Add TFT benchmark

examples/benchmarks/TFT/data_formatters/__init__.py

@@ -0,0 +1,15 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+

examples/benchmarks/TFT/data_formatters/base.py

@@ -0,0 +1,235 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Default data formatting functions for experiments.
+
+For new datasets, inherit form GenericDataFormatter and implement
+all abstract functions.
+
+These dataset-specific methods:
+1) Define the column and input types for tabular dataframes used by model
+2) Perform the necessary input feature engineering & normalisation steps
+3) Reverts the normalisation for predictions
+4) Are responsible for train, validation and test splits
+
+
+"""
+
+import abc
+import enum
+
+
+# Type defintions
+class DataTypes(enum.IntEnum):
+  """Defines numerical types of each column."""
+  REAL_VALUED = 0
+  CATEGORICAL = 1
+  DATE = 2
+
+
+class InputTypes(enum.IntEnum):
+  """Defines input types of each column."""
+  TARGET = 0
+  OBSERVED_INPUT = 1
+  KNOWN_INPUT = 2
+  STATIC_INPUT = 3
+  ID = 4  # Single column used as an entity identifier
+  TIME = 5  # Single column exclusively used as a time index
+
+
+class GenericDataFormatter(abc.ABC):
+  """Abstract base class for all data formatters.
+
+  User can implement the abstract methods below to perform dataset-specific
+  manipulations.
+
+  """
+
+  @abc.abstractmethod
+  def set_scalers(self, df):
+    """Calibrates scalers using the data supplied."""
+    raise NotImplementedError()
+
+  @abc.abstractmethod
+  def transform_inputs(self, df):
+    """Performs feature transformation."""
+    raise NotImplementedError()
+
+  @abc.abstractmethod
+  def format_predictions(self, df):
+    """Reverts any normalisation to give predictions in original scale."""
+    raise NotImplementedError()
+
+  @abc.abstractmethod
+  def split_data(self, df):
+    """Performs the default train, validation and test splits."""
+    raise NotImplementedError()
+
+  @property
+  @abc.abstractmethod
+  def _column_definition(self):
+    """Defines order, input type and data type of each column."""
+    raise NotImplementedError()
+
+  @abc.abstractmethod
+  def get_fixed_params(self):
+    """Defines the fixed parameters used by the model for training.
+
+    Requires the following keys:
+      'total_time_steps': Defines the total number of time steps used by TFT
+      'num_encoder_steps': Determines length of LSTM encoder (i.e. history)
+      'num_epochs': Maximum number of epochs for training
+      'early_stopping_patience': Early stopping param for keras
+      'multiprocessing_workers': # of cpus for data processing
+
+
+    Returns:
+      A dictionary of fixed parameters, e.g.:
+
+      fixed_params = {
+          'total_time_steps': 252 + 5,
+          'num_encoder_steps': 252,
+          'num_epochs': 100,
+          'early_stopping_patience': 5,
+          'multiprocessing_workers': 5,
+      }
+    """
+    raise NotImplementedError
+
+  # Shared functions across data-formatters
+  @property
+  def num_classes_per_cat_input(self):
+    """Returns number of categories per relevant input.
+
+    This is seqeuently required for keras embedding layers.
+    """
+    return self._num_classes_per_cat_input
+
+  def get_num_samples_for_calibration(self):
+    """Gets the default number of training and validation samples.
+
+    Use to sub-sample the data for network calibration and a value of -1 uses
+    all available samples.
+
+    Returns:
+      Tuple of (training samples, validation samples)
+    """
+    return -1, -1
+
+  def get_column_definition(self):
+    """"Returns formatted column definition in order expected by the TFT."""
+
+    column_definition = self._column_definition
+
+    # Sanity checks first.
+    # Ensure only one ID and time column exist
+    def _check_single_column(input_type):
+
+      length = len([tup for tup in column_definition if tup[2] == input_type])
+
+      if length != 1:
+        raise ValueError('Illegal number of inputs ({}) of type {}'.format(
+            length, input_type))
+
+    _check_single_column(InputTypes.ID)
+    _check_single_column(InputTypes.TIME)
+
+    identifier = [tup for tup in column_definition if tup[2] == InputTypes.ID]
+    time = [tup for tup in column_definition if tup[2] == InputTypes.TIME]
+    real_inputs = [
+        tup for tup in column_definition if tup[1] == DataTypes.REAL_VALUED and
+        tup[2] not in {InputTypes.ID, InputTypes.TIME}
+    ]
+    categorical_inputs = [
+        tup for tup in column_definition if tup[1] == DataTypes.CATEGORICAL and
+        tup[2] not in {InputTypes.ID, InputTypes.TIME}
+    ]
+
+    return identifier + time + real_inputs + categorical_inputs
+
+  def _get_input_columns(self):
+    """Returns names of all input columns."""
