evaluate.py

import argparse
import io
import json
import logging
import os
import warnings
from collections import defaultdict
from collections import namedtuple
from typing import List, Optional, Any, Text, Dict, Tuple

import numpy as np
from sklearn.exceptions import UndefinedMetricWarning
from tqdm import tqdm

from rasa_core import training, cli
from rasa_core import utils
from rasa_core.agent import Agent
from rasa_core.events import ActionExecuted, UserUttered
from rasa_core.interpreter import NaturalLanguageInterpreter
from rasa_core.policies import SimplePolicyEnsemble
from rasa_core.trackers import DialogueStateTracker
from rasa_core.training.generator import TrainingDataGenerator
from rasa_core.utils import (
    AvailableEndpoints, pad_list_to_size,
    set_default_subparser)
from rasa_nlu import utils as nlu_utils
from rasa_nlu.evaluate import plot_confusion_matrix, get_evaluation_metrics

logger = logging.getLogger(__name__)

StoryEvalution = namedtuple("StoryEvaluation",
                            "evaluation_store "
                            "failed_stories "
                            "action_list "
                            "in_training_data_fraction")


def create_argument_parser():
    """Create argument parser for the evaluate script."""

    parser = argparse.ArgumentParser(
        description='evaluates a dialogue model')
    parent_parser = argparse.ArgumentParser(add_help=False)
    add_args_to_parser(parent_parser)
    cli.arguments.add_model_and_story_group(parent_parser,
                                            allow_pretrained_model=False)
    utils.add_logging_option_arguments(parent_parser)
    subparsers = parser.add_subparsers(help='mode', dest='mode')
    subparsers.add_parser('default',
                          help='default mode: evaluate a dialogue'
                               ' model',
                          parents=[parent_parser])
    subparsers.add_parser('compare',
                          help='compare mode: evaluate multiple'
                               ' dialogue models to compare '
                               'policies',
                          parents=[parent_parser])

    return parser


def add_args_to_parser(parser):
    parser.add_argument(
        '-m', '--max_stories',
        type=int,
        help="maximum number of stories to test on")
    parser.add_argument(
        '-u', '--nlu',
        type=str,
        help="nlu model to run with the server. None for regex interpreter")
    parser.add_argument(
        '-o', '--output',
        type=str,
        default="results",
        help="output path for the any files created from the evaluation")
    parser.add_argument(
        '--e2e', '--end-to-end',
        action='store_true',
        help="Run an end-to-end evaluation for combined action and "
             "intent prediction. Requires a story file in end-to-end "
             "format.")
    parser.add_argument(
        '--endpoints',
        default=None,
        help="Configuration file for the connectors as a yml file")
    parser.add_argument(
        '--fail_on_prediction_errors',
        action='store_true',
        help="If a prediction error is encountered, an exception "
             "is thrown. This can be used to validate stories during "
             "tests, e.g. on travis.")

    cli.arguments.add_core_model_arg(parser)

    return parser


class EvaluationStore(object):
    """Class storing action, intent and entity predictions and targets."""

    def __init__(
            self,
            action_predictions: Optional[List[str]] = None,
            action_targets: Optional[List[str]] = None,
            intent_predictions: Optional[List[str]] = None,
            intent_targets: Optional[List[str]] = None,
            entity_predictions: Optional[List[Dict[Text, Any]]] = None,
            entity_targets: Optional[List[Dict[Text, Any]]] = None
    ) -> None:
        self.action_predictions = action_predictions or []
        self.action_targets = action_targets or []
        self.intent_predictions = intent_predictions or []
        self.intent_targets = intent_targets or []
        self.entity_predictions = entity_predictions or []
        self.entity_targets = entity_targets or []

    def add_to_store(
            self,
            action_predictions: Optional[List[str]] = None,
            action_targets: Optional[List[str]] = None,
            intent_predictions: Optional[List[str]] = None,
            intent_targets: Optional[List[str]] = None,
            entity_predictions: Optional[List[Dict[Text, Any]]] = None,
            entity_targets: Optional[List[Dict[Text, Any]]] = None
    ) -> None:
        """Add items or lists of items to the store"""
        for k, v in locals().items():
            if k != 'self' and v:
                attr = getattr(self, k)
                if isinstance(v, list):
                    attr.extend(v)
                else:
                    attr.append(v)

