Merge pull request #10626 from ryan0507:master

PiperOrigin-RevId: 461948867

official/projects/yt8m/configs/yt8m_test.py

+# Copyright 2022 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.
+
+from absl.testing import parameterized
+import tensorflow as tf
+
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.modeling import hyperparams
+from official.projects.yt8m.configs import yt8m  # pylint: disable=unused-import
+from official.projects.yt8m.configs.yt8m import yt8m as exp_cfg
+
+
+class YT8MTest(tf.test.TestCase, parameterized.TestCase):
+
+  @parameterized.parameters(
+      ('yt8m_experiment',),)
+  def test_yt8m_configs(self, config_name):
+    config = exp_factory.get_exp_config(config_name)
+    self.assertIsInstance(config, cfg.ExperimentConfig)
+    self.assertIsInstance(config.task, cfg.TaskConfig)
+    self.assertIsInstance(config.task.model, hyperparams.Config)
+    self.assertIsInstance(config.task.train_data, cfg.DataConfig)
+    config.task.train_data.is_training = None
+    with self.assertRaises(KeyError):
+      config.validate()
+
+if __name__ == '__main__':
+  tf.test.main()

official/projects/yt8m/dataloaders/utils.py

@@ -20,7 +20,7 @@ import tensorflow as tf
 from official.vision.dataloaders import tfexample_utils
 
 
-def Dequantize(feat_vector, max_quantized_value=2, min_quantized_value=-2):
+def dequantize(feat_vector, max_quantized_value=2, min_quantized_value=-2):
   """Dequantize the feature from the byte format to the float format.
 
   Args:
@@ -38,7 +38,7 @@ def Dequantize(feat_vector, max_quantized_value=2, min_quantized_value=-2):
   return feat_vector * scalar + bias
 
 
-def MakeSummary(name, value):
+def make_summary(name, value):
   """Creates a tf.Summary proto with the given name and value."""
   summary = tf.Summary()
   val = summary.value.add()
@@ -47,7 +47,7 @@ def MakeSummary(name, value):
   return summary
 
 
-def AddGlobalStepSummary(summary_writer,
+def add_global_step_summary(summary_writer,
                             global_step_val,
                             global_step_info_dict,
                             summary_scope="Eval"):
@@ -69,18 +69,18 @@ def AddGlobalStepSummary(summary_writer,
   examples_per_second = global_step_info_dict.get("examples_per_second", -1)
 
   summary_writer.add_summary(
-      MakeSummary("GlobalStep/" + summary_scope + "_Hit@1", this_hit_at_one),
+      make_summary("GlobalStep/" + summary_scope + "_Hit@1", this_hit_at_one),
       global_step_val)
   summary_writer.add_summary(
-      MakeSummary("GlobalStep/" + summary_scope + "_Perr", this_perr),
+      make_summary("GlobalStep/" + summary_scope + "_Perr", this_perr),
       global_step_val)
   summary_writer.add_summary(
-      MakeSummary("GlobalStep/" + summary_scope + "_Loss", this_loss),
+      make_summary("GlobalStep/" + summary_scope + "_Loss", this_loss),
       global_step_val)
 
   if examples_per_second != -1:
     summary_writer.add_summary(
-        MakeSummary("GlobalStep/" + summary_scope + "_Example_Second",
+        make_summary("GlobalStep/" + summary_scope + "_Example_Second",
                      examples_per_second), global_step_val)
 
   summary_writer.flush()
@@ -92,7 +92,7 @@ def AddGlobalStepSummary(summary_writer,
   return info
 
 
-def AddEpochSummary(summary_writer,
+def add_epoch_summary(summary_writer,
                       global_step_val,
                       epoch_info_dict,
                       summary_scope="Eval"):
@@ -117,18 +117,18 @@ def AddEpochSummary(summary_writer,
   mean_ap = np.mean(aps)
 
   summary_writer.add_summary(
-      MakeSummary("Epoch/" + summary_scope + "_Avg_Hit@1", avg_hit_at_one),
+      make_summary("Epoch/" + summary_scope + "_Avg_Hit@1", avg_hit_at_one),
       global_step_val)
   summary_writer.add_summary(
-      MakeSummary("Epoch/" + summary_scope + "_Avg_Perr", avg_perr),
+      make_summary("Epoch/" + summary_scope + "_Avg_Perr", avg_perr),
       global_step_val)
   summary_writer.add_summary(
-      MakeSummary("Epoch/" + summary_scope + "_Avg_Loss", avg_loss),
+      make_summary("Epoch/" + summary_scope + "_Avg_Loss", avg_loss),
       global_step_val)
   summary_writer.add_summary(
-      MakeSummary("Epoch/" + summary_scope + "_MAP", mean_ap), global_step_val)
+      make_summary("Epoch/" + summary_scope + "_MAP", mean_ap), global_step_val)
   summary_writer.add_summary(
-      MakeSummary("Epoch/" + summary_scope + "_GAP", gap), global_step_val)
+      make_summary("Epoch/" + summary_scope + "_GAP", gap), global_step_val)
   summary_writer.flush()
 
   info = ("epoch/eval number {0} | Avg_Hit@1: {1:.3f} | Avg_PERR: {2:.3f} "
@@ -138,7 +138,7 @@ def AddEpochSummary(summary_writer,
   return info
 
 
-def GetListOfFeatureNamesAndSizes(feature_names, feature_sizes):
+def get_list_of_feature_names_and_sizes(feature_names, feature_sizes):
   """Extract the list of feature names and the dimensionality.
 
