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

examples/benchmarks/TFT/expt_settings/__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/expt_settings/configs.py

@@ -0,0 +1,111 @@
+# 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 configs for TFT experiments.
+
+Contains the default output paths for data, serialised models and predictions
+for the main experiments used in the publication.
+"""
+
+import os
+
+import data_formatters.electricity
+import data_formatters.favorita
+import data_formatters.traffic
+import data_formatters.volatility
+import data_formatters.qlib_Alpha158
+
+
+class ExperimentConfig(object):
+  """Defines experiment configs and paths to outputs.
+
+  Attributes:
+    root_folder: Root folder to contain all experimental outputs.
+    experiment: Name of experiment to run.
+    data_folder: Folder to store data for experiment.
+    model_folder: Folder to store serialised models.
+    results_folder: Folder to store results.
+    data_csv_path: Path to primary data csv file used in experiment.
+    hyperparam_iterations: Default number of random search iterations for
+      experiment.
+  """
+
+  default_experiments = ['volatility', 'electricity', 'traffic', 'favorita', 'Alpha158']
+
+  def __init__(self, experiment='volatility', root_folder=None):
+    """Creates configs based on default experiment chosen.
+
+    Args:
+      experiment: Name of experiment.
+      root_folder: Root folder to save all outputs of training.
+    """
+
+    if experiment not in self.default_experiments:
+      raise ValueError('Unrecognised experiment={}'.format(experiment))
+
+    # Defines all relevant paths
+    if root_folder is None:
+      root_folder = os.path.join(
+          os.path.dirname(os.path.realpath(__file__)), '..', 'outputs')
+      print('Using root folder {}'.format(root_folder))
+
+    self.root_folder = root_folder
+    self.experiment = experiment
+    self.data_folder = os.path.join(root_folder, 'data', experiment)
+    self.model_folder = os.path.join(root_folder, 'saved_models', experiment)
+    self.results_folder = os.path.join(root_folder, 'results', experiment)
+
+    # Creates folders if they don't exist
+    for relevant_directory in [
+        self.root_folder, self.data_folder, self.model_folder,
+        self.results_folder
+    ]:
+      if not os.path.exists(relevant_directory):
+        os.makedirs(relevant_directory)
+
+  @property
+  def data_csv_path(self):
+    csv_map = {
+        'volatility': 'formatted_omi_vol.csv',
+        'electricity': 'hourly_electricity.csv',
+        'traffic': 'hourly_data.csv',
+        'favorita': 'favorita_consolidated.csv',
+        'Alpha158': 'Alpha158.csv',
+    }
+
+    return os.path.join(self.data_folder, csv_map[self.experiment])
+
+  @property
+  def hyperparam_iterations(self):
+
+    return 240 if self.experiment == 'volatility' else 60
+
+  def make_data_formatter(self):
+    """Gets a data formatter object for experiment.
+
+    Returns:
+      Default DataFormatter per experiment.
+    """
+
+    data_formatter_class = {
+        'volatility': data_formatters.volatility.VolatilityFormatter,
+        'electricity': data_formatters.electricity.ElectricityFormatter,
+        'traffic': data_formatters.traffic.TrafficFormatter,
+        'favorita': data_formatters.favorita.FavoritaFormatter,
+        'Alpha158': data_formatters.qlib_Alpha158.Alpha158Formatter,
+    }
+
+    return data_formatter_class[self.experiment]()

examples/benchmarks/TFT/libs/__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/libs/hyperparam_opt.py

@@ -0,0 +1,438 @@
+# 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
+"""Classes used for hyperparameter optimisation.
+
+Two main classes exist:
+1) HyperparamOptManager used for optimisation on a single machine/GPU.
+2) DistributedHyperparamOptManager for multiple GPUs on different machines.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import os
+import shutil
+import libs.utils as utils
+import numpy as np
+import pandas as pd
+
+Deque = collections.deque
+
+
+class HyperparamOptManager:
+  """Manages hyperparameter optimisation using random search for a single GPU.
+
+  Attributes:
+    param_ranges: Discrete hyperparameter range for random search.
+    results: Dataframe of validation results.
+    fixed_params: Fixed model parameters per experiment.
+    saved_params: Dataframe of parameters trained.
+    best_score: Minimum validation loss observed thus far.
+    optimal_name: Key to best configuration.
+    hyperparam_folder: Where to save optimisation outputs.
+  """
+
+  def __init__(self,
+               param_ranges,
+               fixed_params,
+               model_folder,
+               override_w_fixed_params=True):
+    """Instantiates model.
+
+    Args:
+      param_ranges: Discrete hyperparameter range for random search.
+      fixed_params: Fixed model parameters per experiment.
+      model_folder: Folder to store optimisation artifacts.
