model_data_utils.py

import typing
import copy
import numpy as np
import scipy.sparse
from collections import defaultdict, OrderedDict
from typing import List, Optional, Text, Dict, Tuple, Union, Any, DefaultDict, cast

from rasa.nlu.constants import TOKENS_NAMES
from rasa.utils.tensorflow.model_data import Data, FeatureArray
from rasa.utils.tensorflow.constants import MASK, IDS
from rasa.shared.nlu.training_data.message import Message
from rasa.shared.nlu.constants import (
    TEXT,
    ENTITIES,
    ENTITY_ATTRIBUTE_TYPE,
    ENTITY_ATTRIBUTE_GROUP,
    ENTITY_ATTRIBUTE_ROLE,
)

if typing.TYPE_CHECKING:
    from rasa.shared.nlu.training_data.features import Features
    from rasa.nlu.extractors.extractor import EntityTagSpec

TAG_ID_ORIGIN = "tag_id_origin"


def featurize_training_examples(
    training_examples: List[Message],
    attributes: List[Text],
    entity_tag_specs: Optional[List["EntityTagSpec"]] = None,
    featurizers: Optional[List[Text]] = None,
    bilou_tagging: bool = False,
) -> Tuple[List[Dict[Text, List["Features"]]], Dict[Text, Dict[Text, List[int]]]]:
    """Converts training data into a list of attribute to features.

    Possible attributes are, for example, INTENT, RESPONSE, TEXT, ACTION_TEXT,
    ACTION_NAME or ENTITIES.
    Also returns sparse feature sizes for each attribute. It could look like this:
    {TEXT: {FEATURE_TYPE_SEQUENCE: [16, 32], FEATURE_TYPE_SENTENCE: [16, 32]}}.

    Args:
        training_examples: the list of training examples
        attributes: the attributes to consider
        entity_tag_specs: the entity specs
        featurizers: the featurizers to consider
        bilou_tagging: indicates whether BILOU tagging should be used or not

    Returns:
        A list of attribute to features.
        A dictionary of attribute to feature sizes.
    """
    output = []
    if not entity_tag_specs:
        entity_tag_specs = []

    for example in training_examples:
        attribute_to_features: Dict[Text, List["Features"]] = {}
        for attribute in attributes:
            if attribute == ENTITIES:
                attribute_to_features[attribute] = []
                # in case of entities add the tag_ids
                for tag_spec in entity_tag_specs:
                    attribute_to_features[attribute].append(
                        get_tag_ids(example, tag_spec, bilou_tagging)
                    )
            elif attribute in example.data:
                attribute_to_features[attribute] = example.get_all_features(
                    attribute, featurizers
                )
        output.append(attribute_to_features)

    sparse_feature_sizes = {}
    if output and training_examples:
        sparse_feature_sizes = _collect_sparse_feature_sizes(
            featurized_example=output[0],
            training_example=training_examples[0],
            featurizers=featurizers,
        )
    return output, sparse_feature_sizes


def _collect_sparse_feature_sizes(
    featurized_example: Dict[Text, List["Features"]],
    training_example: Message,
    featurizers: Optional[List[Text]] = None,
) -> Dict[Text, Dict[Text, List[int]]]:
    """Collects sparse feature sizes for all attributes that have sparse features.

    Returns sparse feature sizes for each attribute. It could look like this:
    {TEXT: {FEATURE_TYPE_SEQUENCE: [16, 32], FEATURE_TYPE_SENTENCE: [16, 32]}}.

    Args:
        featurized_example: a featurized example
        training_example: a training example
        featurizers: the featurizers to consider

    Returns:
        A dictionary of attribute to feature sizes.
    """
    sparse_feature_sizes = {}
    sparse_attributes = []
    for attribute, features in featurized_example.items():
        if features and features[0].is_sparse():
            sparse_attributes.append(attribute)
    for attribute in sparse_attributes:
        sparse_feature_sizes[attribute] = training_example.get_sparse_feature_sizes(
            attribute=attribute, featurizers=featurizers
        )
    return sparse_feature_sizes


def get_tag_ids(
    example: Message, tag_spec: "EntityTagSpec", bilou_tagging: bool
) -> "Features":
    """Creates a feature array containing the entity tag ids of the given example.

