def _make_common_stats_proto(
common_stats: _PartialCommonStats,
parent_common_stats: Optional[_PartialCommonStats],
q_combiner: quantiles_util.QuantilesCombiner,
num_values_histogram_buckets: int,
has_weights: bool
) -> statistics_pb2.CommonStatistics:
"""Convert the partial common stats into a CommonStatistics proto."""
result = statistics_pb2.CommonStatistics()
result.num_non_missing = common_stats.num_non_missing
if parent_common_stats is not None:
result.num_missing = (
parent_common_stats.total_num_values - common_stats.num_non_missing)
result.tot_num_values = common_stats.total_num_values
# TODO(b/79685042): Need to decide on what is the expected values for
# statistics like min_num_values, max_num_values, avg_num_values, when
# all the values for the feature are missing.
if common_stats.num_non_missing > 0:
result.min_num_values = common_stats.min_num_values
result.max_num_values = common_stats.max_num_values
result.avg_num_values = (
common_stats.total_num_values / common_stats.num_non_missing)
# Add num_values_histogram to the common stats proto.
num_values_quantiles = q_combiner.extract_output(
common_stats.num_values_summary)
histogram = quantiles_util.generate_quantiles_histogram(
num_values_quantiles, common_stats.num_non_missing,
num_values_histogram_buckets)
result.num_values_histogram.CopyFrom(histogram)
# Add weighted common stats to the proto.
if has_weights:
weighted_common_stats_proto = statistics_pb2.WeightedCommonStatistics(
num_non_missing=common_stats.weighted_num_non_missing,
tot_num_values=common_stats.weighted_total_num_values)
if parent_common_stats is not None:
weighted_common_stats_proto.num_missing = (
parent_common_stats.weighted_total_num_values -
common_stats.weighted_num_non_missing)
if common_stats.weighted_num_non_missing > 0:
weighted_common_stats_proto.avg_num_values = (
common_stats.weighted_total_num_values /
common_stats.weighted_num_non_missing)
result.weighted_common_stats.CopyFrom(
weighted_common_stats_proto)
return result
def get_dataset_feature_statistics(builder, split):
"""Calculate statistics for the specified split."""
statistics = statistics_pb2.DatasetFeatureStatistics()
# Make this to the best of our abilities.
schema = schema_pb2.Schema()
dataset = builder.as_dataset(split=split)
# Just computing the number of examples for now.
statistics.num_examples = 0
# Feature dictionaries.
feature_to_shape = {}
feature_to_dtype = {}
feature_to_num_examples = collections.defaultdict(int)
feature_to_min = {}
feature_to_max = {}
for example in dataset:
statistics.num_examples += 1
assert isinstance(example, dict)
feature_names = example.keys()
for feature_name in feature_names:
# Update the number of examples this feature appears in.
feature_to_num_examples[feature_name] += 1
feature_shape = example[feature_name].shape
feature_dtype = example[feature_name].dtype
feature_np = example[feature_name].numpy()
feature_min, feature_max = None, None
is_numeric = (
feature_dtype.is_floating or feature_dtype.is_integer or
feature_dtype.is_bool)
if is_numeric:
feature_min = np.min(feature_np)
feature_max = np.max(feature_np)
# TODO(afrozm): What if shapes don't match? Populate ValueCount? Add
# logic for that.
# Set the shape, or assert shapes match.
if feature_name not in feature_to_shape:
feature_to_shape[feature_name] = feature_shape
else:
assert feature_to_shape[feature_name] == feature_shape
# Set the shape, or assert shapes match.
if feature_name not in feature_to_dtype:
feature_to_dtype[feature_name] = feature_dtype
else:
assert feature_to_dtype[feature_name] == feature_dtype
# Set or update the min, max.
if is_numeric:
if ((feature_name not in feature_to_min) or
(feature_to_min[feature_name] > feature_min)):
feature_to_min[feature_name] = feature_min
if ((feature_name not in feature_to_max) or
(feature_to_max[feature_name] < feature_max)):
feature_to_max[feature_name] = feature_max
# Start here, we've processed all examples.
# Assert that the keys match up.
assert feature_to_shape.keys() == feature_to_dtype.keys()
assert feature_to_shape.keys() == feature_to_num_examples.keys()
for feature_name in feature_to_shape:
# Try to fill in the schema.
feature = schema.feature.add()