+      override_w_fixed_params: Whether to override serialsed fixed model
+        parameters with new supplied values.
+    """
+
+    self.param_ranges = param_ranges
+
+    self._max_tries = 1000
+    self.results = pd.DataFrame()
+    self.fixed_params = fixed_params
+    self.saved_params = pd.DataFrame()
+
+    self.best_score = np.Inf
+    self.optimal_name = ""
+
+    # Setup
+    # Create folder for saving if its not there
+    self.hyperparam_folder = model_folder
+    utils.create_folder_if_not_exist(self.hyperparam_folder)
+
+    self._override_w_fixed_params = override_w_fixed_params
+
+  def load_results(self):
+    """Loads results from previous hyperparameter optimisation.
+
+    Returns:
+      A boolean indicating if previous results can be loaded.
+    """
+    print("Loading results from", self.hyperparam_folder)
+
+    results_file = os.path.join(self.hyperparam_folder, "results.csv")
+    params_file = os.path.join(self.hyperparam_folder, "params.csv")
+
+    if os.path.exists(results_file) and os.path.exists(params_file):
+
+      self.results = pd.read_csv(results_file, index_col=0)
+      self.saved_params = pd.read_csv(params_file, index_col=0)
+
+      if not self.results.empty:
+        self.results.at["loss"] = self.results.loc["loss"].apply(float)
+        self.best_score = self.results.loc["loss"].min()
+
+        is_optimal = self.results.loc["loss"] == self.best_score
+        self.optimal_name = self.results.T[is_optimal].index[0]
+
+        return True
+
+    return False
+
+  def _get_params_from_name(self, name):
+    """Returns previously saved parameters given a key."""
+    params = self.saved_params
+
+    selected_params = dict(params[name])
+
+    if self._override_w_fixed_params:
+      for k in self.fixed_params:
+        selected_params[k] = self.fixed_params[k]
+
+    return selected_params
+
+  def get_best_params(self):
+    """Returns the optimal hyperparameters thus far."""
+
+    optimal_name = self.optimal_name
+
+    return self._get_params_from_name(optimal_name)
+
+  def clear(self):
+    """Clears all previous results and saved parameters."""
+    shutil.rmtree(self.hyperparam_folder)
+    os.makedirs(self.hyperparam_folder)
+    self.results = pd.DataFrame()
+    self.saved_params = pd.DataFrame()
+
+  def _check_params(self, params):
+    """Checks that parameter map is properly defined."""
+
+    valid_fields = list(self.param_ranges.keys()) + list(
+        self.fixed_params.keys())
+    invalid_fields = [k for k in params if k not in valid_fields]
+    missing_fields = [k for k in valid_fields if k not in params]
+
+    if invalid_fields:
+      raise ValueError("Invalid Fields Found {} - Valid ones are {}".format(
+          invalid_fields, valid_fields))
+    if missing_fields:
+      raise ValueError("Missing Fields Found {} - Valid ones are {}".format(
+          missing_fields, valid_fields))
+
+  def _get_name(self, params):
+    """Returns a unique key for the supplied set of params."""
+
+    self._check_params(params)
+
+    fields = list(params.keys())
+    fields.sort()
+
+    return "_".join([str(params[k]) for k in fields])
+
+  def get_next_parameters(self, ranges_to_skip=None):
+    """Returns the next set of parameters to optimise.
+
+    Args:
+      ranges_to_skip: Explicitly defines a set of keys to skip.
+    """
+    if ranges_to_skip is None:
+      ranges_to_skip = set(self.results.index)
+
+    if not isinstance(self.param_ranges, dict):
+      raise ValueError("Only works for random search!")
+
+    param_range_keys = list(self.param_ranges.keys())
+    param_range_keys.sort()
+
+    def _get_next():
+      """Returns next hyperparameter set per try."""
+
+      parameters = {
+          k: np.random.choice(self.param_ranges[k]) for k in param_range_keys
+      }
+
+      # Adds fixed params
+      for k in self.fixed_params:
+        parameters[k] = self.fixed_params[k]
+
+      return parameters
+
+    for _ in range(self._max_tries):
+
+      parameters = _get_next()
+      name = self._get_name(parameters)
+
+      if name not in ranges_to_skip:
+        return parameters
+
+    raise ValueError("Exceeded max number of hyperparameter searches!!")
+
+  def update_score(self, parameters, loss, model, info=""):
+    """Updates the results from last optimisation run.
+
+    Args:
+      parameters: Hyperparameters used in optimisation.
+      loss: Validation loss obtained.
+      model: Model to serialised if required.
+      info: Any ancillary information to tag on to results.
+
+    Returns:
+      Boolean flag indicating if the model is the best seen so far.
