def testMergeTensorValuesSparseOriginalsUnmodified(self):
value1 = tf.compat.v1.SparseTensorValue(
indices=np.array([]).reshape([0, 2]),
values=np.array([]).reshape([0, 1]),
dense_shape=np.array([1, 4]))
value2 = tf.compat.v1.SparseTensorValue(
indices=np.array([]).reshape([0, 2]),
values=np.array([]).reshape([0, 1]),
dense_shape=np.array([1, 4]))
merged_tensor_values = util.merge_tensor_values(
tensor_values=[value1, value2])
# Check that the original SparseTensorValues were not mutated.
self.assertSparseTensorValueEqual(
tf.compat.v1.SparseTensorValue(indices=np.array([]).reshape([0,
2]),
values=np.array([]).reshape([0, 1]),
dense_shape=np.array([1, 4])),
value1)
self.assertSparseTensorValueEqual(
tf.compat.v1.SparseTensorValue(indices=np.array([]).reshape([0,
2]),
values=np.array([]).reshape([0, 1]),
dense_shape=np.array([1, 4])),
value2)
# Check the merged SparseTensorValue.
self.assertSparseTensorValueEqual(
tf.compat.v1.SparseTensorValue(indices=np.array([]).reshape([0,
2]),
values=np.array([]).reshape([0, 1]),
dense_shape=np.array([2, 4])),
merged_tensor_values)
def testMergeTensorValueDense(self):
merged_tensor_values = util.merge_tensor_values(tensor_values=[
np.ndarray(shape=(1, 2), buffer=np.array([1, 2])),
np.ndarray(shape=(1, 2), buffer=np.array([3, 4])),
np.ndarray(shape=(1, 2), buffer=np.array([5, 6])),
])
self.assertAllEqual(
np.ndarray(shape=(3, 2), buffer=np.array([1, 2, 3, 4, 5, 6])),
merged_tensor_values)
def _create_feed_for_features_predictions_labels_list(
self, features_predictions_labels_list):
"""Create feed list for feeding a list of features, predictions, labels."""
result = []
for which_map, key, _ in self._iterate_fpl_maps_in_canonical_order():
result.append(
util.merge_tensor_values([
getattr(fpl, which_map)[key][encoding.NODE_SUFFIX]
for fpl in features_predictions_labels_list
]))
return result
def _create_feed_for_features_predictions_labels_list(
self, features_predictions_labels_list):
"""Create feed dict for feeding a list of features, predictions, labels."""
result = {}
for label_key, label_dict in self._labels_map.items():
result[label_dict[
encoding.NODE_SUFFIX]] = util.merge_tensor_values([
fpl.labels[label_key][encoding.NODE_SUFFIX]
for fpl in features_predictions_labels_list
])
for feature_key, feature_dict in self._features_map.items():
result[feature_dict[
encoding.NODE_SUFFIX]] = util.merge_tensor_values([
fpl.features[feature_key][encoding.NODE_SUFFIX]
for fpl in features_predictions_labels_list
])
for prediction_key, prediction_dict in self._predictions_map.items():
result[prediction_dict[encoding.NODE_SUFFIX]] = (
util.merge_tensor_values([
fpl.predictions[prediction_key][encoding.NODE_SUFFIX]
for fpl in features_predictions_labels_list
]))
return result
def testMergeTensorValueDenseDifferentShapesStrings(self):
merged_tensor_values = util.merge_tensor_values(tensor_values=[
np.array([[['apple'], ['banana']]]),
np.array([[['cherry', 'date', 'elderberry']]]),
np.array([[['fig', 'guava'], ['honeydew', 'imbe'],
['jackfruit', 'kiwi']]])
])
self.assertAllEqual(
np.array([
# Row 0
[['apple', '', ''], ['banana', '', ''], ['', '', '']],
# Row 1
[['cherry', 'date', 'elderberry'], ['', '', ''], ['', '', '']],
# Row 2
[['fig', 'guava', ''], ['honeydew', 'imbe', ''],
['jackfruit', 'kiwi', '']]
]),
merged_tensor_values)
def testMergeTensorValueSparse(self):
merged_tensor_values = util.merge_tensor_values(tensor_values=[
tf.SparseTensorValue(indices=np.array([[0, 0], [0, 1]]),
values=np.array([1, 2]),
dense_shape=np.array([1, 2])),
tf.SparseTensorValue(indices=np.array([[0, 0], [0, 1]]),
values=np.array([3, 4]),
dense_shape=np.array([1, 2])),
tf.SparseTensorValue(indices=np.array([[0, 0], [0, 1]]),
values=np.array([5, 6]),
dense_shape=np.array([1, 2])),
])
self.assertSparseTensorValueEqual(
tf.SparseTensorValue(indices=np.array([[0, 0], [0, 1], [1, 0],
[1, 1], [2, 0], [2, 1]]),
values=np.array([1, 2, 3, 4, 5, 6]),
dense_shape=np.array([3, 2])),
merged_tensor_values)
def _create_feed_for_features_predictions_labels_list(
self, features_predictions_labels_list):
"""Create feed list for a list of FeaturesPredictionsLabels."""
# Feed in the tensors in the following order:
# Features -> Predictions -> Labels and using the standard key ordering
# within each bucket. This should match the placeholder definitions when
# generating the graph.
# Note that we need to merge all examples into one Tensor before feeding.
tensor_feed = []
for which_map, key, _ in self._iterate_fpl_maps_in_canonical_order():
tensor_feed.append(
util.merge_tensor_values([
getattr(fpl, which_map)[key][encoding.NODE_SUFFIX]
for fpl in features_predictions_labels_list # pytype: disable=wrong-arg-types
]))
return tensor_feed
def testMergeTensorValueDenseDifferentShapesInts(self):
merged_tensor_values = util.merge_tensor_values(tensor_values=[
np.array([[[10], [11]]]),
np.array([[[20, 21, 22]]]),
np.array([[[30, 31], [32, 33], [34, 35]]]),
np.array([[[40, 41]]]),
])
self.assertAllEqual(
np.array([
# Row 0
[[10, 0, 0], [11, 0, 0], [0, 0, 0]],
# Row 1
[[20, 21, 22], [0, 0, 0], [0, 0, 0]],
# Row 2
[[30, 31, 0], [32, 33, 0], [34, 35, 0]],
# Row 3
[[40, 41, 0], [0, 0, 0], [0, 0, 0]],
]),
merged_tensor_values)
def testMergeTensorValueSparseDifferentShapes(self):
merged_tensor_values = util.merge_tensor_values(tensor_values=[
tf.SparseTensorValue(indices=np.array([[0, 0, 0], [0, 1, 1]]),
values=np.array([10, 12]),
dense_shape=np.array([1, 2, 2])),
tf.SparseTensorValue(indices=np.array([[0, 2, 2]]),
values=np.array([22]),
dense_shape=np.array([1, 3, 3])),
tf.SparseTensorValue(indices=np.array([[0, 0, 4]]),
values=np.array([33]),
dense_shape=np.array([1, 1, 5]))
])
self.assertSparseTensorValueEqual(
tf.SparseTensorValue(indices=np.array([[0, 0, 0], [0, 1, 1],
[1, 2, 2], [2, 0, 4]]),
values=np.array([10, 12, 22, 33]),
dense_shape=np.array([3, 3, 5])),
merged_tensor_values)