   Args:
@@ -164,59 +164,7 @@ def GetListOfFeatureNamesAndSizes(feature_names, feature_sizes):
   return list_of_feature_names, list_of_feature_sizes
 
 
-def ClipGradientNorms(gradients_to_variables, max_norm):
-  """Clips the gradients by the given value.
-
-  Args:
-    gradients_to_variables: A list of gradient to variable pairs (tuples).
-    max_norm: the maximum norm value.
-
-  Returns:
-    A list of clipped gradient to variable pairs.
-  """
-  clipped_grads_and_vars = []
-  for grad, var in gradients_to_variables:
-    if grad is not None:
-      if isinstance(grad, tf.IndexedSlices):
-        tmp = tf.clip_by_norm(grad.values, max_norm)
-        grad = tf.IndexedSlices(tmp, grad.indices, grad.dense_shape)
-      else:
-        grad = tf.clip_by_norm(grad, max_norm)
-    clipped_grads_and_vars.append((grad, var))
-  return clipped_grads_and_vars
-
-
-def CombineGradients(tower_grads):
-  """Calculate the combined gradient for each shared variable across all towers.
-
-  Note that this function provides a synchronization point across all towers.
-
-  Args:
-    tower_grads: List of lists of (gradient, variable) tuples. The outer list is
-      over individual gradients. The inner list is over the gradient calculation
-      for each tower.
-
-  Returns:
-     List of pairs of (gradient, variable) where the gradient has been summed
-     across all towers.
-  """
-  filtered_grads = [
-      [x for x in grad_list if x[0] is not None] for grad_list in tower_grads
-  ]
-  final_grads = []
-  for i in range(len(filtered_grads[0])):
-    grads = [filtered_grads[t][i] for t in range(len(filtered_grads))]
-    grad = tf.stack([x[0] for x in grads], 0)
-    grad = tf.reduce_sum(grad, 0)
-    final_grads.append((
-        grad,
-        filtered_grads[0][i][1],
-    ))
-
-  return final_grads
-
-
-def MakeYt8mExample(num_segment: int = 5) -> tf.train.SequenceExample:
+def make_yt8m_example(num_segment: int = 5) -> tf.train.SequenceExample:
   """Generate fake data for unit tests."""
   rgb = np.random.randint(low=256, size=1024, dtype=np.uint8)
   audio = np.random.randint(low=256, size=128, dtype=np.uint8)
@@ -240,7 +188,7 @@ def MakeYt8mExample(num_segment: int = 5) -> tf.train.SequenceExample:
 
 
 # TODO(yeqing): Move the test related functions to test_utils.
-def MakeExampleWithFloatFeatures(
+def make_example_with_float_features(
     num_segment: int = 5) -> tf.train.SequenceExample:
   """Generate fake data for unit tests."""
   rgb = np.random.rand(1, 2048).astype(np.float32)

official/projects/yt8m/dataloaders/yt8m_input.py

@@ -175,7 +175,7 @@ def _get_video_matrix(features, feature_size, dtype, max_frames,
 
   num_frames = tf.math.minimum(tf.shape(decoded_features)[0], max_frames)
   if dtype.is_integer:
-    feature_matrix = utils.Dequantize(decoded_features, max_quantized_value,
+    feature_matrix = utils.dequantize(decoded_features, max_quantized_value,
                                       min_quantized_value)
   else:
     feature_matrix = decoded_features

official/projects/yt8m/dataloaders/yt8m_input_test.py

@@ -37,7 +37,7 @@ class Yt8mInputTest(parameterized.TestCase, tf.test.TestCase):
     tf.io.gfile.makedirs(data_dir)
     self.data_path = os.path.join(data_dir, 'data.tfrecord')
     self.num_segment = 6
-    examples = [utils.MakeYt8mExample(self.num_segment) for _ in range(8)]
+    examples = [utils.make_yt8m_example(self.num_segment) for _ in range(8)]
     tfexample_utils.dump_to_tfrecord(self.data_path, tf_examples=examples)
 
