llmeval-env/lib/python3.10/site-packages/evaluate/evaluator/base.py

# Copyright 2022 The HuggingFace Datasets Authors and the TensorFlow Datasets 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.

from abc import ABC, abstractmethod
from numbers import Number
from typing import Any, Callable, Dict, List, Optional, Union

# Lint as: python3
from datasets import Dataset, load_dataset

from evaluate.evaluator.utils import choose_split


try:
    from scipy.stats import bootstrap

    SCIPY_AVAILABLE = True
except ImportError:
    SCIPY_AVAILABLE = False

try:
    import transformers
    from transformers import Pipeline, pipeline

    TRANSFORMERS_AVAILABLE = True
except ImportError:
    TRANSFORMERS_AVAILABLE = False

from time import perf_counter

from typing_extensions import Literal

from ..loading import load
from ..module import EvaluationModule
from ..utils.logging import get_logger
from .utils import DatasetColumn


logger = get_logger(__name__)


EVALUTOR_COMPUTE_START_DOCSTRING = r"""
    Compute the metric for a given pipeline and dataset combination.
    Args:
        model_or_pipeline (`str` or `Pipeline` or `Callable` or `PreTrainedModel` or `TFPreTrainedModel`, defaults to `None`):
            If the argument in not specified, we initialize the default pipeline for the task (in this case
            `text-classification` or its alias - `sentiment-analysis`). If the argument is of the type `str` or
            is a model instance, we use it to initialize a new `Pipeline` with the given model. Otherwise we assume the
            argument specifies a pre-initialized pipeline.
        data (`str` or `Dataset`, defaults to `None`):
            Specifies the dataset we will run evaluation on. If it is of type `str`, we treat it as the dataset
            name, and load it. Otherwise we assume it represents a pre-loaded dataset.
        subset (`str`, defaults to `None`):
            Defines which dataset subset to load. If `None` is passed the default subset is loaded.
        split (`str`, defaults to `None`):
            Defines which dataset split to load. If `None` is passed, infers based on the `choose_split` function.
        metric (`str` or `EvaluationModule`, defaults to `None`):
            Specifies the metric we use in evaluator. If it is of type `str`, we treat it as the metric name, and
            load it. Otherwise we assume it represents a pre-loaded metric.
        tokenizer (`str` or `PreTrainedTokenizer`, *optional*, defaults to `None`):
            Argument can be used to overwrite a default tokenizer if `model_or_pipeline` represents a model for
            which we build a pipeline. If `model_or_pipeline` is `None` or a pre-initialized pipeline, we ignore
            this argument.
        strategy (`Literal["simple", "bootstrap"]`, defaults to "simple"):
            specifies the evaluation strategy. Possible values are:
            - `"simple"` - we evaluate the metric and return the scores.
            - `"bootstrap"` - on top of computing the metric scores, we calculate the confidence interval for each
            of the returned metric keys, using `scipy`'s `bootstrap` method
            https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.bootstrap.html.
        confidence_level (`float`, defaults to `0.95`):
            The `confidence_level` value passed to `bootstrap` if `"bootstrap"` strategy is chosen.
        n_resamples (`int`, defaults to `9999`):
            The `n_resamples` value passed to `bootstrap` if `"bootstrap"` strategy is chosen.
        device (`int`, defaults to `None`):
            Device ordinal for CPU/GPU support of the pipeline. Setting this to -1 will leverage CPU, a positive
            integer will run the model on the associated CUDA device ID. If `None` is provided it will be inferred and
            CUDA:0 used if available, CPU otherwise.
        random_state (`int`, *optional*, defaults to `None`):
            The `random_state` value passed to `bootstrap` if `"bootstrap"` strategy is chosen. Useful for
            debugging.
"""