    def merge_store(self, other: 'EvaluationStore') -> None:
        """Add the contents of other to self"""
        self.add_to_store(action_predictions=other.action_predictions,
                          action_targets=other.action_targets,
                          intent_predictions=other.intent_predictions,
                          intent_targets=other.intent_targets,
                          entity_predictions=other.entity_predictions,
                          entity_targets=other.entity_targets)

    def has_prediction_target_mismatch(self):
        return (self.intent_predictions != self.intent_targets or
                self.entity_predictions != self.entity_targets or
                self.action_predictions != self.action_targets)

    def serialise_targets(self,
                          include_actions=True,
                          include_intents=True,
                          include_entities=False):
        targets = []
        if include_actions:
            targets += self.action_targets
        if include_intents:
            targets += self.intent_targets
        if include_entities:
            targets += self.entity_targets

        return [json.dumps(t) if isinstance(t, dict) else t for t in targets]

    def serialise_predictions(self,
                              include_actions=True,
                              include_intents=True,
                              include_entities=False):
        predictions = []

        if include_actions:
            predictions += self.action_predictions
        if include_intents:
            predictions += self.intent_predictions
        if include_entities:
            predictions += self.entity_predictions

        return [json.dumps(t) if isinstance(t, dict) else t
                for t in predictions]


class WronglyPredictedAction(ActionExecuted):
    """The model predicted the wrong action.

    Mostly used to mark wrong predictions and be able to
    dump them as stories."""

    type_name = "wrong_action"

    def __init__(self, correct_action, predicted_action,
                 policy, confidence, timestamp=None):
        self.predicted_action = predicted_action
        super(WronglyPredictedAction, self).__init__(correct_action,
                                                     policy,
                                                     confidence,
                                                     timestamp=timestamp)

    def as_story_string(self):
        return "{}   <!-- predicted: {} -->".format(self.action_name,
                                                    self.predicted_action)


class EndToEndUserUtterance(UserUttered):
    """End-to-end user utterance.

    Mostly used to print the full end-to-end user message in the
    `failed_stories.md` output file."""

    def as_story_string(self, e2e=True):
        return super(EndToEndUserUtterance, self).as_story_string(e2e=True)


class WronglyClassifiedUserUtterance(UserUttered):
    """The NLU model predicted the wrong user utterance.

    Mostly used to mark wrong predictions and be able to
    dump them as stories."""

    type_name = "wrong_utterance"

    def __init__(self,
                 text,
                 correct_intent,
                 correct_entities,
                 parse_data=None,
                 timestamp=None,
                 input_channel=None,
                 predicted_intent=None,
                 predicted_entities=None):
        self.predicted_intent = predicted_intent
        self.predicted_entities = predicted_entities

        intent = {"name": correct_intent}

        super(WronglyClassifiedUserUtterance, self).__init__(text,
                                                             intent,
                                                             correct_entities,
                                                             parse_data,
                                                             timestamp,
                                                             input_channel)

    def as_story_string(self, e2e=True):
        from rasa_core.events import md_format_message
        correct_message = md_format_message(self.text,
                                            self.intent,
                                            self.entities)
        predicted_message = md_format_message(self.text,
                                              self.predicted_intent,
                                              self.predicted_entities)
        return ("{}: {}   <!-- predicted: {}: {} -->"
                "").format(self.intent.get("name"),
                           correct_message,
                           self.predicted_intent,
                           predicted_message)


def _generate_trackers(resource_name, agent, max_stories=None, use_e2e=False):
    story_graph = training.extract_story_graph(resource_name, agent.domain,
                                               agent.interpreter, use_e2e)
    g = TrainingDataGenerator(story_graph, agent.domain,
                              use_story_concatenation=False,
                              augmentation_factor=0,
                              tracker_limit=max_stories)
    return g.generate()


def _clean_entity_results(entity_results):
    return [{k: r[k] for k in ("start", "end", "entity", "value") if k in r}
            for r in entity_results]


def _collect_user_uttered_predictions(event,
                                      partial_tracker,
                                      fail_on_prediction_errors):
    user_uttered_eval_store = EvaluationStore()

