image_shape = [3, 2, 6]
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
    num_boxes = 2

    expected_output = []

    # Expected output, when crop_size = [3, 2].
    expected_output.append(np.mean(
        np.expand_dims(
            np.tile(np.array([[1, 2],
                              [3, 4],
                              [5, 6]]), (num_boxes, 1, 1)),
            axis=-1),
        axis=(1, 2), keepdims=True))

    # Expected output, when crop_size = [6, 4].
    expected_output.append(np.mean(
        np.expand_dims(
            np.tile(np.array([[1, 1, 2, 2],
                              [1, 1, 2, 2],
                              [3, 3, 4, 4],
                              [3, 3, 4, 4],
                              [5, 5, 6, 6],
                              [5, 5, 6, 6]]), (num_boxes, 1, 1)),
            axis=-1),
        axis=(1, 2), keepdims=True))

    for crop_size_mult in range(1, 3):
      crop_size = [3 * crop_size_mult, 2 * crop_size_mult]

1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
      def graph_fn():
        # pylint:disable=cell-var-from-loop
        # First channel is 1's, second channel is 2's, etc.
        image = tf.constant(
            list(range(1, 3 * 2 + 1)) * 6, dtype=tf.float32, shape=image_shape)
        boxes = tf.random_uniform((num_boxes, 4))

        # Perform global_pooling after running the function with
        # global_pool=False.
        ps_crop = ops.position_sensitive_crop_regions(
            image, boxes, crop_size, num_spatial_bins, global_pool=False)
        ps_crop_and_pool = tf.reduce_mean(
            ps_crop, reduction_indices=(1, 2), keepdims=True)
        return ps_crop_and_pool

      output = self.execute(graph_fn, [])
1085
1086
1087
1088
      self.assertAllEqual(output, expected_output[crop_size_mult - 1])

  def test_raise_value_error_on_non_square_block_size(self):
    num_spatial_bins = [3, 2]
1089
    image_shape = [3, 2, 6]
1090
1091
1092
1093
1094
1095
1096
1097
    crop_size = [6, 2]

    image = tf.constant(1, dtype=tf.float32, shape=image_shape)
    boxes = tf.constant([[0, 0, 1, 1]], dtype=tf.float32)

    with self.assertRaisesRegexp(
        ValueError, 'Only support square bin crop size for now.'):
      ops.position_sensitive_crop_regions(
1098
1099
1100
          image, boxes, crop_size, num_spatial_bins, global_pool=False)


1101
class OpsTestBatchPositionSensitiveCropRegions(test_case.TestCase):
1102
1103
1104
1105
1106
1107

  def test_position_sensitive_with_single_bin(self):
    num_spatial_bins = [1, 1]
    image_shape = [2, 3, 3, 4]
    crop_size = [2, 2]

1108
1109
1110
1111
    def graph_fn():
      image = tf.random_uniform(image_shape)
      boxes = tf.random_uniform((2, 3, 4))
      box_ind = tf.constant([0, 0, 0, 1, 1, 1], dtype=tf.int32)
1112

1113
1114
1115
1116
1117
1118
1119
      # When a single bin is used, position-sensitive crop and pool should be
      # the same as non-position sensitive crop and pool.
      crop = tf.image.crop_and_resize(image,
                                      tf.reshape(boxes, [-1, 4]), box_ind,
                                      crop_size)
      crop_and_pool = tf.reduce_mean(crop, [1, 2], keepdims=True)
      crop_and_pool = tf.reshape(crop_and_pool, [2, 3, 1, 1, 4])
1120

1121
1122
1123
      ps_crop_and_pool = ops.batch_position_sensitive_crop_regions(
          image, boxes, crop_size, num_spatial_bins, global_pool=True)
      return crop_and_pool, ps_crop_and_pool
1124

1125
1126
1127
    # Crop and resize is not supported on TPUs.
    expected_output, output = self.execute_cpu(graph_fn, [])
    self.assertAllClose(output, expected_output)
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152

  def test_position_sensitive_with_global_pool_false_and_known_boxes(self):
    num_spatial_bins = [2, 2]
    image_shape = [2, 2, 2, 4]
    crop_size = [2, 2]