EVALUATOR_COMPUTE_RETURN_DOCSTRING = r"""
    Return:
        A `Dict`. The keys represent metric keys calculated for the `metric` spefied in function arguments. For the
        `"simple"` strategy, the value is the metric score. For the `"bootstrap"` strategy, the value is a `Dict`
        containing the score, the confidence interval and the standard error calculated for each metric key.
"""


class Evaluator(ABC):
    """
    The [`Evaluator`] class is the class from which all evaluators inherit. Refer to this class for methods shared across
    different evaluators.
    Base class implementing evaluator operations.
    """

    PIPELINE_KWARGS = {}
    METRIC_KWARGS = {}

    def __init__(self, task: str, default_metric_name: str = None):
        if not TRANSFORMERS_AVAILABLE:
            raise ImportError(
                "If you want to use the `Evaluator` you need `transformers`. Run `pip install evaluate[evaluator]`."
            )
        if not SCIPY_AVAILABLE:
            raise ImportError(
                "If you want to use the `Evaluator` you need `scipy>=1.7.1`. Run `pip install evaluate[evaluator]`."
            )
        self.task = task
        self.default_metric_name = default_metric_name

    @staticmethod
    def _compute_confidence_interval(
        metric,
        metric_inputs,
        metric_keys: List[str],
        confidence_level: float = 0.95,
        n_resamples: int = 9999,
        random_state: Optional[int] = None,
    ) -> Dict[str, Any]:
        """
        A utility function enabling the confidence interval calculation for metrics computed
        by the evaluator based on `scipy`'s `bootstrap` method.
        """

        # bootstrap only works with functions that use args and no kwargs
        def build_args_metric(metric, key, **kwargs):
            def args_metric(*args):
                return metric.compute(**{k: v for k, v in zip(kwargs.keys(), args)})[key]

            return args_metric

        bootstrap_dict = {}
        for key in metric_keys:
            bs = bootstrap(
                data=list(metric_inputs.values()),
                statistic=build_args_metric(metric, key, **metric_inputs),
                paired=True,
                vectorized=False,
                confidence_level=confidence_level,
                n_resamples=n_resamples,
                random_state=random_state,
            )
            bootstrap_dict[key] = {
                "confidence_interval": (bs.confidence_interval.low, bs.confidence_interval.high),
                "standard_error": bs.standard_error,
            }
        return bootstrap_dict

    @staticmethod
    def _compute_time_perf(start_time: float, end_time: float, num_samples: int) -> Dict[str, Any]:
        """
        A utility function computing time performance metrics:
            - `total_time_in_seconds` - pipeline inference runtime for the evaluation data in seconds,
            - `samples_per_second` - pipeline throughput in the number of samples per second.
            - `latency_in_seconds` - pipeline inference runtime for the evaluation data in seconds per sample,

        """
        latency = end_time - start_time
        throughput = num_samples / latency
        latency_sample = 1.0 / throughput

        return {
            "total_time_in_seconds": latency,
            "samples_per_second": throughput,
            "latency_in_seconds": latency_sample,
        }

    @staticmethod
    def _infer_device() -> int:
        """Helper function to check if GPU or CPU is available for inference."""
        # try infer with torch first
        try:
            import torch

            if torch.cuda.is_available():
                device = 0  # first GPU
            else:
                device = -1  # CPU
        except ImportError:
            # if not available try TF
            try:
                import tensorflow as tf

                if len(tf.config.list_physical_devices("GPU")) > 0:
                    device = 0  # first GPU
                else:
                    device = -1  # CPU
            except ImportError:
                device = -1

        if device == -1:
            logger.info("No GPU found. The default device for pipeline inference is set to CPU.")
        else:
            logger.info("GPU found. The default device for pipeline inference is set to GPU (CUDA:0).")