+    return [
+        tup[0]
+        for tup in self.get_column_definition()
+        if tup[2] not in {InputTypes.ID, InputTypes.TIME}
+    ]
+
+  def _get_tft_input_indices(self):
+    """Returns the relevant indexes and input sizes required by TFT."""
+
+    # Functions
+    def _extract_tuples_from_data_type(data_type, defn):
+      return [
+          tup for tup in defn if tup[1] == data_type and
+          tup[2] not in {InputTypes.ID, InputTypes.TIME}
+      ]
+
+    def _get_locations(input_types, defn):
+      return [i for i, tup in enumerate(defn) if tup[2] in input_types]
+
+    # Start extraction
+    column_definition = [
+        tup for tup in self.get_column_definition()
+        if tup[2] not in {InputTypes.ID, InputTypes.TIME}
+    ]
+
+    categorical_inputs = _extract_tuples_from_data_type(DataTypes.CATEGORICAL,
+                                                        column_definition)
+    real_inputs = _extract_tuples_from_data_type(DataTypes.REAL_VALUED,
+                                                 column_definition)
+
+    locations = {
+        'input_size':
+            len(self._get_input_columns()),
+        'output_size':
+            len(_get_locations({InputTypes.TARGET}, column_definition)),
+        'category_counts':
+            self.num_classes_per_cat_input,
+        'input_obs_loc':
+            _get_locations({InputTypes.TARGET}, column_definition),
+        'static_input_loc':
+            _get_locations({InputTypes.STATIC_INPUT}, column_definition),
+        'known_regular_inputs':
+            _get_locations({InputTypes.STATIC_INPUT, InputTypes.KNOWN_INPUT},
+                           real_inputs),
+        'known_categorical_inputs':
+            _get_locations({InputTypes.STATIC_INPUT, InputTypes.KNOWN_INPUT},
+                           categorical_inputs),
+    }
+
+    return locations
+
+  def get_experiment_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    required_keys = [
+        'total_time_steps', 'num_encoder_steps', 'num_epochs',
+        'early_stopping_patience', 'multiprocessing_workers'
+    ]
+
+    fixed_params = self.get_fixed_params()
+
+    for k in required_keys:
+      if k not in fixed_params:
+        raise ValueError('Field {}'.format(k) +
+                         ' missing from fixed parameter definitions!')
+
+    fixed_params['column_definition'] = self.get_column_definition()
+
+    fixed_params.update(self._get_tft_input_indices())
+
+    return fixed_params

examples/benchmarks/TFT/data_formatters/electricity.py

@@ -0,0 +1,261 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Custom formatting functions for Electricity dataset.
+
+Defines dataset specific column definitions and data transformations. Uses
+entity specific z-score normalization.
+"""
+
+import data_formatters.base
+import libs.utils as utils
+import pandas as pd
+import sklearn.preprocessing
+
+GenericDataFormatter = data_formatters.base.GenericDataFormatter
+DataTypes = data_formatters.base.DataTypes
+InputTypes = data_formatters.base.InputTypes
+
+
+class ElectricityFormatter(GenericDataFormatter):
+  """Defines and formats data for the electricity dataset.
+
+  Note that per-entity z-score normalization is used here, and is implemented
+  across functions.
+
+  Attributes:
+    column_definition: Defines input and data type of column used in the
+      experiment.
+    identifiers: Entity identifiers used in experiments.
+  """
+
+  _column_definition = [
+      ('id', DataTypes.REAL_VALUED, InputTypes.ID),
+      ('hours_from_start', DataTypes.REAL_VALUED, InputTypes.TIME),
+      ('power_usage', DataTypes.REAL_VALUED, InputTypes.TARGET),
+      ('hour', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('day_of_week', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('hours_from_start', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('categorical_id', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+  ]
+
+  def __init__(self):
+    """Initialises formatter."""
+
+    self.identifiers = None
+    self._real_scalers = None
+    self._cat_scalers = None
+    self._target_scaler = None
+    self._num_classes_per_cat_input = None
+    self._time_steps = self.get_fixed_params()['total_time_steps']
+
+  def split_data(self, df, valid_boundary=1315, test_boundary=1339):
+    """Splits data frame into training-validation-test data frames.
+
+    This also calibrates scaling object, and transforms data for each split.
+
+    Args:
+      df: Source data frame to split.
+      valid_boundary: Starting year for validation data
+      test_boundary: Starting year for test data
+
+    Returns:
+      Tuple of transformed (train, valid, test) data.
+    """
+
+    print('Formatting train-valid-test splits.')
+
+    index = df['days_from_start']
+    train = df.loc[index < valid_boundary]
+    valid = df.loc[(index >= valid_boundary - 7) & (index < test_boundary)]
+    test = df.loc[index >= test_boundary - 7]
+
+    self.set_scalers(train)
+
+    return (self.transform_inputs(data) for data in [train, valid, test])
+
+  def set_scalers(self, df):
+    """Calibrates scalers using the data supplied.
+
+    Args:
+      df: Data to use to calibrate scalers.
