def _extract_ts_features(self, ts_data, static_features, locations,
training_window_size):
"""Creates time-series feature dictionaries from data frame.
This is an internal function to allow for feature engineering using both
static and time series features.
Args:
ts_data: Time series DataFrame with columns constants.FEATURE_NAME_COLUMN,
`ModelDefinition._LOCATION_COLUMN_NAME`, constants.DATE_COLUMN, and
constants.FEATURE_VALUE_COLUMN.
static_features: Static features.
locations: Locations to be extracted.
training_window_size: Time-series data points to use for training.
Returns:
The time series dictionary mapping features to values where the values are
a mapping of locations to time series values, and the fitted scalers.
"""
all_dates = preprocessing.get_all_valid_dates(ts_data)
ts_features = preprocessing.ts_feature_df_to_nested_dict(
ts_data,
locations,
all_dates,
self.get_ts_features(),
self._LOCATION_COLUMN_NAME,
)
proc_features, feature_scalers = self.transform_ts_features(
ts_features=ts_features,
static_features=static_features,
initial_train_window_size=training_window_size)
return proc_features, feature_scalers
def extract_ts_overrides(ts_data, locations, ts_categorical_features):
"""Extract time-series overrides.
Args:
ts_data: A dataframe that contains all time-series overrides.
locations: A list of locations for which the features are needed.
ts_categorical_features: A list of names of categorical features.
Returns:
A mapping from the feature name to its value, the value of each feature
is map from location to np.ndarray.
"""
all_dates = preprocessing.get_all_valid_dates(ts_data)
all_feature_names = ts_data["feature_name"].unique().tolist()
ts_features = collections.defaultdict(
functools.partial(
collections.defaultdict,
functools.partial(np.zeros,
shape=(len(all_dates)),
dtype="float32")))
# 3 level dictionary with defaults representing feature_name, location and
# date_index.
# pylint: disable=g-long-lambda
dct = collections.defaultdict(lambda: collections.defaultdict(
lambda: collections.defaultdict(lambda: 1.0)))
# Default value for ts categorical overrides must be -1.0 to be no-op.
for feature_name in ts_categorical_features:
for location in locations:
for date_index, _ in enumerate(all_dates):
dct[feature_name][location][date_index] = -1.0
dt_index = {
pd.Timestamp(dt).to_pydatetime(): idx
for idx, dt in enumerate(all_dates)
}
for _, row in ts_data.iterrows():
if row[constants.DATE_COLUMN] not in dt_index:
continue
dct[row[constants.FEATURE_NAME_COLUMN]][row[constants.GEO_ID_COLUMN]][
dt_index[row[constants.DATE_COLUMN]]] = row[
constants.FEATURE_VALUE_COLUMN]
for feature_key in all_feature_names:
for location in locations:
for date_index, _ in enumerate(all_dates):
ts_features[feature_key][location][date_index] = dct[
feature_key][location][date_index]
key_list = list(ts_features.keys())
if key_list:
# pylint: disable=protected-access
# noinspection PyProtectedMember
preprocessing._assert_feature_lengths_for_location(
ts_features, location, reference_key=key_list[0])
# pylint: enable=protected-access
return ts_features
def extract_all_features(
self,
static_data,
ts_data,
locations,
training_window_size,
):
(static_features,
static_scalers), (ts_features,
ts_scalers) = (super().extract_all_features(
static_data, ts_data, locations,
training_window_size))
if generic_seir_specs_county.INCLUDE_AGE_AS_TIME_COVARIATE:
if static_data is None or ts_data is None:
raise ValueError(
"Both static and time series data must be provided.")
all_dates = preprocessing.get_all_valid_dates(ts_data)
preprocessing.convert_static_features_to_constant_ts(
static_features, static_scalers, ts_features, ts_scalers,
constants.COUNTY_POP_AGE_RANGES.keys(), len(all_dates))
return (static_features, static_scalers), (ts_features, ts_scalers)