    Args:
        example: the message
        tag_spec: entity tag spec
        bilou_tagging: indicates whether BILOU tagging should be used or not

    Returns:
        A list of features.
    """
    from rasa.nlu.test import determine_token_labels
    from rasa.nlu.utils.bilou_utils import bilou_tags_to_ids
    from rasa.shared.nlu.training_data.features import Features

    if bilou_tagging:
        _tags = bilou_tags_to_ids(example, tag_spec.tags_to_ids, tag_spec.tag_name)
    else:
        _tags = []
        for token in example.get(TOKENS_NAMES[TEXT]):
            _tag = determine_token_labels(
                token, example.get(ENTITIES), attribute_key=tag_spec.tag_name
            )
            _tags.append(tag_spec.tags_to_ids[_tag])

    # transpose to have seq_len x 1
    return Features(np.array([_tags]).T, IDS, tag_spec.tag_name, TAG_ID_ORIGIN)


def _surface_attributes(
    features: List[List[Dict[Text, List["Features"]]]],
    featurizers: Optional[List[Text]] = None,
) -> DefaultDict[Text, List[List[Optional[List["Features"]]]]]:
    """Restructure the input.

    "features" can, for example, be a dictionary of attributes (INTENT,
    TEXT, ACTION_NAME, ACTION_TEXT, ENTITIES, SLOTS, FORM) to a list of features for
    all dialogue turns in all training trackers.
    For NLU training it would just be a dictionary of attributes (either INTENT or
    RESPONSE, TEXT, and potentially ENTITIES) to a list of features for all training
    examples.

    The incoming "features" contain a dictionary as inner most value. This method
    surfaces this dictionary, so that it becomes the outer most value.

    Args:
        features: a dictionary of attributes to a list of features for all
            examples in the training data
        featurizers: the featurizers to consider

    Returns:
        A dictionary of attributes to a list of features for all examples.
    """
    # collect all attributes
    attributes = set(
        attribute
        for list_of_attribute_to_features in features
        for attribute_to_features in list_of_attribute_to_features
        for attribute in attribute_to_features.keys()
    )

    output = defaultdict(list)
    for list_of_attribute_to_features in features:
        intermediate_features = defaultdict(list)
        for attribute_to_features in list_of_attribute_to_features:
            for attribute in attributes:
                attribute_features = attribute_to_features.get(attribute)
                if featurizers:
                    attribute_features = _filter_features(
                        attribute_features, featurizers
                    )

                # if attribute is not present in the example, populate it with None
                intermediate_features[attribute].append(attribute_features)

        for key, collection_of_feature_collections in intermediate_features.items():
            output[key].append(collection_of_feature_collections)

    return output


def _filter_features(
    features: Optional[List["Features"]], featurizers: List[Text]
) -> Optional[List["Features"]]:
    """Filter the given features.

    Return only those features that are coming from one of the given featurizers.

    Args:
        features: list of features
        featurizers: names of featurizers to consider

    Returns:
        The filtered list of features.
    """
    if features is None or not featurizers:
        return features

    # it might be that the list of features also contains some tag_ids
    # the origin of the tag_ids is set to TAG_ID_ORIGIN
    # add TAG_ID_ORIGIN to the list of featurizers to make sure that we keep the
    # tag_ids
    featurizers.append(TAG_ID_ORIGIN)

    # filter the features
    return [f for f in features if f.origin in featurizers]


def _create_fake_features(
    all_features: List[List[List["Features"]]],
) -> List["Features"]:
    """Computes default feature values.

    All given features should have the same type, e.g. dense or sparse.

    Args:
        all_features: list containing all feature values encountered in the dataset
        for an attribute.