+    """
+
+    if np.isnan(loss):
+      loss = np.Inf
+
+    if not os.path.isdir(self.hyperparam_folder):
+      os.makedirs(self.hyperparam_folder)
+
+    name = self._get_name(parameters)
+
+    is_optimal = self.results.empty or loss < self.best_score
+
+    # save the first model
+    if is_optimal:
+      # Try saving first, before updating info
+      if model is not None:
+        print("Optimal model found, updating")
+        model.save(self.hyperparam_folder)
+      self.best_score = loss
+      self.optimal_name = name
+
+    self.results[name] = pd.Series({"loss": loss, "info": info})
+    self.saved_params[name] = pd.Series(parameters)
+
+    self.results.to_csv(os.path.join(self.hyperparam_folder, "results.csv"))
+    self.saved_params.to_csv(os.path.join(self.hyperparam_folder, "params.csv"))
+
+    return is_optimal
+
+
+class DistributedHyperparamOptManager(HyperparamOptManager):
+  """Manages distributed hyperparameter optimisation across many gpus."""
+
+  def __init__(self,
+               param_ranges,
+               fixed_params,
+               root_model_folder,
+               worker_number,
+               search_iterations=1000,
+               num_iterations_per_worker=5,
+               clear_serialised_params=False):
+    """Instantiates optimisation manager.
+
+    This hyperparameter optimisation pre-generates #search_iterations
+    hyperparameter combinations and serialises them
+    at the start. At runtime, each worker goes through their own set of
+    parameter ranges. The pregeneration
+    allows for multiple workers to run in parallel on different machines without
+    resulting in parameter overlaps.
+
+    Args:
+      param_ranges: Discrete hyperparameter range for random search.
+      fixed_params: Fixed model parameters per experiment.
+      root_model_folder: Folder to store optimisation artifacts.
+      worker_number: Worker index definining which set of hyperparameters to
+        test.
+      search_iterations: Maximum numer of random search iterations.
+      num_iterations_per_worker: How many iterations are handled per worker.
+      clear_serialised_params: Whether to regenerate hyperparameter
+        combinations.
+    """
+
+    max_workers = int(np.ceil(search_iterations / num_iterations_per_worker))
+
+    # Sanity checks
+    if worker_number > max_workers:
+      raise ValueError(
+          "Worker number ({}) cannot be larger than the total number of workers!"
+          .format(max_workers))
+    if worker_number > search_iterations:
+      raise ValueError(
+          "Worker number ({}) cannot be larger than the max search iterations ({})!"
+          .format(worker_number, search_iterations))
+
+    print("*** Creating hyperparameter manager for worker {} ***".format(
+        worker_number))
+
+    hyperparam_folder = os.path.join(root_model_folder, str(worker_number))
+    super().__init__(
+        param_ranges,
+        fixed_params,
+        hyperparam_folder,
+        override_w_fixed_params=True)
+
+    serialised_ranges_folder = os.path.join(root_model_folder, "hyperparams")
+    if clear_serialised_params:
+      print("Regenerating hyperparameter list")
+      if os.path.exists(serialised_ranges_folder):
+        shutil.rmtree(serialised_ranges_folder)
+
+    utils.create_folder_if_not_exist(serialised_ranges_folder)
+
+    self.serialised_ranges_path = os.path.join(
+        serialised_ranges_folder, "ranges_{}.csv".format(search_iterations))
+    self.hyperparam_folder = hyperparam_folder  # override
+    self.worker_num = worker_number
+    self.total_search_iterations = search_iterations
+    self.num_iterations_per_worker = num_iterations_per_worker
+    self.global_hyperparam_df = self.load_serialised_hyperparam_df()
+    self.worker_search_queue = self._get_worker_search_queue()
+
+  @property
+  def optimisation_completed(self):
+    return False if self.worker_search_queue else True
+
+  def get_next_parameters(self):
+    """Returns next dictionary of hyperparameters to optimise."""
+    param_name = self.worker_search_queue.pop()
+
+    params = self.global_hyperparam_df.loc[param_name, :].to_dict()
+
+    # Always override!
+    for k in self.fixed_params:
+      print("Overriding saved {}: {}".format(k, self.fixed_params[k]))
+
+      params[k] = self.fixed_params[k]
+
+    return params
+
+  def load_serialised_hyperparam_df(self):
+    """Loads serialsed hyperparameter ranges from file.
+
+    Returns:
+      DataFrame containing hyperparameter combinations.
+    """
+    print("Loading params for {} search iterations form {}".format(
+        self.total_search_iterations, self.serialised_ranges_path))
+
+    if os.path.exists(self.serialised_ranges_folder):
+      df = pd.read_csv(self.serialised_ranges_path, index_col=0)
+    else:
+      print("Unable to load - regenerating serach ranges instead")
+      df = self.update_serialised_hyperparam_df()
+
+    return df
+
+  def update_serialised_hyperparam_df(self):
+    """Regenerates hyperparameter combinations and saves to file.