        return device

    @abstractmethod
    def predictions_processor(self, *args, **kwargs):
        """
        A core method of the `Evaluator` class, which processes the pipeline outputs for compatibility with the metric.
        """
        raise NotImplementedError()

    def compute(
        self,
        model_or_pipeline: Union[
            str, "Pipeline", Callable, "PreTrainedModel", "TFPreTrainedModel"  # noqa: F821
        ] = None,
        data: Union[str, Dataset] = None,
        subset: Optional[str] = None,
        split: Optional[str] = None,
        metric: Union[str, EvaluationModule] = None,
        tokenizer: Optional[Union[str, "PreTrainedTokenizer"]] = None,  # noqa: F821
        feature_extractor: Optional[Union[str, "FeatureExtractionMixin"]] = None,  # noqa: F821
        strategy: Literal["simple", "bootstrap"] = "simple",
        confidence_level: float = 0.95,
        n_resamples: int = 9999,
        device: int = None,
        random_state: Optional[int] = None,
        input_column: str = "text",
        label_column: str = "label",
        label_mapping: Optional[Dict[str, Number]] = None,
    ) -> Dict[str, float]:

        result = {}

        self.check_for_mismatch_in_device_setup(device, model_or_pipeline)

        # Prepare inputs
        data = self.load_data(data=data, subset=subset, split=split)
        metric_inputs, pipe_inputs = self.prepare_data(data=data, input_column=input_column, label_column=label_column)
        pipe = self.prepare_pipeline(
            model_or_pipeline=model_or_pipeline,
            tokenizer=tokenizer,
            feature_extractor=feature_extractor,
            device=device,
        )
        metric = self.prepare_metric(metric)

        # Compute predictions
        predictions, perf_results = self.call_pipeline(pipe, pipe_inputs)
        predictions = self.predictions_processor(predictions, label_mapping)

        metric_inputs.update(predictions)

        # Compute metrics from references and predictions
        metric_results = self.compute_metric(
            metric=metric,
            metric_inputs=metric_inputs,
            strategy=strategy,
            confidence_level=confidence_level,
            n_resamples=n_resamples,
            random_state=random_state,
        )

        # TODO: To clarify why `wer` and `cer` return float
        # even though metric.compute contract says that it
        # returns Optional[dict].
        if type(metric_results) is float:
            metric_results = {metric.name: metric_results}

        result.update(metric_results)
        result.update(perf_results)

        return result

    @staticmethod
    def check_for_mismatch_in_device_setup(device, model_or_pipeline):
        if device is not None and device != -1 and isinstance(model_or_pipeline, Pipeline):
            if model_or_pipeline.device.type == "cpu":
                raise ValueError(
                    "The value of the `device` kwarg passed to `compute` suggests that this pipe should be run on an "
                    "accelerator, but the pipe was instantiated on CPU. Pass `device` to the pipeline during "
                    "initialization to use an accelerator, or pass `device=None` to `compute`. "
                )
            elif device != model_or_pipeline.device.index:
                raise ValueError(
                    f"This pipeline was instantiated on device {model_or_pipeline.device.index} but device={device} was passed to `compute`."
                )

    def check_required_columns(self, data: Union[str, Dataset], columns_names: Dict[str, str]):
        """
        Ensure the columns required for the evaluation are present in the dataset.

        Args:
            data (`str` or [`Dataset`]):
                Specifies the dataset we will run evaluation on.
            columns_names (`List[str]`):
                List of column names to check in the dataset. The keys are the arguments to the [`evaluate.EvaluationModule.compute`] method,
                while the values are the column names to check.

        Example:

        ```py
        >>> from datasets import load_dataset
        >>> from evaluate import evaluator
        >>> data = load_dataset("rotten_tomatoes', split="train")
        >>> evaluator.check_required_columns(data, {"input_column": "text", "label_column": "label"})
        ```
        """
        for input_name, column_name in columns_names.items():
            if column_name not in data.column_names:
                raise ValueError(
                    f"Invalid `{input_name}` {column_name} specified. The dataset contains the following columns: {data.column_names}."
                )

    @staticmethod
    def get_dataset_split(data, subset=None, split=None):
        """
        Infers which split to use if `None` is given.