feature.name = feature_name
# TODO(afrozm): What do we do for non fixed size shapes?
# What to do for scalars?
for dim in feature_to_shape[feature_name].as_list():
feature.shape.dim.add().size = dim
feature_type = feature_to_dtype[feature_name]
feature.type = _FEATURE_TYPE_MAP.get(feature_type, schema_pb2.BYTES)
common_statistics = statistics_pb2.CommonStatistics()
common_statistics.num_non_missing = feature_to_num_examples[feature_name]
common_statistics.num_missing = (
statistics.num_examples - common_statistics.num_non_missing)
feature_name_statistics = statistics.features.add()
feature_name_statistics.name = feature_name
# TODO(afrozm): This can be skipped, since type information was added to
# the Schema.
feature_name_statistics.type = _SCHEMA_TYPE_MAP.get(
feature.type, statistics_pb2.FeatureNameStatistics.BYTES)
if feature.type == schema_pb2.INT or feature.type == schema_pb2.FLOAT:
numeric_statistics = statistics_pb2.NumericStatistics()
numeric_statistics.min = feature_to_min[feature_name]
numeric_statistics.max = feature_to_max[feature_name]
numeric_statistics.common_stats.CopyFrom(common_statistics)
feature_name_statistics.num_stats.CopyFrom(numeric_statistics)
else:
# Let's shove it into BytesStatistics for now.
bytes_statistics = statistics_pb2.BytesStatistics()
bytes_statistics.common_stats.CopyFrom(common_statistics)
feature_name_statistics.bytes_stats.CopyFrom(bytes_statistics)
return statistics, schema
def get_dataset_feature_statistics(builder, split):
"""Calculate statistics for the specified split."""
statistics = statistics_pb2.DatasetFeatureStatistics()
# Make this to the best of our abilities.
schema = schema_pb2.Schema()
dataset = builder.as_dataset(split=split)
# Just computing the number of examples for now.
statistics.num_examples = 0
# Feature dictionaries.
feature_to_num_examples = collections.defaultdict(int)
feature_to_min = {}
feature_to_max = {}
np_dataset = dataset_utils.dataset_as_numpy(dataset)
for example in tqdm.tqdm(np_dataset, unit=" examples"):
statistics.num_examples += 1
assert isinstance(example, dict)
feature_names = sorted(example.keys())
for feature_name in feature_names:
# Update the number of examples this feature appears in.
feature_to_num_examples[feature_name] += 1
feature_np = example[feature_name]
# For compatibility in graph and eager mode, we can get PODs here and
# everything may not be neatly wrapped up in numpy's ndarray.
feature_dtype = type(feature_np)
if isinstance(feature_np, np.ndarray):
feature_dtype = feature_np.dtype.type
feature_min, feature_max = None, None
is_numeric = (np.issubdtype(feature_dtype, np.number)
or feature_dtype == np.bool_)
if is_numeric:
feature_min = np.min(feature_np)
feature_max = np.max(feature_np)
# TODO(afrozm): What if shapes don't match? Populate ValueCount? Add
# logic for that.
# Set or update the min, max.
if is_numeric:
if ((feature_name not in feature_to_min)
or (feature_to_min[feature_name] > feature_min)):
feature_to_min[feature_name] = feature_min
if ((feature_name not in feature_to_max)
or (feature_to_max[feature_name] < feature_max)):
feature_to_max[feature_name] = feature_max
# Start here, we've processed all examples.
output_shapes_dict = dataset.output_shapes
output_types_dict = dataset.output_types
for feature_name in sorted(feature_to_num_examples.keys()):
# Try to fill in the schema.
feature = schema.feature.add()
feature.name = feature_name
# TODO(afrozm): Make this work with nested structures, currently the Schema
# proto has no support for it.
maybe_feature_shape = output_shapes_dict[feature_name]
if not isinstance(maybe_feature_shape, tf.TensorShape):
logging.error(
"Statistics generation doesn't work for nested structures yet")
continue
for dim in maybe_feature_shape.as_list():