    # box_ind = tf.constant([0, 1], dtype=tf.int32)

    expected_output = []

    # Expected output, when the box containing whole image.
    expected_output.append(
        np.reshape(np.array([[4, 7],
                             [10, 13]]),
                   (1, 2, 2, 1))
    )

    # Expected output, when the box containing only first row.
    expected_output.append(
        np.reshape(np.array([[3, 6],
                             [7, 10]]),
                   (1, 2, 2, 1))
    )
    expected_output = np.stack(expected_output, axis=0)

1153
1154
1155
1156
    def graph_fn():
      images = tf.constant(
          list(range(1, 2 * 2 * 4 + 1)) * 2, dtype=tf.float32,
          shape=image_shape)
1157

1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
      # First box contains whole image, and second box contains only first row.
      boxes = tf.constant(np.array([[[0., 0., 1., 1.]],
                                    [[0., 0., 0.5, 1.]]]), dtype=tf.float32)

      ps_crop = ops.batch_position_sensitive_crop_regions(
          images, boxes, crop_size, num_spatial_bins, global_pool=False)
      return ps_crop

    output = self.execute(graph_fn, [])
    self.assertAllEqual(output, expected_output)
1168
1169
1170
1171
1172
1173

  def test_position_sensitive_with_global_pool_false_and_single_bin(self):
    num_spatial_bins = [1, 1]
    image_shape = [2, 3, 3, 4]
    crop_size = [1, 1]

1174
1175
1176
1177
    def graph_fn():
      images = tf.random_uniform(image_shape)
      boxes = tf.random_uniform((2, 3, 4))
      # box_ind = tf.constant([0, 0, 0, 1, 1, 1], dtype=tf.int32)
1178

1179
1180
1181
1182
1183
1184
1185
      # Since single_bin is used and crop_size = [1, 1] (i.e., no crop resize),
      # the outputs are the same whatever the global_pool value is.
      ps_crop_and_pool = ops.batch_position_sensitive_crop_regions(
          images, boxes, crop_size, num_spatial_bins, global_pool=True)
      ps_crop = ops.batch_position_sensitive_crop_regions(
          images, boxes, crop_size, num_spatial_bins, global_pool=False)
      return ps_crop_and_pool, ps_crop
1186

1187
1188
    pooled_output, unpooled_output = self.execute(graph_fn, [])
    self.assertAllClose(pooled_output, unpooled_output)
1189
1190


1191
1192
1193
# The following tests are only executed on CPU because the output
# shape is not constant.
class ReframeBoxMasksToImageMasksTest(test_case.TestCase):
1194
1195
1196
1197
1198
1199

  def testZeroImageOnEmptyMask(self):
    np_expected_image_masks = np.array([[[0, 0, 0, 0],
                                         [0, 0, 0, 0],
                                         [0, 0, 0, 0],
                                         [0, 0, 0, 0]]], dtype=np.float32)
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
    def graph_fn():
      box_masks = tf.constant([[[0, 0],
                                [0, 0]]], dtype=tf.float32)
      boxes = tf.constant([[0.0, 0.0, 1.0, 1.0]], dtype=tf.float32)
      image_masks = ops.reframe_box_masks_to_image_masks(box_masks, boxes,
                                                         image_height=4,
                                                         image_width=4)
      return image_masks

    np_image_masks = self.execute_cpu(graph_fn, [])
    self.assertAllClose(np_image_masks, np_expected_image_masks)
1211

1212
  def testZeroBoxMasks(self):
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236

    def graph_fn():
      box_masks = tf.zeros([0, 3, 3], dtype=tf.float32)
      boxes = tf.zeros([0, 4], dtype=tf.float32)
      image_masks = ops.reframe_box_masks_to_image_masks(box_masks, boxes,
                                                         image_height=4,
                                                         image_width=4)
      return image_masks

    np_image_masks = self.execute_cpu(graph_fn, [])
    self.assertAllEqual(np_image_masks.shape, np.array([0, 4, 4]))

  def testBoxWithZeroArea(self):

    def graph_fn():
      box_masks = tf.zeros([1, 3, 3], dtype=tf.float32)
      boxes = tf.constant([[0.1, 0.2, 0.1, 0.7]], dtype=tf.float32)
      image_masks = ops.reframe_box_masks_to_image_masks(box_masks, boxes,
                                                         image_height=4,
                                                         image_width=4)
      return image_masks

    np_image_masks = self.execute_cpu(graph_fn, [])
    self.assertAllEqual(np_image_masks.shape, np.array([1, 4, 4]))
1237