+    """
+    print('Setting scalers with training data...')
+
+    column_definitions = self.get_column_definition()
+    id_column = utils.get_single_col_by_input_type(InputTypes.ID,
+                                                   column_definitions)
+    target_column = utils.get_single_col_by_input_type(InputTypes.TARGET,
+                                                       column_definitions)
+
+    # Format real scalers
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    # Initialise scaler caches
+    self._real_scalers = {}
+    self._target_scaler = {}
+    identifiers = []
+    for identifier, sliced in df.groupby(id_column):
+
+      if len(sliced) >= self._time_steps:
+
+        data = sliced[real_inputs].values
+        targets = sliced[[target_column]].values
+        self._real_scalers[identifier] \
+      = sklearn.preprocessing.StandardScaler().fit(data)
+
+        self._target_scaler[identifier] \
+      = sklearn.preprocessing.StandardScaler().fit(targets)
+      identifiers.append(identifier)
+
+    # Format categorical scalers
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    categorical_scalers = {}
+    num_classes = []
+    for col in categorical_inputs:
+      # Set all to str so that we don't have mixed integer/string columns
+      srs = df[col].apply(str)
+      categorical_scalers[col] = sklearn.preprocessing.LabelEncoder().fit(
+          srs.values)
+      num_classes.append(srs.nunique())
+
+    # Set categorical scaler outputs
+    self._cat_scalers = categorical_scalers
+    self._num_classes_per_cat_input = num_classes
+
+    # Extract identifiers in case required
+    self.identifiers = identifiers
+
+  def transform_inputs(self, df):
+    """Performs feature transformations.
+
+    This includes both feature engineering, preprocessing and normalisation.
+
+    Args:
+      df: Data frame to transform.
+
+    Returns:
+      Transformed data frame.
+
+    """
+
+    if self._real_scalers is None and self._cat_scalers is None:
+      raise ValueError('Scalers have not been set!')
+
+    # Extract relevant columns
+    column_definitions = self.get_column_definition()
+    id_col = utils.get_single_col_by_input_type(InputTypes.ID,
+                                                column_definitions)
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    # Transform real inputs per entity
+    df_list = []
+    for identifier, sliced in df.groupby(id_col):
+
+      # Filter out any trajectories that are too short
+      if len(sliced) >= self._time_steps:
+        sliced_copy = sliced.copy()
+        sliced_copy[real_inputs] = self._real_scalers[identifier].transform(
+            sliced_copy[real_inputs].values)
+        df_list.append(sliced_copy)
+
+    output = pd.concat(df_list, axis=0)
+
+    # Format categorical inputs
+    for col in categorical_inputs:
+      string_df = df[col].apply(str)
+      output[col] = self._cat_scalers[col].transform(string_df)
+
+    return output
+
+  def format_predictions(self, predictions):
+    """Reverts any normalisation to give predictions in original scale.
+
+    Args:
+      predictions: Dataframe of model predictions.
+
+    Returns:
+      Data frame of unnormalised predictions.
+    """
+
+    if self._target_scaler is None:
+      raise ValueError('Scalers have not been set!')
+
+    column_names = predictions.columns
+
+    df_list = []
+    for identifier, sliced in predictions.groupby('identifier'):
+      sliced_copy = sliced.copy()
+      target_scaler = self._target_scaler[identifier]
+
+      for col in column_names:
+        if col not in {'forecast_time', 'identifier'}:
+          sliced_copy[col] = target_scaler.inverse_transform(sliced_copy[col])
+      df_list.append(sliced_copy)
+
+    output = pd.concat(df_list, axis=0)
+
+    return output
+
+  # Default params
+  def get_fixed_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    fixed_params = {
+        'total_time_steps': 8 * 24,
+        'num_encoder_steps': 7 * 24,
+        'num_epochs': 100,
+        'early_stopping_patience': 5,
+        'multiprocessing_workers': 5
+    }
+
+    return fixed_params
+
+  def get_default_model_params(self):
+    """Returns default optimised model parameters."""
+
+    model_params = {
+        'dropout_rate': 0.1,
+        'hidden_layer_size': 160,
+        'learning_rate': 0.001,
+        'minibatch_size': 64,
+        'max_gradient_norm': 0.01,
+        'num_heads': 4,
+        'stack_size': 1
+    }
+
+    return model_params
+
+  def get_num_samples_for_calibration(self):
+    """Gets the default number of training and validation samples.
+
+    Use to sub-sample the data for network calibration and a value of -1 uses
+    all available samples.
+
+    Returns:
+      Tuple of (training samples, validation samples)
+    """
+    return 450000, 50000

examples/benchmarks/TFT/data_formatters/favorita.py

@@ -0,0 +1,327 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Custom formatting functions for Favorita dataset.
+
+Defines dataset specific column definitions and data transformations.