# We denote `None`s as -1 in the shape proto.
feature.shape.dim.add().size = dim if dim else -1
feature_type = output_types_dict[feature_name]
feature.type = _FEATURE_TYPE_MAP.get(feature_type, schema_pb2.BYTES)
common_statistics = statistics_pb2.CommonStatistics()
common_statistics.num_non_missing = feature_to_num_examples[
feature_name]
common_statistics.num_missing = (statistics.num_examples -
common_statistics.num_non_missing)
feature_name_statistics = statistics.features.add()
feature_name_statistics.name = feature_name
# TODO(afrozm): This can be skipped, since type information was added to
# the Schema.
feature_name_statistics.type = _SCHEMA_TYPE_MAP.get(
feature.type, statistics_pb2.FeatureNameStatistics.BYTES)
if feature.type == schema_pb2.INT or feature.type == schema_pb2.FLOAT:
numeric_statistics = statistics_pb2.NumericStatistics()
numeric_statistics.min = feature_to_min[feature_name]
numeric_statistics.max = feature_to_max[feature_name]
numeric_statistics.common_stats.CopyFrom(common_statistics)
feature_name_statistics.num_stats.CopyFrom(numeric_statistics)
else:
# Let's shove it into BytesStatistics for now.
bytes_statistics = statistics_pb2.BytesStatistics()
bytes_statistics.common_stats.CopyFrom(common_statistics)
feature_name_statistics.bytes_stats.CopyFrom(bytes_statistics)
return statistics, schema
def _make_feature_stats_proto(common_stats, feature_name, q_combiner,
num_values_histogram_buckets, is_categorical,
has_weights):
"""Convert the partial common stats into a FeatureNameStatistics proto.
Args:
common_stats: The partial common stats associated with a feature.
feature_name: The name of the feature.
q_combiner: The quantiles combiner used to construct the quantiles
histogram for the number of values in the feature.
num_values_histogram_buckets: Number of buckets in the quantiles
histogram for the number of values per feature.
is_categorical: A boolean indicating whether the feature is categorical.
has_weights: A boolean indicating whether a weight feature is specified.
Returns:
A statistics_pb2.FeatureNameStatistics proto.
"""
common_stats_proto = statistics_pb2.CommonStatistics()
common_stats_proto.num_non_missing = common_stats.num_non_missing
common_stats_proto.num_missing = common_stats.num_missing
common_stats_proto.tot_num_values = common_stats.total_num_values
if common_stats.num_non_missing > 0:
common_stats_proto.min_num_values = common_stats.min_num_values
common_stats_proto.max_num_values = common_stats.max_num_values
common_stats_proto.avg_num_values = (common_stats.total_num_values /
common_stats.num_non_missing)
# Add num_values_histogram to the common stats proto.
num_values_quantiles = q_combiner.extract_output(
common_stats.num_values_summary)
histogram = quantiles_util.generate_quantiles_histogram(
num_values_quantiles, common_stats.min_num_values,
common_stats.max_num_values, common_stats.num_non_missing,
num_values_histogram_buckets)
common_stats_proto.num_values_histogram.CopyFrom(histogram)
# Add weighted common stats to the proto.
if has_weights:
weighted_common_stats_proto = statistics_pb2.WeightedCommonStatistics(
num_non_missing=common_stats.weighted_num_non_missing,
num_missing=common_stats.weighted_num_missing,
tot_num_values=common_stats.weighted_total_num_values)
if common_stats.weighted_num_non_missing > 0:
weighted_common_stats_proto.avg_num_values = (
common_stats.weighted_total_num_values /
common_stats.weighted_num_non_missing)
common_stats_proto.weighted_common_stats.CopyFrom(
weighted_common_stats_proto)
# Create a new FeatureNameStatistics proto.
result = statistics_pb2.FeatureNameStatistics()
result.name = feature_name
# Set the feature type.
# If we have a categorical feature, we preserve the type to be the original
# INT type. Currently we don't set the type if we cannot infer it, which
# happens when all the values are missing. We need to add an UNKNOWN type
# to the stats proto to handle this case.
if is_categorical:
result.type = statistics_pb2.FeatureNameStatistics.INT
elif common_stats.type is None:
# If a feature is completely missing, we assume the type to be STRING.
result.type = statistics_pb2.FeatureNameStatistics.STRING
else:
result.type = common_stats.type
# Copy the common stats into appropriate numeric/string stats.
# If the type is not set, we currently wrap the common stats
# within numeric stats.
if (result.type == statistics_pb2.FeatureNameStatistics.STRING
or is_categorical):
# Add the common stats into string stats.
string_stats_proto = statistics_pb2.StringStatistics()
string_stats_proto.common_stats.CopyFrom(common_stats_proto)
result.string_stats.CopyFrom(string_stats_proto)
else:
# Add the common stats into numeric stats.
numeric_stats_proto = statistics_pb2.NumericStatistics()
numeric_stats_proto.common_stats.CopyFrom(common_stats_proto)
result.num_stats.CopyFrom(numeric_stats_proto)
return result
def _make_common_stats_proto(
common_stats: _PartialCommonStats,
parent_common_stats: Optional[_PartialCommonStats],
make_quantiles_sketch_fn: Callable[[], sketches.QuantilesSketch],
num_values_histogram_buckets: int,
has_weights: bool
) -> statistics_pb2.CommonStatistics:
"""Convert the partial common stats into a CommonStatistics proto."""
result = statistics_pb2.CommonStatistics()
parent_presence_and_valency = None
if parent_common_stats is not None:
parent_presence_and_valency = (
_PresenceAndValencyStats(make_quantiles_sketch_fn)
if parent_common_stats.presence_and_valency_stats is None else
parent_common_stats.presence_and_valency_stats[-1])
presence_and_valency_stats = common_stats.presence_and_valency_stats
# the CommonStatistics already contains the presence and valency
# for a 1-nested feature.
if (presence_and_valency_stats is not None and
len(presence_and_valency_stats) > 1):
result.presence_and_valency_stats.extend(
_make_presence_and_valency_stats_protos(
parent_presence_and_valency,
common_stats.presence_and_valency_stats))
if has_weights:
result.weighted_presence_and_valency_stats.extend(
_make_weighted_presence_and_valency_stats_protos(
parent_presence_and_valency,
common_stats.presence_and_valency_stats))
top_level_presence_and_valency = (
_PresenceAndValencyStats(make_quantiles_sketch_fn)
if common_stats.presence_and_valency_stats is None else
common_stats.presence_and_valency_stats[0])
result.num_non_missing = top_level_presence_and_valency.num_non_missing
if parent_presence_and_valency is not None:
result.num_missing = (
parent_presence_and_valency.total_num_values -
top_level_presence_and_valency.num_non_missing)
result.tot_num_values = top_level_presence_and_valency.total_num_values
# TODO(b/79685042): Need to decide on what is the expected values for
# statistics like min_num_values, max_num_values, avg_num_values, when
# all the values for the feature are missing.
if top_level_presence_and_valency.num_non_missing > 0:
result.min_num_values = top_level_presence_and_valency.min_num_values
result.max_num_values = top_level_presence_and_valency.max_num_values
result.avg_num_values = (
top_level_presence_and_valency.total_num_values /
top_level_presence_and_valency.num_non_missing)
if top_level_presence_and_valency.num_values_summary is not None:
# Add num_values_histogram to the common stats proto.
num_values_quantiles = (
top_level_presence_and_valency.num_values_summary.GetQuantiles(
num_values_histogram_buckets).flatten().to_pylist())
histogram = quantiles_util.generate_quantiles_histogram(
num_values_quantiles, top_level_presence_and_valency.num_non_missing,
num_values_histogram_buckets)
result.num_values_histogram.CopyFrom(histogram)
# Add weighted common stats to the proto.
if has_weights:
weighted_common_stats_proto = statistics_pb2.WeightedCommonStatistics(
num_non_missing=top_level_presence_and_valency.weighted_num_non_missing,
tot_num_values=top_level_presence_and_valency.weighted_total_num_values)
if parent_presence_and_valency is not None:
weighted_common_stats_proto.num_missing = (
parent_presence_and_valency.weighted_total_num_values -
top_level_presence_and_valency.weighted_num_non_missing)
if top_level_presence_and_valency.weighted_num_non_missing > 0:
weighted_common_stats_proto.avg_num_values = (
top_level_presence_and_valency.weighted_total_num_values /
top_level_presence_and_valency.weighted_num_non_missing)
result.weighted_common_stats.CopyFrom(
weighted_common_stats_proto)
return result