    intent_gold = event.parse_data.get("true_intent")
    predicted_intent = event.parse_data.get("intent").get("name")
    if predicted_intent is None:
        predicted_intent = "None"
    user_uttered_eval_store.add_to_store(intent_predictions=predicted_intent,
                                         intent_targets=intent_gold)

    entity_gold = event.parse_data.get("true_entities")
    predicted_entities = event.parse_data.get("entities")

    if entity_gold or predicted_entities:
        if len(entity_gold) > len(predicted_entities):
            predicted_entities = pad_list_to_size(predicted_entities,
                                                  len(entity_gold),
                                                  "None")
        elif len(predicted_entities) > len(entity_gold):
            entity_gold = pad_list_to_size(entity_gold,
                                           len(predicted_entities),
                                           "None")

        user_uttered_eval_store.add_to_store(
            entity_targets=_clean_entity_results(entity_gold),
            entity_predictions=_clean_entity_results(predicted_entities)
        )

    if user_uttered_eval_store.has_prediction_target_mismatch():
        partial_tracker.update(
            WronglyClassifiedUserUtterance(
                event.text, intent_gold,
                user_uttered_eval_store.entity_predictions,
                event.parse_data,
                event.timestamp,
                event.input_channel,
                predicted_intent,
                user_uttered_eval_store.entity_targets)
        )
        if fail_on_prediction_errors:
            raise ValueError(
                "NLU model predicted a wrong intent. Failed Story:"
                " \n\n{}".format(partial_tracker.export_stories()))
    else:
        end_to_end_user_utterance = EndToEndUserUtterance(
            event.text, event.intent, event.entities)
        partial_tracker.update(end_to_end_user_utterance)

    return user_uttered_eval_store


def _collect_action_executed_predictions(processor, partial_tracker, event,
                                         fail_on_prediction_errors):
    action_executed_eval_store = EvaluationStore()

    action, policy, confidence = processor.predict_next_action(partial_tracker)

    predicted = action.name()
    gold = event.action_name

    action_executed_eval_store.add_to_store(action_predictions=predicted,
                                            action_targets=gold)

    if action_executed_eval_store.has_prediction_target_mismatch():
        partial_tracker.update(WronglyPredictedAction(gold, predicted,
                                                      event.policy,
                                                      event.confidence,
                                                      event.timestamp))
        if fail_on_prediction_errors:
            raise ValueError(
                "Model predicted a wrong action. Failed Story: "
                "\n\n{}".format(partial_tracker.export_stories()))
    else:
        partial_tracker.update(event)

    return action_executed_eval_store, policy, confidence


def _predict_tracker_actions(tracker, agent, fail_on_prediction_errors=False,
                             use_e2e=False):
    processor = agent.create_processor()
    tracker_eval_store = EvaluationStore()

    events = list(tracker.events)

    partial_tracker = DialogueStateTracker.from_events(tracker.sender_id,
                                                       events[:1],
                                                       agent.domain.slots)

    tracker_actions = []

    for event in events[1:]:
        if isinstance(event, ActionExecuted):
            action_executed_result, policy, confidence = \
                _collect_action_executed_predictions(
                    processor, partial_tracker, event,
                    fail_on_prediction_errors
                )
            tracker_eval_store.merge_store(action_executed_result)
            tracker_actions.append(
                {"action": action_executed_result.action_targets[0],
                 "predicted": action_executed_result.action_predictions[0],
                 "policy": policy,
                 "confidence": confidence}
            )
        elif use_e2e and isinstance(event, UserUttered):
            user_uttered_result = \
                _collect_user_uttered_predictions(
                    event, partial_tracker, fail_on_prediction_errors)

            tracker_eval_store.merge_store(user_uttered_result)
        else:
            partial_tracker.update(event)

    return tracker_eval_store, partial_tracker, tracker_actions


def _in_training_data_fraction(action_list):
    """Given a list of action items, returns the fraction of actions

    that were predicted using one of the Memoization policies."""
    in_training_data = [
        a["action"] for a in action_list
        if not SimplePolicyEnsemble.is_not_memo_policy(a["policy"])
    ]

    return len(in_training_data) / len(action_list)


def collect_story_predictions(
        completed_trackers: List[DialogueStateTracker],
        agent: Agent,
        fail_on_prediction_errors: bool = False,
        use_e2e: bool = False
) -> Tuple[StoryEvalution, int]:
    """Test the stories from a file, running them through the stored model."""