+"""
+
+import data_formatters.base
+import libs.utils as utils
+import pandas as pd
+import sklearn.preprocessing
+
+DataTypes = data_formatters.base.DataTypes
+InputTypes = data_formatters.base.InputTypes
+
+
+class FavoritaFormatter(data_formatters.base.GenericDataFormatter):
+  """Defines and formats data for the Favorita dataset.
+
+  Attributes:
+    column_definition: Defines input and data type of column used in the
+      experiment.
+    identifiers: Entity identifiers used in experiments.
+  """
+
+  _column_definition = [
+      ('traj_id', DataTypes.REAL_VALUED, InputTypes.ID),
+      ('date', DataTypes.DATE, InputTypes.TIME),
+      ('log_sales', DataTypes.REAL_VALUED, InputTypes.TARGET),
+      ('onpromotion', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('transactions', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('oil', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('day_of_week', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('day_of_month', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('month', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('national_hol', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('regional_hol', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('local_hol', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('open', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('item_nbr', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('store_nbr', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('city', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('state', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('type', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('cluster', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('family', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('class', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+      ('perishable', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT)
+  ]
+
+  def __init__(self):
+    """Initialises formatter."""
+
+    self.identifiers = None
+    self._real_scalers = None
+    self._cat_scalers = None
+    self._target_scaler = None
+    self._num_classes_per_cat_input = None
+
+  def split_data(self, df, valid_boundary=None, test_boundary=None):
+    """Splits data frame into training-validation-test data frames.
+
+    This also calibrates scaling object, and transforms data for each split.
+
+    Args:
+      df: Source data frame to split.
+      valid_boundary: Starting year for validation data
+      test_boundary: Starting year for test data
+
+    Returns:
+      Tuple of transformed (train, valid, test) data.
+    """
+
+    print('Formatting train-valid-test splits.')
+
+    if valid_boundary is None:
+      valid_boundary = pd.datetime(2015, 12, 1)
+
+    fixed_params = self.get_fixed_params()
+    time_steps = fixed_params['total_time_steps']
+    lookback = fixed_params['num_encoder_steps']
+    forecast_horizon = time_steps - lookback
+
+    df['date'] = pd.to_datetime(df['date'])
+    df_lists = {'train': [], 'valid': [], 'test': []}
+    for _, sliced in df.groupby('traj_id'):
+      index = sliced['date']
+      train = sliced.loc[index < valid_boundary]
+      train_len = len(train)
+      valid_len = train_len + forecast_horizon
+      valid = sliced.iloc[train_len - lookback:valid_len, :]
+      test = sliced.iloc[valid_len - lookback:valid_len + forecast_horizon, :]
+
+      sliced_map = {'train': train, 'valid': valid, 'test': test}
+
+      for k in sliced_map:
+        item = sliced_map[k]
+
+        if len(item) >= time_steps:
+          df_lists[k].append(item)
+
+    dfs = {k: pd.concat(df_lists[k], axis=0) for k in df_lists}
+
+    train = dfs['train']
+    self.set_scalers(train, set_real=True)
+
+    # Use all data for label encoding  to handle labels not present in training.
+    self.set_scalers(df, set_real=False)
+
+    # Filter out identifiers not present in training (i.e. cold-started items).
+    def filter_ids(frame):
+      identifiers = set(self.identifiers)
+      index = frame['traj_id']
+      return frame.loc[index.apply(lambda x: x in identifiers)]
+
+    valid = filter_ids(dfs['valid'])
+    test = filter_ids(dfs['test'])
+
+    return (self.transform_inputs(data) for data in [train, valid, test])
+
+  def set_scalers(self, df, set_real=True):
+    """Calibrates scalers using the data supplied.
+
+    Label encoding is applied to the entire dataset (i.e. including test),
+    so that unseen labels can be handled at run-time.
+
+    Args:
+      df: Data to use to calibrate scalers.
+      set_real: Whether to fit set real-valued or categorical scalers
+    """
+    print('Setting scalers with training data...')
+
+    column_definitions = self.get_column_definition()
+    id_column = utils.get_single_col_by_input_type(InputTypes.ID,
+                                                   column_definitions)
+    target_column = utils.get_single_col_by_input_type(InputTypes.TARGET,
+                                                       column_definitions)
+
+    if set_real:
+
+      # Extract identifiers in case required
+      self.identifiers = list(df[id_column].unique())
+
+      # Format real scalers
+      self._real_scalers = {}
+      for col in ['oil', 'transactions', 'log_sales']:
+        self._real_scalers[col] = (df[col].mean(), df[col].std())
+
+      self._target_scaler = (df[target_column].mean(), df[target_column].std())
+
+    else:
+      # Format categorical scalers
+      categorical_inputs = utils.extract_cols_from_data_type(
+          DataTypes.CATEGORICAL, column_definitions,
+          {InputTypes.ID, InputTypes.TIME})
+
+      categorical_scalers = {}
+      num_classes = []
+      if self.identifiers is None:
+        raise ValueError('Scale real-valued inputs first!')