    Returns:
        The default features
    """
    example_features = next(
        iter(
            [
                list_of_features
                for list_of_list_of_features in all_features
                for list_of_features in list_of_list_of_features
                if list_of_features is not None
            ]
        )
    )

    # create fake_features for Nones
    fake_features = []
    for _features in example_features:
        new_features = copy.deepcopy(_features)
        if _features.is_dense():
            new_features.features = np.zeros(
                (0, _features.features.shape[-1]), _features.features.dtype
            )
        if _features.is_sparse():
            new_features.features = scipy.sparse.coo_matrix(
                (0, _features.features.shape[-1]), _features.features.dtype
            )
        fake_features.append(new_features)

    return fake_features


def convert_to_data_format(
    features: Union[
        List[List[Dict[Text, List["Features"]]]], List[Dict[Text, List["Features"]]]
    ],
    fake_features: Optional[Dict[Text, List["Features"]]] = None,
    consider_dialogue_dimension: bool = True,
    featurizers: Optional[List[Text]] = None,
) -> Tuple[Data, Dict[Text, List["Features"]]]:
    """Converts the input into "Data" format.

    "features" can, for example, be a dictionary of attributes (INTENT,
    TEXT, ACTION_NAME, ACTION_TEXT, ENTITIES, SLOTS, FORM) to a list of features for
    all dialogue turns in all training trackers.
    For NLU training it would just be a dictionary of attributes (either INTENT or
    RESPONSE, TEXT, and potentially ENTITIES) to a list of features for all training
    examples.

    The "Data" format corresponds to Dict[Text, Dict[Text, List[FeatureArray]]]. It's
    a dictionary of attributes (e.g. TEXT) to a dictionary of secondary attributes
    (e.g. SEQUENCE or SENTENCE) to the list of actual features.

    Args:
        features: a dictionary of attributes to a list of features for all
            examples in the training data
        fake_features: Contains default feature values for attributes
        consider_dialogue_dimension: If set to false the dialogue dimension will be
            removed from the resulting sequence features.
        featurizers: the featurizers to consider

    Returns:
        Input in "Data" format and fake features
    """
    training = False
    if not fake_features:
        training = True
        fake_features = defaultdict(list)

    # unify format of incoming features
    if isinstance(features[0], Dict):
        features = cast(
            List[List[Dict[Text, List["Features"]]]], [[dicts] for dicts in features]
        )

    attribute_to_features = _surface_attributes(features, featurizers)

    attribute_data = {}

    # During prediction we need to iterate over the fake features attributes to

    # have all keys in the resulting model data
    if training:
        attributes = list(attribute_to_features.keys())
    else:
        attributes = list(fake_features.keys())

    # In case an attribute is not present during prediction, replace it with
    # None values that will then be replaced by fake features
    dialogue_length = 1
    num_examples = 1
    for _features in attribute_to_features.values():
        num_examples = max(num_examples, len(_features))
        dialogue_length = max(dialogue_length, len(_features[0]))
    absent_features = [[None] * dialogue_length] * num_examples

    for attribute in attributes:
        attribute_data[attribute] = _feature_arrays_for_attribute(
            attribute,
            absent_features,
            attribute_to_features,
            training,
            fake_features,
            consider_dialogue_dimension,
        )

    # ensure that all attributes are in the same order
    attribute_data = OrderedDict(sorted(attribute_data.items()))

    return attribute_data, fake_features


def _feature_arrays_for_attribute(
    attribute: Text,
    absent_features: List[Any],
    attribute_to_features: Dict[Text, List[List[List["Features"]]]],
    training: bool,
    fake_features: Dict[Text, List["Features"]],
    consider_dialogue_dimension: bool,
) -> Dict[Text, List[FeatureArray]]:
    """Create the features for the given attribute from the all examples features.

    Args:
        attribute: the attribute of Message to be featurized
        absent_features: list of Nones, used as features if `attribute_to_features`
            does not contain the `attribute`
        attribute_to_features: features for every example
        training: boolean indicating whether we are currently in training or not
        fake_features: zero features
        consider_dialogue_dimension: If set to false the dialogue dimension will be
          removed from the resulting sequence features.