+
+    Returns:
+      DataFrame containing hyperparameter combinations.
+    """
+    search_df = self._generate_full_hyperparam_df()
+
+    print("Serialising params for {} search iterations to {}".format(
+        self.total_search_iterations, self.serialised_ranges_path))
+
+    search_df.to_csv(self.serialised_ranges_path)
+
+    return search_df
+
+  def _generate_full_hyperparam_df(self):
+    """Generates actual hyperparameter combinations.
+
+    Returns:
+      DataFrame containing hyperparameter combinations.
+    """
+
+    np.random.seed(131)  # for reproducibility of hyperparam list
+
+    name_list = []
+    param_list = []
+    for _ in range(self.total_search_iterations):
+      params = super().get_next_parameters(name_list)
+
+      name = self._get_name(params)
+
+      name_list.append(name)
+      param_list.append(params)
+
+    full_search_df = pd.DataFrame(param_list, index=name_list)
+
+    return full_search_df
+
+  def clear(self):  # reset when cleared
+    """Clears results for hyperparameter manager and resets."""
+    super().clear()
+    self.worker_search_queue = self._get_worker_search_queue()
+
+  def load_results(self):
+    """Load results from file and queue parameter combinations to try.
+
+    Returns:
+      Boolean indicating if results were successfully loaded.
+    """
+    success = super().load_results()
+
+    if success:
+      self.worker_search_queue = self._get_worker_search_queue()
+
+    return success
+
+  def _get_worker_search_queue(self):
+    """Generates the queue of param combinations for current worker.
+
+    Returns:
+      Queue of hyperparameter combinations outstanding.
+    """
+    global_df = self.assign_worker_numbers(self.global_hyperparam_df)
+    worker_df = global_df[global_df["worker"] == self.worker_num]
+
+    left_overs = [s for s in worker_df.index if s not in self.results.columns]
+
+    return Deque(left_overs)
+
+  def assign_worker_numbers(self, df):
+    """Updates parameter combinations with the index of the worker used.
+
+    Args:
+      df: DataFrame of parameter combinations.
+
+    Returns:
+      Updated DataFrame with worker number.
+    """
+    output = df.copy()
+
+    n = self.total_search_iterations
+    batch_size = self.num_iterations_per_worker
+
+    max_worker_num = int(np.ceil(n / batch_size))
+
+    worker_idx = np.concatenate([
+        np.tile(i + 1, self.num_iterations_per_worker)
+        for i in range(max_worker_num)
+    ])
+
+    output["worker"] = worker_idx[:len(output)]
+
+    return output

examples/benchmarks/TFT/libs/utils.py

@@ -0,0 +1,236 @@
+# 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
+"""Generic helper functions used across codebase."""
+
+import os
+import pathlib
+
+import numpy as np
+import tensorflow as tf
+from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file
+
+
+# Generic.
+def get_single_col_by_input_type(input_type, column_definition):
+  """Returns name of single column.
+
+  Args:
+    input_type: Input type of column to extract
+    column_definition: Column definition list for experiment
+  """
+
+  l = [tup[0] for tup in column_definition if tup[2] == input_type]
+
+  if len(l) != 1:
+    raise ValueError('Invalid number of columns for {}'.format(input_type))
+
+  return l[0]
+
+
+def extract_cols_from_data_type(data_type, column_definition,
+                                excluded_input_types):
+  """Extracts the names of columns that correspond to a define data_type.
+
+  Args:
+    data_type: DataType of columns to extract.
+    column_definition: Column definition to use.
+    excluded_input_types: Set of input types to exclude
+
+  Returns:
+    List of names for columns with data type specified.
+  """
+  return [
+      tup[0]
+      for tup in column_definition
+      if tup[1] == data_type and tup[2] not in excluded_input_types
+  ]
+
+
+# Loss functions.
+def tensorflow_quantile_loss(y, y_pred, quantile):
+  """Computes quantile loss for tensorflow.
+
+  Standard quantile loss as defined in the "Training Procedure" section of
+  the main TFT paper
+
+  Args:
+    y: Targets
+    y_pred: Predictions
+    quantile: Quantile to use for loss calculations (between 0 & 1)
+
+  Returns:
+    Tensor for quantile loss.
+  """
+
+  # Checks quantile
+  if quantile < 0 or quantile > 1:
+    raise ValueError(
+        'Illegal quantile value={}! Values should be between 0 and 1.'.format(
+            quantile))
+
+  prediction_underflow = y - y_pred
+  q_loss = quantile * tf.maximum(prediction_underflow, 0.) + (
+      1. - quantile) * tf.maximum(-prediction_underflow, 0.)
