def _normalize_feature(feature: schema_pb2.Feature,
schema: schema_pb2.Schema) -> None:
"""Make each feature self-contained.
If the feature references a global domain, copy the global domain locally.
Also do this for any child features.
Note: the name of the domain is retained, so if we want to, we could attempt
to "unnormalize" the feature, recreating global domains.
Args:
feature: feature to modify in place.
schema: schema containing any global domains.
"""
if feature.HasField("struct_domain"):
for x in feature.struct_domain.feature:
_normalize_feature(x, schema)
if feature.HasField("domain"):
for string_domain in schema.string_domain:
if string_domain.name == feature.domain:
feature.string_domain.CopyFrom(string_domain)
return
for int_domain in schema.int_domain:
if int_domain.name == feature.domain:
feature.int_domain.CopyFrom(int_domain)
return
for float_domain in schema.float_domain:
if float_domain.name == feature.domain:
feature.float_domain.CopyFrom(float_domain)
return
raise ValueError("Did not find domain {} in schema {}".format(
feature.domain, schema))
def is_categorical_feature(feature: schema_pb2.Feature):
"""Checks if the input feature is categorical."""
if feature.type == schema_pb2.BYTES:
return True
elif feature.type == schema_pb2.INT:
return ((feature.HasField('int_domain')
and feature.int_domain.is_categorical)
or feature.HasField('bool_domain'))
else:
return False
def _apply_feature(original_child: expression.Expression,
feature: schema_pb2.Feature):
"""Apply a feature to an expression. Feature should be "unclean"."""
feature_copy = [x for x in feature.struct_domain.feature
] if feature.HasField("struct_domain") else []
return _SchemaExpression(original_child, feature_copy,
_clean_feature(feature))
def is_categorical_feature(feature: schema_pb2.Feature):
"""Checks if the input feature is categorical."""
if feature.type == schema_pb2.BYTES:
return True
elif feature.type == schema_pb2.INT:
return ((feature.HasField('int_domain')
and feature.int_domain.is_categorical)
or feature.WhichOneof('domain_info')
in ['bool_domain', 'natural_language_domain'])
else:
return False
def _infer_feature_shape(feature: schema_pb2.Feature):
if feature.HasField('struct_domain'):
for struct_domain_feature in feature.struct_domain.feature:
_infer_feature_shape(struct_domain_feature)
# Currently we infer shape only for required features.
if feature.presence.min_fraction == 1:
if (feature.HasField('value_count')
and feature.value_count.min != 0
and feature.value_count.min == feature.value_count.max):
feature.shape.dim.add().size = feature.value_count.min
elif feature.HasField('value_counts'):
# Infer shape for a feature that has a nestedness level > 1 if and only
# if the min value count equals the max value count at each nestedness
# level.
dimension_sizes = list()
for value_count in feature.value_counts.value_count:
if (value_count.min == 0
or value_count.min != value_count.max):
return
dimension_sizes.append(value_count.min)
if len(dimension_sizes) == len(
feature.value_counts.value_count):
for size in dimension_sizes:
feature.shape.dim.add().size = size