1238
  def testMaskIsCenteredInImageWhenBoxIsCentered(self):
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248

    def graph_fn():
      box_masks = tf.constant([[[1, 1],
                                [1, 1]]], dtype=tf.float32)
      boxes = tf.constant([[0.25, 0.25, 0.75, 0.75]], dtype=tf.float32)
      image_masks = ops.reframe_box_masks_to_image_masks(box_masks, boxes,
                                                         image_height=4,
                                                         image_width=4)
      return image_masks

1249
1250
1251
1252
    np_expected_image_masks = np.array([[[0, 0, 0, 0],
                                         [0, 1, 1, 0],
                                         [0, 1, 1, 0],
                                         [0, 0, 0, 0]]], dtype=np.float32)
1253
1254
    np_image_masks = self.execute_cpu(graph_fn, [])
    self.assertAllClose(np_image_masks, np_expected_image_masks)
1255
1256

  def testMaskOffCenterRemainsOffCenterInImage(self):
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266

    def graph_fn():
      box_masks = tf.constant([[[1, 0],
                                [0, 1]]], dtype=tf.float32)
      boxes = tf.constant([[0.25, 0.5, 0.75, 1.0]], dtype=tf.float32)
      image_masks = ops.reframe_box_masks_to_image_masks(box_masks, boxes,
                                                         image_height=4,
                                                         image_width=4)
      return image_masks

1267
1268
1269
1270
    np_expected_image_masks = np.array([[[0, 0, 0, 0],
                                         [0, 0, 0.6111111, 0.16666669],
                                         [0, 0, 0.3888889, 0.83333337],
                                         [0, 0, 0, 0]]], dtype=np.float32)
1271
1272
    np_image_masks = self.execute_cpu(graph_fn, [])
    self.assertAllClose(np_image_masks, np_expected_image_masks)
1273
1274


1275
class MergeBoxesWithMultipleLabelsTest(test_case.TestCase):
1276
1277

  def testMergeBoxesWithMultipleLabels(self):
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293

    def graph_fn():
      boxes = tf.constant(
          [[0.25, 0.25, 0.75, 0.75], [0.0, 0.0, 0.5, 0.75],
           [0.25, 0.25, 0.75, 0.75]],
          dtype=tf.float32)
      class_indices = tf.constant([0, 4, 2], dtype=tf.int32)
      class_confidences = tf.constant([0.8, 0.2, 0.1], dtype=tf.float32)
      num_classes = 5
      merged_boxes, merged_classes, merged_confidences, merged_box_indices = (
          ops.merge_boxes_with_multiple_labels(
              boxes, class_indices, class_confidences, num_classes))

      return (merged_boxes, merged_classes, merged_confidences,
              merged_box_indices)

1294
1295
1296
1297
    expected_merged_boxes = np.array(
        [[0.25, 0.25, 0.75, 0.75], [0.0, 0.0, 0.5, 0.75]], dtype=np.float32)
    expected_merged_classes = np.array(
        [[1, 0, 1, 0, 0], [0, 0, 0, 0, 1]], dtype=np.int32)
1298
1299
    expected_merged_confidences = np.array(
        [[0.8, 0, 0.1, 0, 0], [0, 0, 0, 0, 0.2]], dtype=np.float32)
1300
    expected_merged_box_indices = np.array([0, 1], dtype=np.int32)
1301
1302
1303
1304
1305
1306
1307
1308