+
+  return tf.reduce_sum(q_loss, axis=-1)
+
+
+def numpy_normalised_quantile_loss(y, y_pred, quantile):
+  """Computes normalised quantile loss for numpy arrays.
+
+  Uses the q-Risk metric as defined in the "Training Procedure" section of the
+  main TFT paper.
+
+  Args:
+    y: Targets
+    y_pred: Predictions
+    quantile: Quantile to use for loss calculations (between 0 & 1)
+
+  Returns:
+    Float for normalised quantile loss.
+  """
+  prediction_underflow = y - y_pred
+  weighted_errors = quantile * np.maximum(prediction_underflow, 0.) \
+      + (1. - quantile) * np.maximum(-prediction_underflow, 0.)
+
+  quantile_loss = weighted_errors.mean()
+  normaliser = y.abs().mean()
+
+  return 2 * quantile_loss / normaliser
+
+
+# OS related functions.
+def create_folder_if_not_exist(directory):
+  """Creates folder if it doesn't exist.
+
+  Args:
+    directory: Folder path to create.
+  """
+  # Also creates directories recursively
+  pathlib.Path(directory).mkdir(parents=True, exist_ok=True)
+
+
+# Tensorflow related functions.
+def get_default_tensorflow_config(tf_device='gpu', gpu_id=0):
+  """Creates tensorflow config for graphs to run on CPU or GPU.
+
+  Specifies whether to run graph on gpu or cpu and which GPU ID to use for multi
+  GPU machines.
+
+  Args:
+    tf_device: 'cpu' or 'gpu'
+    gpu_id: GPU ID to use if relevant
+
+  Returns:
+    Tensorflow config.
+  """
+
+  if tf_device == 'cpu':
+    os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # for training on cpu
+    tf_config = tf.ConfigProto(
+        log_device_placement=False, device_count={'GPU': 0})
+
+  else:
+    os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id)
+
+    print('Selecting GPU ID={}'.format(gpu_id))
+
+    tf_config = tf.ConfigProto(log_device_placement=False)
+    tf_config.gpu_options.allow_growth = True
+
+  return tf_config
+
+
+def save(tf_session, model_folder, cp_name, scope=None):
+  """Saves Tensorflow graph to checkpoint.
+
+  Saves all trainiable variables under a given variable scope to checkpoint.
+
+  Args:
+    tf_session: Session containing graph
+    model_folder: Folder to save models
+    cp_name: Name of Tensorflow checkpoint
+    scope: Variable scope containing variables to save
+  """
+  # Save model
+  if scope is None:
+    saver = tf.train.Saver()
+  else:
+    var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope)
+    saver = tf.train.Saver(var_list=var_list, max_to_keep=100000)
+
+  save_path = saver.save(tf_session,
+                         os.path.join(model_folder, '{0}.ckpt'.format(cp_name)))
+  print('Model saved to: {0}'.format(save_path))
+
+
+def load(tf_session, model_folder, cp_name, scope=None, verbose=False):
+  """Loads Tensorflow graph from checkpoint.
+
+  Args:
+    tf_session: Session to load graph into
+    model_folder: Folder containing serialised model
+    cp_name: Name of Tensorflow checkpoint
+    scope: Variable scope to use.
+    verbose: Whether to print additional debugging information.
+  """
+  # Load model proper
+  load_path = os.path.join(model_folder, '{0}.ckpt'.format(cp_name))
+
+  print('Loading model from {0}'.format(load_path))
+
+  print_weights_in_checkpoint(model_folder, cp_name)
+
+  initial_vars = set(
+      [v.name for v in tf.get_default_graph().as_graph_def().node])
+
+  # Saver
+  if scope is None:
+    saver = tf.train.Saver()
+  else:
+    var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=scope)
+    saver = tf.train.Saver(var_list=var_list, max_to_keep=100000)
+  # Load
+  saver.restore(tf_session, load_path)
+  all_vars = set([v.name for v in tf.get_default_graph().as_graph_def().node])
+
+  if verbose:
+    print('Restored {0}'.format(','.join(initial_vars.difference(all_vars))))
+    print('Existing {0}'.format(','.join(all_vars.difference(initial_vars))))
+    print('All {0}'.format(','.join(all_vars)))
+
+  print('Done.')
+
+
+def print_weights_in_checkpoint(model_folder, cp_name):
+  """Prints all weights in Tensorflow checkpoint.
