def insert_transformed_feature(self, columns_to_tensors):
# Bucketize the source column.
if self.source_column not in columns_to_tensors:
self.source_column.insert_transformed_feature(columns_to_tensors)
columns_to_tensors[self] = bucketization_op.bucketize(
columns_to_tensors[self.source_column],
boundaries=list(self.boundaries))
def test_normal_usecase(self):
op = bucketization_op.bucketize(
constant_op.constant([-5, 0, 2, 3, 5, 8, 10, 11, 12]),
boundaries=[0, 3, 8, 11])
expected_out = [0, 1, 1, 2, 2, 3, 3, 4, 4]
with self.test_session() as sess:
self.assertAllEqual(expected_out, sess.run(op))
def insert_transformed_feature(self, columns_to_tensors):
# Bucketize the source column.
if self.source_column not in columns_to_tensors:
self.source_column.insert_transformed_feature(columns_to_tensors)
columns_to_tensors[self] = bucketization_op.bucketize(
columns_to_tensors[self.source_column],
boundaries=list(self.boundaries))
def _bucketize(instances, feature, schema, metadata):
"""Applies the bucketize transform to a numeric field.
"""
field = schema[feature.field]
if not field.numeric:
raise ValueError(
'A scale transform cannot be applied to non-numerical field "%s".'
% feature.field)
transform = feature.transform
boundaries = map(float, transform['boundaries'].split(','))
# TODO: Figure out how to use tf.case instead of this contrib op
from tensorflow.contrib.layers.python.ops.bucketization_op import bucketize
# Create a one-hot encoded tensor. The dimension of this tensor is the set of buckets defined
# by N boundaries == N + 1.
# A squeeze is needed to remove the extra dimension added to the shape.
value = instances[feature.field]
value = tf.squeeze(tf.one_hot(bucketize(value, boundaries, name='bucket'),
depth=len(boundaries) + 1,
on_value=1.0,
off_value=0.0,
name='one_hot'),
axis=1,
name='bucketize')
value.set_shape((None, len(boundaries) + 1))
return value
def test_normal_usecase(self):
op = bucketization_op.bucketize(constant_op.constant(
[-5, 0, 2, 3, 5, 8, 10, 11, 12]),
boundaries=[0, 3, 8, 11])
expected_out = [0, 1, 1, 2, 2, 3, 3, 4, 4]
with self.test_session() as sess:
self.assertAllEqual(expected_out, sess.run(op))
def _transform_feature(self, inputs):
"""Handles cross transformation."""
# Bucketize the source column.
if not self.add_random:
return bucketization_op.bucketize(inputs.get(self.source_column),
boundaries=list(self.boundaries),
name="bucketize")
else:
rawts = inputs.get(self.source_column)
tbn = np.asarray(self.boundaries[1:])
if len(tbn) > 30:
# noise = min(np.median(tbn)-tbn[0],tbn[20:-20].std())/2.
noise = tbn[10:-10].std() / 10.
rndts = rawts + random_normal(array_ops.shape(rawts), 0, noise)
return bucketization_op.bucketize(rndts,
boundaries=list(
self.boundaries),
name="bucketize")
else:
return bucketization_op.bucketize(rawts,
boundaries=list(
self.boundaries),
name="bucketize")
def test_invalid_boundaries_order(self):
op = bucketization_op.bucketize(
constant_op.constant([-5, 0]), boundaries=[0, 8, 3, 11])
with self.test_session() as sess:
with self.assertRaises(errors_impl.InvalidArgumentError):
sess.run(op)
def test_invalid_boundaries_order(self):
op = bucketization_op.bucketize(constant_op.constant([-5, 0]),
boundaries=[0, 8, 3, 11])
with self.test_session() as sess:
with self.assertRaises(errors_impl.InvalidArgumentError):
sess.run(op)