    # Running on CPU only as tf.unique is not supported on TPU.
    (np_merged_boxes, np_merged_classes, np_merged_confidences,
     np_merged_box_indices) = self.execute_cpu(graph_fn, [])
    self.assertAllClose(np_merged_boxes, expected_merged_boxes)
    self.assertAllClose(np_merged_classes, expected_merged_classes)
    self.assertAllClose(np_merged_confidences, expected_merged_confidences)
    self.assertAllClose(np_merged_box_indices, expected_merged_box_indices)
1309
1310

  def testMergeBoxesWithMultipleLabelsCornerCase(self):
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326

    def graph_fn():
      boxes = tf.constant(
          [[0, 0, 1, 1], [0, 1, 1, 1], [1, 0, 1, 1], [1, 1, 1, 1],
           [1, 1, 1, 1], [1, 0, 1, 1], [0, 1, 1, 1], [0, 0, 1, 1]],
          dtype=tf.float32)
      class_indices = tf.constant([0, 1, 2, 3, 2, 1, 0, 3], dtype=tf.int32)
      class_confidences = tf.constant([0.1, 0.9, 0.2, 0.8, 0.3, 0.7, 0.4, 0.6],
                                      dtype=tf.float32)
      num_classes = 4
      merged_boxes, merged_classes, merged_confidences, merged_box_indices = (
          ops.merge_boxes_with_multiple_labels(
              boxes, class_indices, class_confidences, num_classes))
      return (merged_boxes, merged_classes, merged_confidences,
              merged_box_indices)

1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
    expected_merged_boxes = np.array(
        [[0, 0, 1, 1], [0, 1, 1, 1], [1, 0, 1, 1], [1, 1, 1, 1]],
        dtype=np.float32)
    expected_merged_classes = np.array(
        [[1, 0, 0, 1], [1, 1, 0, 0], [0, 1, 1, 0], [0, 0, 1, 1]],
        dtype=np.int32)
    expected_merged_confidences = np.array(
        [[0.1, 0, 0, 0.6], [0.4, 0.9, 0, 0],
         [0, 0.7, 0.2, 0], [0, 0, 0.3, 0.8]], dtype=np.float32)
    expected_merged_box_indices = np.array([0, 1, 2, 3], dtype=np.int32)
1337
1338
1339
1340
1341
1342
1343
1344
1345

    # Running on CPU only as tf.unique is not supported on TPU.
    (np_merged_boxes, np_merged_classes, np_merged_confidences,
     np_merged_box_indices) = self.execute_cpu(graph_fn, [])

    self.assertAllClose(np_merged_boxes, expected_merged_boxes)
    self.assertAllClose(np_merged_classes, expected_merged_classes)
    self.assertAllClose(np_merged_confidences, expected_merged_confidences)
    self.assertAllClose(np_merged_box_indices, expected_merged_box_indices)
1346
1347

  def testMergeBoxesWithEmptyInputs(self):
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366

    def graph_fn():
      boxes = tf.zeros([0, 4], dtype=tf.float32)
      class_indices = tf.constant([], dtype=tf.int32)
      class_confidences = tf.constant([], dtype=tf.float32)
      num_classes = 5
      merged_boxes, merged_classes, merged_confidences, merged_box_indices = (
          ops.merge_boxes_with_multiple_labels(
              boxes, class_indices, class_confidences, num_classes))
      return (merged_boxes, merged_classes, merged_confidences,
              merged_box_indices)

    # Running on CPU only as tf.unique is not supported on TPU.
    (np_merged_boxes, np_merged_classes, np_merged_confidences,
     np_merged_box_indices) = self.execute_cpu(graph_fn, [])
    self.assertAllEqual(np_merged_boxes.shape, [0, 4])
    self.assertAllEqual(np_merged_classes.shape, [0, 5])
    self.assertAllEqual(np_merged_confidences.shape, [0, 5])
    self.assertAllEqual(np_merged_box_indices.shape, [0])
1367

1368
  def testMergeBoxesWithMultipleLabelsUsesInt64(self):
1369
1370
1371
1372

    if self.is_tf2():
      self.skipTest('Getting op names is not supported in eager mode.')