+      id_set = set(self.identifiers)
+      valid_idx = df['traj_id'].apply(lambda x: x in id_set)
+      for col in categorical_inputs:
+        # Set all to str so that we don't have mixed integer/string columns
+        srs = df[col].apply(str).loc[valid_idx]
+        categorical_scalers[col] = sklearn.preprocessing.LabelEncoder().fit(
+            srs.values)
+
+        num_classes.append(srs.nunique())
+
+      # Set categorical scaler outputs
+      self._cat_scalers = categorical_scalers
+      self._num_classes_per_cat_input = num_classes
+
+  def transform_inputs(self, df):
+    """Performs feature transformations.
+
+    This includes both feature engineering, preprocessing and normalisation.
+
+    Args:
+      df: Data frame to transform.
+
+    Returns:
+      Transformed data frame.
+
+    """
+    output = df.copy()
+
+    if self._real_scalers is None and self._cat_scalers is None:
+      raise ValueError('Scalers have not been set!')
+
+    column_definitions = self.get_column_definition()
+
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    # Format real inputs
+    for col in ['log_sales', 'oil', 'transactions']:
+      mean, std = self._real_scalers[col]
+      output[col] = (df[col] - mean) / std
+
+      if col == 'log_sales':
+        output[col] = output[col].fillna(0.)  # mean imputation
+
+    # Format categorical inputs
+    for col in categorical_inputs:
+      string_df = df[col].apply(str)
+      output[col] = self._cat_scalers[col].transform(string_df)
+
+    return output
+
+  def format_predictions(self, predictions):
+    """Reverts any normalisation to give predictions in original scale.
+
+    Args:
+      predictions: Dataframe of model predictions.
+
+    Returns:
+      Data frame of unnormalised predictions.
+    """
+    output = predictions.copy()
+
+    column_names = predictions.columns
+    mean, std = self._target_scaler
+    for col in column_names:
+      if col not in {'forecast_time', 'identifier'}:
+        output[col] = (predictions[col] * std) + mean
+
+    return output
+
+  # Default params
+  def get_fixed_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    fixed_params = {
+        'total_time_steps': 120,
+        'num_encoder_steps': 90,
+        'num_epochs': 100,
+        'early_stopping_patience': 5,
+        'multiprocessing_workers': 5
+    }
+
+    return fixed_params
+
+  def get_default_model_params(self):
+    """Returns default optimised model parameters."""
+
+    model_params = {
+        'dropout_rate': 0.1,
+        'hidden_layer_size': 240,
+        'learning_rate': 0.001,
+        'minibatch_size': 128,
+        'max_gradient_norm': 100.,
+        'num_heads': 4,
+        'stack_size': 1
+    }
+
+    return model_params
+
+  def get_num_samples_for_calibration(self):
+    """Gets the default number of training and validation samples.
+
+    Use to sub-sample the data for network calibration and a value of -1 uses
+    all available samples.
+
+    Returns:
+      Tuple of (training samples, validation samples)
+    """
+    return 450000, 50000
+
+  def get_column_definition(self):
+    """"Formats column definition in order expected by the TFT.
+
+    Modified for Favorita to match column order of original experiment.
+
+    Returns:
+      Favorita-specific column definition
+    """
+
+    column_definition = self._column_definition
+
+    # Sanity checks first.
+    # Ensure only one ID and time column exist
+    def _check_single_column(input_type):
+
+      length = len([tup for tup in column_definition if tup[2] == input_type])
+
+      if length != 1:
+        raise ValueError('Illegal number of inputs ({}) of type {}'.format(
+            length, input_type))
+
+    _check_single_column(InputTypes.ID)
+    _check_single_column(InputTypes.TIME)
+
+    identifier = [tup for tup in column_definition if tup[2] == InputTypes.ID]
+    time = [tup for tup in column_definition if tup[2] == InputTypes.TIME]
+    real_inputs = [
+        tup for tup in column_definition if tup[1] == DataTypes.REAL_VALUED and
+        tup[2] not in {InputTypes.ID, InputTypes.TIME}
+    ]
+
+    col_definition_map = {tup[0]: tup for tup in column_definition}
+    col_order = [
+        'item_nbr', 'store_nbr', 'city', 'state', 'type', 'cluster', 'family',
+        'class', 'perishable', 'onpromotion', 'day_of_week', 'national_hol',
+        'regional_hol', 'local_hol'
+    ]
+    categorical_inputs = [
+        col_definition_map[k] for k in col_order if k in col_definition_map
+    ]
+
+    return identifier + time + real_inputs + categorical_inputs

examples/benchmarks/TFT/data_formatters/qlib_Alpha158.py

@@ -0,0 +1,220 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Custom formatting functions for Alpha158 dataset.
+
+Defines dataset specific column definitions and data transformations.