    story_eval_store = EvaluationStore()
    failed = []
    correct_dialogues = []
    num_stories = len(completed_trackers)

    logger.info("Evaluating {} stories\n"
                "Progress:".format(num_stories))

    action_list = []

    for tracker in tqdm(completed_trackers):
        tracker_results, predicted_tracker, tracker_actions = \
            _predict_tracker_actions(tracker, agent,
                                     fail_on_prediction_errors, use_e2e)

        story_eval_store.merge_store(tracker_results)

        action_list.extend(tracker_actions)

        if tracker_results.has_prediction_target_mismatch():
            # there is at least one wrong prediction
            failed.append(predicted_tracker)
            correct_dialogues.append(0)
        else:
            correct_dialogues.append(1)

    logger.info("Finished collecting predictions.")
    report, precision, f1, accuracy = get_evaluation_metrics(
        [1] * len(completed_trackers), correct_dialogues)

    in_training_data_fraction = _in_training_data_fraction(action_list)

    log_evaluation_table([1] * len(completed_trackers),
                         "END-TO-END" if use_e2e else "CONVERSATION",
                         report, precision, f1, accuracy,
                         in_training_data_fraction,
                         include_report=False)

    return (StoryEvalution(evaluation_store=story_eval_store,
                           failed_stories=failed,
                           action_list=action_list,
                           in_training_data_fraction=in_training_data_fraction),
            num_stories)


def log_failed_stories(failed, out_directory):
    """Take stories as a list of dicts."""
    if not out_directory:
        return
    with io.open(os.path.join(out_directory, 'failed_stories.md'), 'w',
                 encoding="utf-8") as f:
        if len(failed) == 0:
            f.write("<!-- All stories passed -->")
        else:
            for failure in failed:
                f.write(failure.export_stories())
                f.write("\n\n")


def run_story_evaluation(resource_name, agent,
                         max_stories=None,
                         out_directory=None,
                         fail_on_prediction_errors=False,
                         use_e2e=False):
    """Run the evaluation of the stories, optionally plots the results."""

    completed_trackers = _generate_trackers(resource_name, agent,
                                            max_stories, use_e2e)

    story_evaluation, _ = collect_story_predictions(completed_trackers, agent,
                                                    fail_on_prediction_errors,
                                                    use_e2e)

    evaluation_store = story_evaluation.evaluation_store

    with warnings.catch_warnings():
        warnings.simplefilter("ignore", UndefinedMetricWarning)
        report, precision, f1, accuracy = get_evaluation_metrics(
            evaluation_store.serialise_targets(),
            evaluation_store.serialise_predictions()
        )

    if out_directory:
        plot_story_evaluation(evaluation_store.action_targets,
                              evaluation_store.action_predictions,
                              report, precision, f1, accuracy,
                              story_evaluation.in_training_data_fraction,
                              out_directory)

    log_failed_stories(story_evaluation.failed_stories, out_directory)

    return {
        "report": report,
        "precision": precision,
        "f1": f1,
        "accuracy": accuracy,
        "actions": story_evaluation.action_list,
        "in_training_data_fraction":
            story_evaluation.in_training_data_fraction,
        "is_end_to_end_evaluation": use_e2e
    }


def log_evaluation_table(golds, name,
                         report, precision, f1, accuracy,
                         in_training_data_fraction,
                         include_report=True):  # pragma: no cover
    """Log the sklearn evaluation metrics."""
    logger.info("Evaluation Results on {} level:".format(name))
    logger.info("\tCorrect:          {} / {}"
                "".format(int(len(golds) * accuracy), len(golds)))
    logger.info("\tF1-Score:         {:.3f}".format(f1))
    logger.info("\tPrecision:        {:.3f}".format(precision))
    logger.info("\tAccuracy:         {:.3f}".format(accuracy))
    logger.info("\tIn-data fraction: {:.3g}"
                "".format(in_training_data_fraction))

    if include_report:
        logger.info("\tClassification report: \n{}".format(report))


def plot_story_evaluation(test_y, predictions,
                          report, precision, f1, accuracy,
                          in_training_data_fraction,
                          out_directory):
    """Plot the results of story evaluation"""
    from sklearn.metrics import confusion_matrix
    from sklearn.utils.multiclass import unique_labels
    import matplotlib.pyplot as plt