1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
    boxes = tf.constant(
        [[0.25, 0.25, 0.75, 0.75], [0.0, 0.0, 0.5, 0.75],
         [0.25, 0.25, 0.75, 0.75]],
        dtype=tf.float32)
    class_indices = tf.constant([0, 4, 2], dtype=tf.int32)
    class_confidences = tf.constant([0.8, 0.2, 0.1], dtype=tf.float32)
    num_classes = 5
    ops.merge_boxes_with_multiple_labels(
        boxes, class_indices, class_confidences, num_classes)

    graph = tf.get_default_graph()

    def assert_dtype_is_int64(op_name):
      op = graph.get_operation_by_name(op_name)
      self.assertEqual(op.get_attr('dtype'), tf.int64)

    def assert_t_is_int64(op_name):
      op = graph.get_operation_by_name(op_name)
      self.assertEqual(op.get_attr('T'), tf.int64)

    assert_dtype_is_int64('map/TensorArray')
    assert_dtype_is_int64('map/TensorArray_1')
    assert_dtype_is_int64('map/while/TensorArrayReadV3')
    assert_t_is_int64('map/while/TensorArrayWrite/TensorArrayWriteV3')
    assert_t_is_int64(
        'map/TensorArrayUnstack/TensorArrayScatter/TensorArrayScatterV3')
    assert_dtype_is_int64('map/TensorArrayStack/TensorArrayGatherV3')

1401

1402
1403
class NearestNeighborUpsamplingTest(test_case.TestCase):

1404
  def test_upsampling_with_single_scale(self):
1405
1406
1407

    def graph_fn(inputs):
      custom_op_output = ops.nearest_neighbor_upsampling(inputs, scale=2)
1408
1409
1410
1411
1412
1413
1414
1415
1416
      return custom_op_output
    inputs = np.reshape(np.arange(4).astype(np.float32), [1, 2, 2, 1])
    custom_op_output = self.execute(graph_fn, [inputs])

    expected_output = [[[[0], [0], [1], [1]],
                        [[0], [0], [1], [1]],
                        [[2], [2], [3], [3]],
                        [[2], [2], [3], [3]]]]
    self.assertAllClose(custom_op_output, expected_output)
1417

1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
  def test_upsampling_with_separate_height_width_scales(self):

    def graph_fn(inputs):
      custom_op_output = ops.nearest_neighbor_upsampling(inputs,
                                                         height_scale=2,
                                                         width_scale=3)
      return custom_op_output
    inputs = np.reshape(np.arange(4).astype(np.float32), [1, 2, 2, 1])
    custom_op_output = self.execute(graph_fn, [inputs])

    expected_output = [[[[0], [0], [0], [1], [1], [1]],
                        [[0], [0], [0], [1], [1], [1]],
                        [[2], [2], [2], [3], [3], [3]],
                        [[2], [2], [2], [3], [3], [3]]]]
    self.assertAllClose(custom_op_output, expected_output)

1434

1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
class MatmulGatherOnZerothAxis(test_case.TestCase):

  def test_gather_2d(self):

    def graph_fn(params, indices):
      return ops.matmul_gather_on_zeroth_axis(params, indices)

    params = np.array([[1, 2, 3, 4],
                       [5, 6, 7, 8],
                       [9, 10, 11, 12],
                       [0, 1, 0, 0]], dtype=np.float32)
1446
    indices = np.array([2, 2, 1], dtype=np.int32)
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
    expected_output = np.array([[9, 10, 11, 12], [9, 10, 11, 12], [5, 6, 7, 8]])
    gather_output = self.execute(graph_fn, [params, indices])
    self.assertAllClose(gather_output, expected_output)

  def test_gather_3d(self):

    def graph_fn(params, indices):
      return ops.matmul_gather_on_zeroth_axis(params, indices)

    params = np.array([[[1, 2], [3, 4]],
                       [[5, 6], [7, 8]],
                       [[9, 10], [11, 12]],
                       [[0, 1], [0, 0]]], dtype=np.float32)
1460
    indices = np.array([0, 3, 1], dtype=np.int32)
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
    expected_output = np.array([[[1, 2], [3, 4]],
                                [[0, 1], [0, 0]],
                                [[5, 6], [7, 8]]])
    gather_output = self.execute(graph_fn, [params, indices])
    self.assertAllClose(gather_output, expected_output)

  def test_gather_with_many_indices(self):

    def graph_fn(params, indices):
      return ops.matmul_gather_on_zeroth_axis(params, indices)

    params = np.array([[1, 2, 3, 4],
                       [5, 6, 7, 8],
                       [9, 10, 11, 12],
                       [0, 1, 0, 0]], dtype=np.float32)
1476
    indices = np.array([0, 0, 0, 0, 0, 0], dtype=np.int32)
1477
1478
1479
1480
    expected_output = np.array(6*[[1, 2, 3, 4]])
    gather_output = self.execute(graph_fn, [params, indices])
    self.assertAllClose(gather_output, expected_output)