+"""
+
+import data_formatters.base
+import libs.utils as utils
+import sklearn.preprocessing
+
+GenericDataFormatter = data_formatters.base.GenericDataFormatter
+DataTypes = data_formatters.base.DataTypes
+InputTypes = data_formatters.base.InputTypes
+
+class Alpha158Formatter(GenericDataFormatter):
+  """Defines and formats data for the Alpha158 dataset.
+
+  Attributes:
+    column_definition: Defines input and data type of column used in the
+      experiment.
+    identifiers: Entity identifiers used in experiments.
+  """
+
+  _column_definition = [
+      ('instrument', DataTypes.CATEGORICAL, InputTypes.ID),
+      ('LABEL0', DataTypes.REAL_VALUED, InputTypes.TARGET),
+      ('date', DataTypes.DATE, InputTypes.TIME),
+      ('month', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('day_of_week', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      # Selected 10 features
+      ('RESI5', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('WVMA5', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('RSQR5', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('KLEN', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('RSQR10', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('CORR5', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('CORD5', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('CORR10', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('ROC60', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('RESI10', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('const', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+    ]
+
+  def __init__(self):
+    """Initialises formatter."""
+
+    self.identifiers = None
+    self._real_scalers = None
+    self._cat_scalers = None
+    self._target_scaler = None
+    self._num_classes_per_cat_input = None
+
+  def split_data(self, df, valid_boundary=2016, test_boundary=2018):
+    """Splits data frame into training-validation-test data frames.
+
+    This also calibrates scaling object, and transforms data for each split.
+
+    Args:
+      df: Source data frame to split.
+      valid_boundary: Starting year for validation data
+      test_boundary: Starting year for test data
+
+    Returns:
+      Tuple of transformed (train, valid, test) data.
+    """
+
+    print('Formatting train-valid-test splits.')
+
+    index = df['year']
+    train = df.loc[index < valid_boundary]
+    valid = df.loc[(index >= valid_boundary) & (index < test_boundary)]
+    test = df.loc[index >= test_boundary]
+
+    self.set_scalers(train)
+
+    return (self.transform_inputs(data) for data in [train, valid, test])
+
+  def set_scalers(self, df):
+    """Calibrates scalers using the data supplied.
+
+    Args:
+      df: Data to use to calibrate scalers.
+    """
+    print('Setting scalers with training data...')
+
+    column_definitions = self.get_column_definition()
+    id_column = utils.get_single_col_by_input_type(InputTypes.ID,
+                                                   column_definitions)
+    target_column = utils.get_single_col_by_input_type(InputTypes.TARGET,
+                                                       column_definitions)
+
+    # Extract identifiers in case required
+    self.identifiers = list(df[id_column].unique())
+
+    # Format real scalers
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    data = df[real_inputs].values
+    self._real_scalers = sklearn.preprocessing.StandardScaler().fit(data)
+    self._target_scaler = sklearn.preprocessing.StandardScaler().fit(
+        df[[target_column]].values)  # used for predictions
+
+    # Format categorical scalers
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    categorical_scalers = {}
+    num_classes = []
+    for col in categorical_inputs:
+      # Set all to str so that we don't have mixed integer/string columns
+      srs = df[col].apply(str)
+      categorical_scalers[col] = sklearn.preprocessing.LabelEncoder().fit(
+          srs.values)
+      num_classes.append(srs.nunique())
+
+    # Set categorical scaler outputs
+    self._cat_scalers = categorical_scalers
+    self._num_classes_per_cat_input = num_classes
+
+  def transform_inputs(self, df):
+    """Performs feature transformations.
+
+    This includes both feature engineering, preprocessing and normalisation.
+
+    Args:
+      df: Data frame to transform.
+
+    Returns:
+      Transformed data frame.
+
+    """
+    output = df.copy()
+
+    if self._real_scalers is None and self._cat_scalers is None:
+      raise ValueError('Scalers have not been set!')
+
+    column_definitions = self.get_column_definition()
+
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    # Format real inputs
+    output[real_inputs] = self._real_scalers.transform(df[real_inputs].values)
+
+    # Format categorical inputs
+    for col in categorical_inputs:
+      string_df = df[col].apply(str)
+      output[col] = self._cat_scalers[col].transform(string_df)
+
+    return output
+
+  def format_predictions(self, predictions):
+    """Reverts any normalisation to give predictions in original scale.
+
+    Args:
+      predictions: Dataframe of model predictions.
+
+    Returns:
+      Data frame of unnormalised predictions.
+    """
+    output = predictions.copy()
+
+    column_names = predictions.columns
+
+    for col in column_names:
+      if col not in {'forecast_time', 'identifier'}:
+        output[col] = self._target_scaler.inverse_transform(predictions[col])
+
+    return output
+
+  # Default params
+  def get_fixed_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    fixed_params = {
+        'total_time_steps': 16 + 6,
+        'num_encoder_steps': 16,
+        'num_epochs': 100,
+        'early_stopping_patience': 5,
+        'multiprocessing_workers': 5,
+    }
+
+    return fixed_params
+
+  def get_default_model_params(self):
+    """Returns default optimised model parameters."""