        Args:
             data (`str`):
                Name of dataset.
             subset (`str`):
                Name of config for datasets with multiple configurations (e.g. 'glue/cola').
             split (`str`, defaults to `None`):
                Split to use.
        Returns:
            `split`: `str` containing which split to use

        Example:

        ```py
        >>> from evaluate import evaluator
        >>> evaluator("text-classification").get_dataset_split(data="rotten_tomatoes")
        WARNING:evaluate.evaluator.base:Dataset split not defined! Automatically evaluating with split: TEST
        'test'
        ```
        """
        if split is None:
            split = choose_split(data, subset)
            logger.warning(f"Dataset split not defined! Automatically evaluating with split: {split.upper()}")
        return split

    def load_data(self, data: Union[str, Dataset], subset: str = None, split: str = None):
        """
        Load dataset with given subset and split.
        Args:
            data ([`Dataset`] or `str`, defaults to `None`):
                Specifies the dataset we will run evaluation on. If it is of
                type `str`, we treat it as the dataset name, and load it. Otherwise we assume it represents a pre-loaded dataset.
            subset (`str`, defaults to `None`):
                Specifies dataset subset to be passed to `name` in `load_dataset`. To be
                used with datasets with several configurations (e.g. glue/sst2).
            split (`str`, defaults to `None`):
                User-defined dataset split by name (e.g. train, validation, test). Supports slice-split (`test[:n]`).
                If not defined and data is a `str` type, will automatically select the best one via `choose_split()`.
        Returns:
            data ([`Dataset`]): Loaded dataset which will be used for evaluation.

        Example:

        ```py
        >>> from evaluate import evaluator
        >>> evaluator("text-classification").load_data(data="rotten_tomatoes", split="train")
        Dataset({
            features: ['text', 'label'],
            num_rows: 8530
        })
        ```
        """
        if isinstance(data, str):
            split = self.get_dataset_split(data, subset, split)
            data = load_dataset(data, name=subset, split=split)
            return data
        elif isinstance(data, Dataset):
            if split is not None or subset is not None:
                logger.warning("`data` is a preloaded Dataset! Ignoring `subset` and `split`.")
            return data
        else:
            raise ValueError(
                "Please specify a valid `data` object - either a `str` with a name or a `Dataset` object."
            )

    def prepare_data(self, data: Dataset, input_column: str, label_column: str, *args, **kwargs):
        """
        Prepare data.

        Args:
            data ([`Dataset`]):
                Specifies the dataset we will run evaluation on.
            input_column (`str`, defaults to `"text"`):
                The name of the column containing the text feature in the dataset specified by `data`.
            second_input_column(`str`, *optional*):
                The name of the column containing the second text feature if there is one. Otherwise, set to `None`.
            label_column (`str`, defaults to `"label"`):
                The name of the column containing the labels in the dataset specified by `data`.
        Returns:
            `dict`:  metric inputs.
            `list`:  pipeline inputs.

        Example:

        ```py
        >>> from evaluate import evaluator
        >>> from datasets import load_dataset

        >>> ds = load_dataset("rotten_tomatoes", split="train")
        >>> evaluator("text-classification").prepare_data(ds, input_column="text", second_input_column=None, label_column="label")
        ```
        """

        self.check_required_columns(data, {"input_column": input_column, "label_column": label_column})

        return {"references": data[label_column]}, DatasetColumn(data, input_column)

    def prepare_pipeline(
        self,
        model_or_pipeline: Union[str, "Pipeline", Callable, "PreTrainedModel", "TFPreTrainedModel"],  # noqa: F821
        tokenizer: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"] = None,  # noqa: F821
        feature_extractor: Union["PreTrainedTokenizerBase", "FeatureExtractionMixin"] = None,  # noqa: F821
        device: int = None,
    ):
        """
        Prepare pipeline.