+
+  Args:
+    model_folder: Folder containing checkpoint
+    cp_name: Name of checkpoint
+
+  Returns:
+
+  """
+  load_path = os.path.join(model_folder, '{0}.ckpt'.format(cp_name))
+
+  print_tensors_in_checkpoint_file(
+      file_name=load_path,
+      tensor_name='',
+      all_tensors=True,
+      all_tensor_names=True)

examples/benchmarks/TFT/tft.py

@@ -0,0 +1,246 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import numpy as np
+import pandas as pd
+import tensorflow.compat.v1 as tf
+import data_formatters.base
+import expt_settings.configs
+import libs.hyperparam_opt
+import libs.tft_model
+import libs.utils as utils
+import os
+import datetime as dte
+
+
+from qlib.model.base import ModelFT
+from qlib.data.dataset import DatasetH
+from qlib.data.dataset.handler import DataHandlerLP
+
+
+
+# To register new datasets, please add them here.
+ALLOW_DATASET = ['Alpha158'] 
+DATASET_SETTING = {
+    'Alpha158': {
+        'feature_col': ['RESI5', 'WVMA5', 'RSQR5', 'KLEN', 'RSQR10', 'CORR5', 'CORD5', 'CORR10', 'ROC60', 'RESI10'],
+        'label_col': ['LABEL0'],
+    },
+}
+# To register new datasets, please add their configurations here.
+
+def get_shifted_label(data_df, shifts=5, col_shift='LABEL0'):
+    return data_df[[col_shift]].groupby('instrument').apply(lambda df: df.shift(shifts))
+
+def fill_test_na(test_df):
+    test_df_res = test_df.copy()
+    feature_cols = ~test_df_res.columns.str.contains('label', case=False)
+    test_feature_fna = test_df_res.loc[:, feature_cols].groupby('datetime').apply(lambda df: df.fillna(df.mean()))
+    test_df_res.loc[:, feature_cols] = test_feature_fna
+    return test_df_res
+
+def process_qlib_data(df, dataset, fillna=False):
+    """Prepare data to fit the TFT model.
+
+        Args:
+          df: Original DataFrame.
+          fillna: Whether to fill the data with the mean values.
+
+        Returns:
+          Transformed DataFrame.
+
+        """
+    # Several features selected manually
+    feature_col = DATASET_SETTING[dataset]['feature_col']
+    label_col = DATASET_SETTING[dataset]['label_col']
+    temp_df = df.loc[:, feature_col+label_col]
+    if fillna:
+        temp_df = fill_test_na(temp_df)
+    temp_df = temp_df.swaplevel()
+    temp_df = temp_df.sort_index()
+    temp_df = temp_df.reset_index(level=0)
+    dates = pd.to_datetime(temp_df.index)
+    temp_df['date'] = dates
+    temp_df['day_of_week'] = dates.dayofweek
+    temp_df['month'] = dates.month
+    temp_df['year'] = dates.year
+    temp_df['const'] = 1.0
+    return temp_df
+
+def process_predicted(df, col_name):
+    """Transform the TFT predicted data into Qlib format.
+
+    Args:
+      df: Original DataFrame.
+      fillna: New column name.
+
+    Returns:
+      Transformed DataFrame.
+
+    """
+    df_res = df.copy()
+    df_res = df_res.rename(columns={"forecast_time": "datetime", "identifier": "instrument", "t+0": col_name})
+    df_res = df_res.set_index(['datetime','instrument']).sort_index()
+    df_res = df_res[[col_name]]
+    return df_res
+
+def format_score(forecast_df, col_name='pred', label_shift=5):
+    pred = process_predicted(forecast_df, col_name=col_name)
+    pred = get_shifted_label(pred, shifts=-label_shift, col_shift=col_name)
+    pred = pred.dropna()[col_name]
+    return pred
+
+def transform_df(df, col_name='LABEL0'):
+    df_res = df['feature']
+    df_res[col_name] = df['label']
+    return df_res
+
+class TFTModel(ModelFT):
+    """TFT Model"""
+
+    def __init__(self, **kwargs):
+        self.model = None
+
+    def _prepare_data(self, dataset: DatasetH):
+        df_train, df_valid = dataset.prepare(
+            ["train", "valid"], col_set=["feature", "label"], data_key=DataHandlerLP.DK_L
+        )
+        return transform_df(df_train), transform_df(df_valid)
+
+    def fit(
+        self,
+        dataset: DatasetH,
+        DATASET = 'Alpha158',
+        MODEL_FOLDER = 'qlib_alpha158_model',
+        LABEL_COL = 'LABEL0',
+        LABEL_SHIFT = 5,
+        USE_GPU_ID = 0,
+        **kwargs
+    ):
+        
+        if DATASET not in ALLOW_DATASET:
+            raise AssertionError("The dataset is not supported, please make a new formatter to fit this dataset")
+        
+        dtrain, dvalid = self._prepare_data(dataset)
+        dtrain.loc[:, LABEL_COL] = get_shifted_label(dtrain, shifts=LABEL_SHIFT, col_shift=LABEL_COL)
+        dvalid.loc[:, LABEL_COL] = get_shifted_label(dvalid, shifts=LABEL_SHIFT, col_shift=LABEL_COL)
+        
+        
+        train = process_qlib_data(dtrain, DATASET, fillna=True).dropna()
+        valid = process_qlib_data(dvalid, DATASET, fillna=True).dropna()
+        
+        ExperimentConfig = expt_settings.configs.ExperimentConfig
+        config = ExperimentConfig(DATASET)
+        self.data_formatter = config.make_data_formatter()
+        self.model_folder = MODEL_FOLDER
+        self.gpu_id = USE_GPU_ID
+        self.label_shift = LABEL_SHIFT