    Returns:
        A dictionary of feature type to actual features for the given attribute.
    """
    features = (
        attribute_to_features[attribute]
        if attribute in attribute_to_features
        else absent_features
    )

    # in case some features for a specific attribute are
    # missing, replace them with a feature vector of zeros
    if training:
        fake_features[attribute] = _create_fake_features(features)

    (attribute_masks, _dense_features, _sparse_features) = _extract_features(
        features, fake_features[attribute], attribute
    )

    sparse_features = {}
    dense_features = {}

    for key, values in _sparse_features.items():
        if consider_dialogue_dimension:
            sparse_features[key] = FeatureArray(
                np.array(values, dtype=object), number_of_dimensions=4
            )
        else:
            sparse_features[key] = FeatureArray(
                np.array([v[0] for v in values], dtype=object), number_of_dimensions=3
            )

    for key, values in _dense_features.items():
        if consider_dialogue_dimension:
            dense_features[key] = FeatureArray(np.array(values, dtype=object), number_of_dimensions=4)
        else:
            dense_features[key] = FeatureArray(
                np.array([v[0] for v in values], dtype=object), number_of_dimensions=3
            )

    attribute_to_feature_arrays = {
        MASK: [FeatureArray(np.array(attribute_masks), number_of_dimensions=3)]
    }

    feature_types = set()
    feature_types.update(list(dense_features.keys()))
    feature_types.update(list(sparse_features.keys()))

    for feature_type in feature_types:
        attribute_to_feature_arrays[feature_type] = []
        if feature_type in sparse_features:
            attribute_to_feature_arrays[feature_type].append(
                sparse_features[feature_type]
            )
        if feature_type in dense_features:
            attribute_to_feature_arrays[feature_type].append(
                dense_features[feature_type]
            )

    return attribute_to_feature_arrays


def _extract_features(
    features: List[List[List["Features"]]],
    fake_features: List["Features"],
    attribute: Text,
) -> Tuple[
    List[np.ndarray],
    Dict[Text, List[List[np.ndarray]]],
    Dict[Text, List[List[scipy.sparse.spmatrix]]],
]:
    """Create masks for feature attributes and split into dense and sparse features.

    Args:
        features: all features
        fake_features: list of zero features

    Returns:
        - a list of attribute masks
        - a map of attribute to dense features
        - a map of attribute to sparse features
    """
    sparse_features = defaultdict(list)
    dense_features = defaultdict(list)
    attribute_masks = []

    for list_of_list_of_features in features:
        dialogue_sparse_features = defaultdict(list)
        dialogue_dense_features = defaultdict(list)

        # create a mask for every state
        # to capture which turn has which input
        attribute_mask = np.ones(len(list_of_list_of_features), np.float32)

        for i, list_of_features in enumerate(list_of_list_of_features):

            if list_of_features is None:
                # use zero features and set mask to zero
                attribute_mask[i] = 0
                list_of_features = fake_features

            for feature in list_of_features:
                # in case of ENTITIES, if the attribute type matches either 'entity',
                # 'role', or 'group' the features correspond to the tag ids of that
                # entity type in order to distinguish later on between the different
                # tag ids, we use the entity type as key
                if attribute == ENTITIES and feature.attribute in [
                    ENTITY_ATTRIBUTE_TYPE,
                    ENTITY_ATTRIBUTE_GROUP,
                    ENTITY_ATTRIBUTE_ROLE,
                ]:
                    key = feature.attribute
                else:
                    key = feature.type

                # all features should have the same types
                if feature.is_sparse():
                    dialogue_sparse_features[key].append(feature.features)
                else:
                    dialogue_dense_features[key].append(feature.features)

        for key, value in dialogue_sparse_features.items():
            sparse_features[key].append(value)
        for key, value in dialogue_dense_features.items():
            dense_features[key].append(value)

        # add additional dimension to attribute mask
        # to get a vector of shape (dialogue length x 1),
        # the batch dim will be added later
        attribute_mask = np.expand_dims(attribute_mask, -1)
        attribute_masks.append(attribute_mask)

    return attribute_masks, dense_features, sparse_features