        Args:
            model_or_pipeline (`str` or [`~transformers.Pipeline`] or `Callable` or [`~transformers.PreTrainedModel`] or [`~transformers.TFPreTrainedModel`], defaults to `None`):
                If the argument in not specified, we initialize the default pipeline for the task. If the argument is of the type `str` or
                is a model instance, we use it to initialize a new [`~transformers.Pipeline`] with the given model. Otherwise we assume the
                argument specifies a pre-initialized pipeline.
            preprocessor ([`~transformers.PreTrainedTokenizerBase`] or [`~transformers.FeatureExtractionMixin`], *optional*, defaults to `None`):
                Argument can be used to overwrite a default preprocessor if `model_or_pipeline` represents a model for
                which we build a pipeline. If `model_or_pipeline` is `None` or a pre-initialized pipeline, we ignore
                this argument.
        Returns:
            The initialized pipeline.

        Example:

        ```py
        >>> from evaluate import evaluator
        >>> evaluator("text-classification").prepare_pipeline(model_or_pipeline="distilbert-base-uncased")
        ```
        """

        if device is None:
            device = self._infer_device()

        if (
            isinstance(model_or_pipeline, str)
            or isinstance(model_or_pipeline, transformers.PreTrainedModel)
            or isinstance(model_or_pipeline, transformers.TFPreTrainedModel)
        ):
            pipe = pipeline(
                self.task,
                model=model_or_pipeline,
                tokenizer=tokenizer,
                feature_extractor=feature_extractor,
                device=device,
            )
        else:
            if model_or_pipeline is None:
                pipe = pipeline(self.task, device=device)
            else:
                pipe = model_or_pipeline
            if tokenizer is not None and feature_extractor is not None:
                logger.warning("Ignoring the value of the preprocessor argument (`tokenizer` or `feature_extractor`).")
        if (pipe.task != self.task) and not (self.task == "translation" and pipe.task.startswith("translation")):
            raise ValueError(
                f"Incompatible `model_or_pipeline`. Please specify `model_or_pipeline` compatible with the `{self.task}` task."
            )
        return pipe

    def prepare_metric(self, metric: Union[str, EvaluationModule]):
        """
        Prepare metric.

        Args:
            metric (`str` or [`EvaluationModule`], defaults to `None`):
                Specifies the metric we use in evaluator. If it is of type `str`, we treat it as the metric name, and
                load it. Otherwise we assume it represents a pre-loaded metric.

        Returns:
            The loaded metric.

        Example:

        ```py
        >>> from evaluate import evaluator
        >>> evaluator("text-classification").prepare_metric("accuracy")
        ```
        """
        # Prepare metric.
        if metric is None:
            if self.default_metric_name is None:
                raise ValueError(
                    "`Evaluator` doesn't specify a default metric. Please specify a valid `metric` argument."
                )
            metric = load(self.default_metric_name)
        elif isinstance(metric, str):
            metric = load(metric)

        return metric

    def call_pipeline(self, pipe, *args, **kwargs):
        start_time = perf_counter()
        pipe_output = pipe(*args, **kwargs, **self.PIPELINE_KWARGS)
        end_time = perf_counter()
        return pipe_output, self._compute_time_perf(start_time, end_time, len(pipe_output))

    def compute_metric(
        self,
        metric: EvaluationModule,
        metric_inputs: Dict,
        strategy: Literal["simple", "bootstrap"] = "simple",
        confidence_level: float = 0.95,
        n_resamples: int = 9999,
        random_state: Optional[int] = None,
    ):
        """Compute and return metrics."""
        result = metric.compute(**metric_inputs, **self.METRIC_KWARGS)

        if strategy == "bootstrap":
            metric_keys = result.keys()
            bootstrap_dict = self._compute_confidence_interval(
                metric,
                metric_inputs,
                metric_keys,
                confidence_level,
                n_resamples,
                random_state,
            )
            for key in metric_keys:
                bootstrap_dict[key]["score"] = result[key]

            return bootstrap_dict

        return result