+        self.expt_name = DATASET
+        self.label_col = LABEL_COL
+        
+        use_gpu = (True, self.gpu_id)
+        #===========================Training Process===========================
+        ModelClass = libs.tft_model.TemporalFusionTransformer
+        if not isinstance(self.data_formatter, data_formatters.base.GenericDataFormatter):
+             raise ValueError(
+            "Data formatters should inherit from" +
+            "AbstractDataFormatter! Type={}".format(type(self.data_formatter)))
+
+        default_keras_session = tf.keras.backend.get_session()
+
+        if use_gpu[0]:
+            self.tf_config = utils.get_default_tensorflow_config(tf_device="gpu", gpu_id=use_gpu[1])
+        else:
+            self.tf_config = utils.get_default_tensorflow_config(tf_device="cpu")
+
+        self.data_formatter.set_scalers(train)
+
+        # Sets up default params
+        fixed_params = self.data_formatter.get_experiment_params()
+        params = self.data_formatter.get_default_model_params()
+
+        # Wendi: 合并调优的参数和非调优的参数
+        params = {**params, **fixed_params}
+
+        if not os.path.exists(self.model_folder):
+            os.makedirs(self.model_folder)
+        params['model_folder'] = self.model_folder
+
+        print("*** Begin training ***")
+        best_loss = np.Inf
+
+        tf.reset_default_graph()
+        
+        self.tf_graph = tf.Graph()
+        with self.tf_graph.as_default():
+            self.sess = tf.Session(config=self.tf_config)
+            tf.keras.backend.set_session(self.sess)
+            self.model = ModelClass(params, use_cudnn=use_gpu[0])
+            self.sess.run(tf.global_variables_initializer())
+            self.model.fit(train_df=train, valid_df=valid)
+            print("*** Finished training ***")
+            saved_model_dir = self.model_folder+'/'+'saved_model'
+            if not os.path.exists(saved_model_dir):
+                os.makedirs(saved_model_dir)
+            self.model.save(saved_model_dir)
+            
+            def extract_numerical_data(data):
+                """Strips out forecast time and identifier columns."""
+                return data[[
+                    col for col in data.columns
+                    if col not in {"forecast_time", "identifier"}
+                    ]]
+            
+            #p50_loss = utils.numpy_normalised_quantile_loss(
+            #    extract_numerical_data(targets), extract_numerical_data(p50_forecast),
+            #    0.5)
+            #p90_loss = utils.numpy_normalised_quantile_loss(
+            #    extract_numerical_data(targets), extract_numerical_data(p90_forecast),
+            #    0.9)
+            tf.keras.backend.set_session(default_keras_session)
+        print("Training completed.".format(dte.datetime.now()))
+        #===========================Training Process===========================
+
+    def predict(self, dataset):
+        if self.model is None:
+            raise ValueError("model is not fitted yet!")
+        d_test = dataset.prepare("test", col_set=["feature", "label"])
+        d_test = transform_df(d_test)
+        d_test.loc[:, self.label_col] = get_shifted_label(d_test, shifts=self.label_shift, col_shift=self.label_col)
+        test = process_qlib_data(d_test, self.expt_name, fillna=True).dropna()
+        
+        use_gpu = (True, self.gpu_id)
+        #===========================Predicting Process===========================
+        default_keras_session = tf.keras.backend.get_session()
+        
+        # Sets up default params
+        fixed_params = self.data_formatter.get_experiment_params()
+        params = self.data_formatter.get_default_model_params()
+        params = {**params, **fixed_params}
+        
+        
+        print("*** Begin predicting ***")
+        tf.reset_default_graph()
+        
+        with self.tf_graph.as_default():
+            tf.keras.backend.set_session(self.sess)
+            output_map = self.model.predict(test, return_targets=True)
+            targets = self.data_formatter.format_predictions(output_map["targets"])
+            p50_forecast = self.data_formatter.format_predictions(output_map["p50"])
+            p90_forecast = self.data_formatter.format_predictions(output_map["p90"])
+            tf.keras.backend.set_session(default_keras_session)
+            
+        predict = format_score(p90_forecast, 'pred', self.label_shift)
+        label = format_score(targets, 'label', self.label_shift)
+        #===========================Predicting Process===========================
+        return predict, label
+
+    def finetune(self, dataset: DatasetH):
+        """
+        finetune model
+        Parameters
+        ----------
+        dataset : DatasetH
+            dataset for finetuning
+        """
+        pass

examples/benchmarks/TFT/workflow_by_code_tft.py

@@ -0,0 +1,130 @@
+ #Copyright (c) Microsoft Corporation.