+
+    model_params = {
+        'dropout_rate': 0.3,
+        'hidden_layer_size': 160,
+        'learning_rate': 0.01,
+        'minibatch_size': 64,
+        'max_gradient_norm': 0.01,
+        'num_heads': 1,
+        'stack_size': 1
+    }
+
+    return model_params

examples/benchmarks/TFT/data_formatters/traffic.py

@@ -0,0 +1,117 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Custom formatting functions for Traffic dataset.
+
+Defines dataset specific column definitions and data transformations. This also
+performs z-score normalization across the entire dataset, hence re-uses most of
+the same functions as volatility.
+"""
+
+import data_formatters.base
+import data_formatters.volatility
+
+VolatilityFormatter = data_formatters.volatility.VolatilityFormatter
+DataTypes = data_formatters.base.DataTypes
+InputTypes = data_formatters.base.InputTypes
+
+
+class TrafficFormatter(VolatilityFormatter):
+  """Defines and formats data for the traffic dataset.
+
+  This also performs z-score normalization across the entire dataset, hence
+  re-uses most of the same functions as volatility.
+
+  Attributes:
+    column_definition: Defines input and data type of column used in the
+      experiment.
+    identifiers: Entity identifiers used in experiments.
+  """
+
+  _column_definition = [
+      ('id', DataTypes.REAL_VALUED, InputTypes.ID),
+      ('hours_from_start', DataTypes.REAL_VALUED, InputTypes.TIME),
+      ('values', DataTypes.REAL_VALUED, InputTypes.TARGET),
+      ('time_on_day', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('day_of_week', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('hours_from_start', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('categorical_id', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+  ]
+
+  def split_data(self, df, valid_boundary=151, test_boundary=166):
+    """Splits data frame into training-validation-test data frames.
+
+    This also calibrates scaling object, and transforms data for each split.
+
+    Args:
+      df: Source data frame to split.
+      valid_boundary: Starting year for validation data
+      test_boundary: Starting year for test data
+
+    Returns:
+      Tuple of transformed (train, valid, test) data.
+    """
+
+    print('Formatting train-valid-test splits.')
+
+    index = df['sensor_day']
+    train = df.loc[index < valid_boundary]
+    valid = df.loc[(index >= valid_boundary - 7) & (index < test_boundary)]
+    test = df.loc[index >= test_boundary - 7]
+
+    self.set_scalers(train)
+
+    return (self.transform_inputs(data) for data in [train, valid, test])
+
+  # Default params
+  def get_fixed_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    fixed_params = {
+        'total_time_steps': 8 * 24,
+        'num_encoder_steps': 7 * 24,
+        'num_epochs': 100,
+        'early_stopping_patience': 5,
+        'multiprocessing_workers': 5
+    }
+
+    return fixed_params
+
+  def get_default_model_params(self):
+    """Returns default optimised model parameters."""
+
+    model_params = {
+        'dropout_rate': 0.3,
+        'hidden_layer_size': 320,
+        'learning_rate': 0.001,
+        'minibatch_size': 128,
+        'max_gradient_norm': 100.,
+        'num_heads': 4,
+        'stack_size': 1
+    }
+
+    return model_params
+
+  def get_num_samples_for_calibration(self):
+    """Gets the default number of training and validation samples.
+
+    Use to sub-sample the data for network calibration and a value of -1 uses
+    all available samples.
+
+    Returns:
+      Tuple of (training samples, validation samples)
+    """
+    return 450000, 50000

examples/benchmarks/TFT/data_formatters/volatility.py

@@ -0,0 +1,214 @@
+# coding=utf-8
+# Copyright 2020 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Custom formatting functions for Volatility dataset.
+
+Defines dataset specific column definitions and data transformations.
+"""
+
+import data_formatters.base
+import libs.utils as utils
+import sklearn.preprocessing
+
+GenericDataFormatter = data_formatters.base.GenericDataFormatter
+DataTypes = data_formatters.base.DataTypes
+InputTypes = data_formatters.base.InputTypes
+
+
+class VolatilityFormatter(GenericDataFormatter):
+  """Defines and formats data for the volatility dataset.
+
+  Attributes:
+    column_definition: Defines input and data type of column used in the
+      experiment.
+    identifiers: Entity identifiers used in experiments.
+  """
+
+  _column_definition = [
+      ('Symbol', DataTypes.CATEGORICAL, InputTypes.ID),
+      ('date', DataTypes.DATE, InputTypes.TIME),
+      ('log_vol', DataTypes.REAL_VALUED, InputTypes.TARGET),
+      ('open_to_close', DataTypes.REAL_VALUED, InputTypes.OBSERVED_INPUT),
+      ('days_from_start', DataTypes.REAL_VALUED, InputTypes.KNOWN_INPUT),
+      ('day_of_week', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('day_of_month', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('week_of_year', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('month', DataTypes.CATEGORICAL, InputTypes.KNOWN_INPUT),
+      ('Region', DataTypes.CATEGORICAL, InputTypes.STATIC_INPUT),
+  ]
+
+  def __init__(self):
+    """Initialises formatter."""