   def create_input_reader(self, params):
@@ -130,7 +130,8 @@ class Yt8mInputTest(parameterized.TestCase, tf.test.TestCase):
     tf.io.gfile.makedirs(data_dir)
     data_path = os.path.join(data_dir, 'data2.tfrecord')
     examples = [
-        utils.MakeExampleWithFloatFeatures(self.num_segment) for _ in range(8)
+        utils.make_example_with_float_features(self.num_segment)
+        for _ in range(8)
     ]
     tfexample_utils.dump_to_tfrecord(data_path, tf_examples=examples)
 

official/projects/yt8m/eval_utils/average_precision_calculator.py

@@ -268,6 +268,5 @@ class AveragePrecisionCalculator(object):
       The normalized prediction.
     """
     denominator = numpy.max(predictions) - numpy.min(predictions)
-    ret = (predictions - numpy.min(predictions)) / numpy.max(
-        denominator, epsilon)
+    ret = (predictions - numpy.min(predictions)) / max(denominator, epsilon)
     return ret

official/projects/yt8m/eval_utils/eval_util_test.py

+# Copyright 2022 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.
+
+from absl import logging
+from absl.testing import parameterized
+import numpy as np
+import tensorflow as tf
+
+from official.projects.yt8m.eval_utils.average_precision_calculator import AveragePrecisionCalculator
+
+
+class YT8MAveragePrecisionCalculatorTest(parameterized.TestCase,
+                                         tf.test.TestCase):
+
+  def setUp(self):
+    super().setUp()
+    self.prediction = np.array([
+        [0.98, 0.88, 0.77, 0.65, 0.64, 0.59, 0.45, 0.43, 0.20, 0.05],
+        [0.878, 0.832, 0.759, 0.621, 0.458, 0.285, 0.134],
+        [0.98],
+        [0.56],
+    ])
+    self.raw_prediction = np.random.rand(5, 10) + np.random.randint(
+        low=0, high=10, size=(5, 10))
+    self.ground_truth = np.array([[1, 1, 0, 0, 0, 1, 1, 0, 0, 1],
+                                  [1, 0, 1, 0, 0, 1, 0], [1], [0]])
+
+    self.expected_ap = np.array([
+        0.714,
+        0.722,
+        1.000,
+        0.000,
+    ])
+
+  def test_ap_calculator_ap(self):
+
+    # Compare Expected Average Precision with function expected
+    for i, _ in enumerate(self.ground_truth):
+      calculator = AveragePrecisionCalculator()
+      ap = calculator.ap(self.prediction[i], self.ground_truth[i])
+      logging.info('DEBUG %dth AP: %r', i + 1, ap)
+
+  def test_ap_calculator_zero_one_normalize(self):
+    for i, _ in enumerate(self.raw_prediction):
+      calculator = AveragePrecisionCalculator()
+      logging.error('%r', self.raw_prediction[i])
+      normalized_score = calculator._zero_one_normalize(self.raw_prediction[i])
+      self.assertAllInRange(normalized_score, lower_bound=0.0, upper_bound=1.0)
+
+  @parameterized.parameters((None,), (3,), (5,), (10,), (20,))
+  def test_ap_calculator_ap_at_n(self, n):
+    for i, _ in enumerate(self.ground_truth):
+      calculator = AveragePrecisionCalculator(n)
+      ap = calculator.ap_at_n(self.prediction[i], self.ground_truth[i], n)
+      logging.info('DEBUG %dth AP: %r', i + 1, ap)
+
+
+if __name__ == '__main__':
+  tf.test.main()

official/projects/yt8m/modeling/yt8m_model.py

@@ -16,9 +16,10 @@
 from typing import Optional
 
 import tensorflow as tf
+
 from official.modeling import tf_utils
 from official.projects.yt8m.configs import yt8m as yt8m_cfg
-from official.projects.yt8m.modeling import yt8m_agg_models
+from official.projects.yt8m.modeling import nn_layers
 from official.projects.yt8m.modeling import yt8m_model_utils as utils
 
 layers = tf.keras.layers
@@ -171,7 +172,7 @@ class DbofModel(tf.keras.Model):
     activation = self._act_fn(activation)
     tf.summary.histogram("hidden1_output", activation)
 
-    aggregated_model = getattr(yt8m_agg_models,
+    aggregated_model = getattr(nn_layers,
                                params.yt8m_agg_classifier_model)
     norm_args = dict(axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)
     output = aggregated_model().create_model(

official/projects/yt8m/train_test.py

@@ -36,7 +36,7 @@ class TrainTest(parameterized.TestCase, tf.test.TestCase):
     data_dir = os.path.join(self.get_temp_dir(), 'data')
     tf.io.gfile.makedirs(data_dir)
     self._data_path = os.path.join(data_dir, 'data.tfrecord')
-    examples = [utils.MakeYt8mExample() for _ in range(8)]
+    examples = [utils.make_yt8m_example() for _ in range(8)]
     tfexample_utils.dump_to_tfrecord(self._data_path, tf_examples=examples)
 
   @parameterized.named_parameters(