+ #Licensed under the MIT License.
+
+import sys
+from pathlib import Path
+
+import qlib
+import pandas as pd
+from qlib.config import REG_CN
+from qlib.contrib.model.pytorch_lstm import LSTM
+from qlib.contrib.data.handler import ALPHA360_Denoise
+from qlib.contrib.strategy.strategy import TopkDropoutStrategy
+from qlib.contrib.evaluate import (
+    backtest as normal_backtest,
+    risk_analysis,
+)
+from qlib.utils import exists_qlib_data
+
+# from qlib.model.learner import train_model
+from qlib.utils import init_instance_by_config
+
+import pickle
+from tft import TFTModel
+
+if __name__ == "__main__":
+
+    # use default data
+    provider_uri = "~/.qlib/qlib_data/cn_data"  # target_dir
+    if not exists_qlib_data(provider_uri):
+        print(f"Qlib data is not found in {provider_uri}")
+        sys.path.append(str(Path(__file__).resolve().parent.parent.joinpath("scripts")))
+        from get_data import GetData
+
+        GetData().qlib_data_cn(target_dir=provider_uri)
+
+    qlib.init(provider_uri=provider_uri, region=REG_CN)
+
+    MARKET = "csi300"
+    BENCHMARK = "SH000300"
+
+    ###################################
+    # train model
+    ###################################
+    DATA_HANDLER_CONFIG = {
+        "start_time": "2008-01-01",
+        "end_time": "2020-08-01",
+        "fit_start_time": "2008-01-01",
+        "fit_end_time": "2014-12-31",
+        "instruments": MARKET,
+    }
+
+    TRAINER_CONFIG = {
+        "train_start_time": "2008-01-01",
+        "train_end_time": "2014-12-31",
+        "validate_start_time": "2015-01-01",
+        "validate_end_time": "2016-12-31",
+        "test_start_time": "2017-01-01",
+        "test_end_time": "2020-08-01",
+    }
+
+    task = {
+        "dataset": {
+            "class": "DatasetH",
+            "module_path": "qlib.data.dataset",
+            "kwargs": {
+                'handler': {
+                    "class": "Alpha158",
+                    "module_path": "qlib.contrib.data.handler",
+                    "kwargs": DATA_HANDLER_CONFIG
+                },
+                'segments': {
+                    'train': ("2008-01-01", "2014-12-31"),
+                    'valid': ("2015-01-01", "2016-12-31",),
+                    'test': ("2017-01-01", "2020-08-01",),
+                }
+            }
+        }
+        # You shoud record the data in specific sequence
+        # "record": ['SignalRecord', 'SigAnaRecord', 'PortAnaRecord'],
+    }
+    
+
+    model = TFTModel()
+    dataset = init_instance_by_config(task["dataset"])
+    model.fit(dataset)
+
+    pred_score, label_score = model.predict(dataset)
+
+    # save pred_score to file
+    pred_score_path = Path("~/tmp/qlib/pred_score.pkl").expanduser()
+    pred_score_path.parent.mkdir(exist_ok=True, parents=True)
+    pred_score.to_pickle(pred_score_path)
+
+
+    ###################################
+    # backtest
+    ###################################
+    STRATEGY_CONFIG = {
+        "topk": 50,
+        "n_drop": 5,
+    }
+    BACKTEST_CONFIG = {
+        "verbose": False,
+        "limit_threshold": 0.095,
+        "account": 100000000,
+        "benchmark": BENCHMARK,
+        "deal_price": "close",
+        "open_cost": 0.0005,
+        "close_cost": 0.0015,
+        "min_cost": 5,
+    }
+
+    # use default strategy
+    # custom Strategy, refer to: TODO: Strategy API url
+    strategy = TopkDropoutStrategy(**STRATEGY_CONFIG)
+    report_normal, positions_normal = normal_backtest(pred_score, strategy=strategy, **BACKTEST_CONFIG)
+
+    ###################################
+    # analyze
+    # If need a more detailed analysis, refer to: examples/train_and_bakctest.ipynb
+    ###################################
+    analysis = dict()
+    analysis["excess_return_without_cost"] = risk_analysis(report_normal["return"] - report_normal["bench"])
+    analysis["excess_return_with_cost"] = risk_analysis(
+        report_normal["return"] - report_normal["bench"] - report_normal["cost"]
+    )
+    analysis_df = pd.concat(analysis)  # type: pd.DataFrame
+    print(analysis_df)
+
+