+
+    self.identifiers = None
+    self._real_scalers = None
+    self._cat_scalers = None
+    self._target_scaler = None
+    self._num_classes_per_cat_input = None
+
+  def split_data(self, df, valid_boundary=2016, test_boundary=2018):
+    """Splits data frame into training-validation-test data frames.
+
+    This also calibrates scaling object, and transforms data for each split.
+
+    Args:
+      df: Source data frame to split.
+      valid_boundary: Starting year for validation data
+      test_boundary: Starting year for test data
+
+    Returns:
+      Tuple of transformed (train, valid, test) data.
+    """
+
+    print('Formatting train-valid-test splits.')
+
+    index = df['year']
+    train = df.loc[index < valid_boundary]
+    valid = df.loc[(index >= valid_boundary) & (index < test_boundary)]
+    test = df.loc[index >= test_boundary]
+
+    self.set_scalers(train)
+
+    return (self.transform_inputs(data) for data in [train, valid, test])
+
+  def set_scalers(self, df):
+    """Calibrates scalers using the data supplied.
+
+    Args:
+      df: Data to use to calibrate scalers.
+    """
+    print('Setting scalers with training data...')
+
+    column_definitions = self.get_column_definition()
+    id_column = utils.get_single_col_by_input_type(InputTypes.ID,
+                                                   column_definitions)
+    target_column = utils.get_single_col_by_input_type(InputTypes.TARGET,
+                                                       column_definitions)
+
+    # Extract identifiers in case required
+    self.identifiers = list(df[id_column].unique())
+
+    # Format real scalers
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    data = df[real_inputs].values
+    self._real_scalers = sklearn.preprocessing.StandardScaler().fit(data)
+    self._target_scaler = sklearn.preprocessing.StandardScaler().fit(
+        df[[target_column]].values)  # used for predictions
+
+    # Format categorical scalers
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    categorical_scalers = {}
+    num_classes = []
+    for col in categorical_inputs:
+      # Set all to str so that we don't have mixed integer/string columns
+      srs = df[col].apply(str)
+      categorical_scalers[col] = sklearn.preprocessing.LabelEncoder().fit(
+          srs.values)
+      num_classes.append(srs.nunique())
+
+    # Set categorical scaler outputs
+    self._cat_scalers = categorical_scalers
+    self._num_classes_per_cat_input = num_classes
+
+  def transform_inputs(self, df):
+    """Performs feature transformations.
+
+    This includes both feature engineering, preprocessing and normalisation.
+
+    Args:
+      df: Data frame to transform.
+
+    Returns:
+      Transformed data frame.
+
+    """
+    output = df.copy()
+
+    if self._real_scalers is None and self._cat_scalers is None:
+      raise ValueError('Scalers have not been set!')
+
+    column_definitions = self.get_column_definition()
+
+    real_inputs = utils.extract_cols_from_data_type(
+        DataTypes.REAL_VALUED, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+    categorical_inputs = utils.extract_cols_from_data_type(
+        DataTypes.CATEGORICAL, column_definitions,
+        {InputTypes.ID, InputTypes.TIME})
+
+    # Format real inputs
+    output[real_inputs] = self._real_scalers.transform(df[real_inputs].values)
+
+    # Format categorical inputs
+    for col in categorical_inputs:
+      string_df = df[col].apply(str)
+      output[col] = self._cat_scalers[col].transform(string_df)
+
+    return output
+
+  def format_predictions(self, predictions):
+    """Reverts any normalisation to give predictions in original scale.
+
+    Args:
+      predictions: Dataframe of model predictions.
+
+    Returns:
+      Data frame of unnormalised predictions.
+    """
+    output = predictions.copy()
+
+    column_names = predictions.columns
+
+    for col in column_names:
+      if col not in {'forecast_time', 'identifier'}:
+        output[col] = self._target_scaler.inverse_transform(predictions[col])
+
+    return output
+
+  # Default params
+  def get_fixed_params(self):
+    """Returns fixed model parameters for experiments."""
+
+    fixed_params = {
+        'total_time_steps': 252 + 5,
+        'num_encoder_steps': 252,
+        'num_epochs': 100,
+        'early_stopping_patience': 5,
+        'multiprocessing_workers': 5,
+    }
+
+    return fixed_params
+
+  def get_default_model_params(self):
+    """Returns default optimised model parameters."""
+
+    model_params = {
+        'dropout_rate': 0.3,
+        'hidden_layer_size': 160,
+        'learning_rate': 0.01,
+        'minibatch_size': 64,
+        'max_gradient_norm': 0.01,
+        'num_heads': 1,
+        'stack_size': 1
+    }
+
+    return model_params