    log_evaluation_table(test_y, "ACTION",
                         report, precision, f1, accuracy,
                         in_training_data_fraction,
                         include_report=True)

    cnf_matrix = confusion_matrix(test_y, predictions)

    plot_confusion_matrix(cnf_matrix,
                          classes=unique_labels(test_y, predictions),
                          title='Action Confusion matrix')

    fig = plt.gcf()
    fig.set_size_inches(int(20), int(20))
    fig.savefig(os.path.join(out_directory, "story_confmat.pdf"),
                bbox_inches='tight')


def run_comparison_evaluation(models: Text,
                              stories_file: Text,
                              output: Text) -> None:
    """Evaluates multiple trained models on a test set"""

    num_correct = defaultdict(list)

    for run in nlu_utils.list_subdirectories(models):
        num_correct_run = defaultdict(list)

        for model in sorted(nlu_utils.list_subdirectories(run)):
            logger.info("Evaluating model {}".format(model))

            agent = Agent.load(model)

            completed_trackers = _generate_trackers(stories_file, agent)

            story_eval_store, no_of_stories = \
                collect_story_predictions(completed_trackers,
                                          agent)

            failed_stories = story_eval_store.failed_stories
            policy_name = ''.join(
                [i for i in os.path.basename(model) if not i.isdigit()])
            num_correct_run[policy_name].append(no_of_stories -
                                                len(failed_stories))

        for k, v in num_correct_run.items():
            num_correct[k].append(v)

    utils.dump_obj_as_json_to_file(os.path.join(output, 'results.json'),
                                   num_correct)


def plot_curve(output: Text, no_stories: List[int]) -> None:
    """Plot the results from run_comparison_evaluation.

    Args:
        output: Output directory to save resulting plots to
        no_stories: Number of stories per run
    """
    import matplotlib.pyplot as plt

    ax = plt.gca()

    # load results from file
    data = utils.read_json_file(os.path.join(output, 'results.json'))
    x = no_stories

    # compute mean of all the runs for keras/embed policies
    for label in data.keys():
        if len(data[label]) == 0:
            continue
        mean = np.mean(data[label], axis=0)
        std = np.std(data[label], axis=0)
        ax.plot(x, mean, label=label, marker='.')
        ax.fill_between(x,
                        [m - s for m, s in zip(mean, std)],
                        [m + s for m, s in zip(mean, std)],
                        color='#6b2def',
                        alpha=0.2)
    ax.legend(loc=4)
    ax.set_xlabel("Number of stories present during training")
    ax.set_ylabel("Number of correct test stories")
    plt.savefig(os.path.join(output, 'model_comparison_graph.pdf'),
                format='pdf')
    plt.show()


if __name__ == '__main__':
    # Running as standalone python application
    arg_parser = create_argument_parser()
    set_default_subparser(arg_parser, 'default')
    cmdline_arguments = arg_parser.parse_args()

    logging.basicConfig(level=cmdline_arguments.loglevel)
    _endpoints = AvailableEndpoints.read_endpoints(cmdline_arguments.endpoints)

    if cmdline_arguments.output:
        nlu_utils.create_dir(cmdline_arguments.output)

    if not cmdline_arguments.core:
        raise ValueError("you must provide a core model directory to evaluate "
                         "using -d / --core")
    if cmdline_arguments.mode == 'default':

        _interpreter = NaturalLanguageInterpreter.create(cmdline_arguments.nlu,
                                                         _endpoints.nlu)

        _agent = Agent.load(cmdline_arguments.core,
                            interpreter=_interpreter)

        stories = cli.stories_from_cli_args(cmdline_arguments)

        run_story_evaluation(stories,
                             _agent,
                             cmdline_arguments.max_stories,
                             cmdline_arguments.output,
                             cmdline_arguments.fail_on_prediction_errors,
                             cmdline_arguments.e2e)

    elif cmdline_arguments.mode == 'compare':
        run_comparison_evaluation(cmdline_arguments.core,
                                  cmdline_arguments.stories,
                                  cmdline_arguments.output)

        story_n_path = os.path.join(cmdline_arguments.core, 'num_stories.json')

        number_of_stories = utils.read_json_file(story_n_path)
        plot_curve(cmdline_arguments.output, number_of_stories)

    logger.info("Finished evaluation")