1481
  def test_gather_with_dynamic_shape_input(self):
1482
1483
1484
1485

    def graph_fn(params, indices):
      return ops.matmul_gather_on_zeroth_axis(params, indices)

1486
1487
1488
1489
1490
1491
    params = np.array([[1, 2, 3, 4],
                       [5, 6, 7, 8],
                       [9, 10, 11, 12],
                       [0, 1, 0, 0]], dtype=np.float32)
    indices = np.array([0, 0, 0, 0, 0, 0])
    expected_output = np.array(6*[[1, 2, 3, 4]])
1492
1493
    gather_output = self.execute(graph_fn, [params, indices])
    self.assertAllClose(gather_output, expected_output)
1494

1495

1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
class FpnFeatureLevelsTest(test_case.TestCase):

  def test_correct_fpn_levels(self):
    image_size = 640
    pretraininig_image_size = 224
    image_ratio = image_size * 1.0 / pretraininig_image_size
    boxes = np.array(
        [
            [
                [0, 0, 111, 111],  # Level 0.
                [0, 0, 113, 113],  # Level 1.
                [0, 0, 223, 223],  # Level 1.
                [0, 0, 225, 225],  # Level 2.
                [0, 0, 449, 449]   # Level 3.
            ],
        ],
        dtype=np.float32) / image_size

    def graph_fn(boxes):
      return ops.fpn_feature_levels(
          num_levels=5, unit_scale_index=2, image_ratio=image_ratio,
          boxes=boxes)

    levels = self.execute(graph_fn, [boxes])
    self.assertAllEqual([[0, 1, 1, 2, 3]], levels)
1521
1522


1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
class TestBfloat16ToFloat32(test_case.TestCase):

  def test_convert_list(self):
    var_list = [
        tf.constant([1.], dtype=tf.bfloat16),
        tf.constant([2], dtype=tf.int32)
    ]
    casted_var_list = ops.bfloat16_to_float32_nested(var_list)
    self.assertEqual(casted_var_list[0].dtype, tf.float32)
    self.assertEqual(casted_var_list[1].dtype, tf.int32)

  def test_convert_tensor_dict(self):
    tensor_dict = {
        'key1': tf.constant([1.], dtype=tf.bfloat16),
        'key2': [
            tf.constant([0.5], dtype=tf.bfloat16),
            tf.constant([7], dtype=tf.int32),
        ],
        'key3': tf.constant([2], dtype=tf.uint8),
    }
    tensor_dict = ops.bfloat16_to_float32_nested(tensor_dict)

    self.assertEqual(tensor_dict['key1'].dtype, tf.float32)
    self.assertEqual(tensor_dict['key2'][0].dtype, tf.float32)
    self.assertEqual(tensor_dict['key2'][1].dtype, tf.int32)
    self.assertEqual(tensor_dict['key3'].dtype, tf.uint8)


class TestGatherWithPaddingValues(test_case.TestCase):

  def test_gather_with_padding_values(self):
    expected_gathered_tensor = [
        [0, 0, 0.2, 0.2],
        [0, 0, 0, 0],
        [0, 0, 0.1, 0.1],
        [0, 0, 0, 0],
    ]
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574

    def graph_fn():
      indices = tf.constant([1, -1, 0, -1])
      input_tensor = tf.constant([[0, 0, 0.1, 0.1], [0, 0, 0.2, 0.2]],
                                 dtype=tf.float32)

      gathered_tensor = ops.gather_with_padding_values(
          input_tensor,
          indices=indices,
          padding_value=tf.zeros_like(input_tensor[0]))
      self.assertEqual(gathered_tensor.dtype, tf.float32)

      return gathered_tensor

    gathered_tensor_np = self.execute(graph_fn, [])
1575
1576
1577
    self.assertAllClose(expected_gathered_tensor, gathered_tensor_np)


1578
1579
1580



1581
1582


1583
1584
if __name__ == '__main__':
  tf.test.main()