The evaluator classes for automatic evaluation.
The main entry point for using the evaluator:
( task: str = None ) → Evaluator
Parameters
str) —
The task defining which evaluator will be returned. Currently accepted tasks are:"image-classification": will return a ImageClassificationEvaluator."question-answering": will return a QuestionAnsweringEvaluator."text-classification" (alias "sentiment-analysis" available): will return a TextClassificationEvaluator."token-classification": will return a TokenClassificationEvaluator.Returns
An evaluator suitable for the task.
Utility factory method to build an Evaluator.
Evaluators encapsulate a task and a default metric name. They leverage pipeline functionalify from transformers
to simplify the evaluation of multiple combinations of models, datasets and metrics for a given task.
The base class for all evaluator classes:
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.
( metric: EvaluationModule metric_inputs: typing.Dict strategy: typing.Literal['simple', 'bootstrap'] = 'simple' confidence_level: float = 0.95 n_resamples: int = 9999 random_state: typing.Optional[int] = None )
Compute and return metrics.
A core method of the Evaluator class, which processes the pipeline outputs for compatibility with the metric.
(
data: typing.Union[str, datasets.arrow_dataset.Dataset]
input_column: str
label_column: str
)
→
dict
Parameters
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.
str, defaults to "text") —
the name of the column containing the text feature in the dataset specified by data.
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.
Prepare data.
( metric: typing.Union[str, evaluate.module.EvaluationModule] )
Prepare metric.
( model_or_pipeline: typing.Union[str, ForwardRef('Pipeline'), typing.Callable, ForwardRef('PreTrainedModel'), ForwardRef('TFPreTrainedModel')] tokenizer: typing.Union[ForwardRef('PreTrainedTokenizerBase'), ForwardRef('FeatureExtractionMixin')] = None feature_extractor: typing.Union[ForwardRef('PreTrainedTokenizerBase'), ForwardRef('FeatureExtractionMixin')] = None device: int = None )
Parameters
str or Pipeline or Callable or PreTrainedModel or TFPreTrainedModel, —
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 Pipeline with the given model. Otherwise we assume the
argument specifies a pre-initialized pipeline.
PreTrainedTokenizerBase or 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.
Prepare pipeline.
( task = 'image-classification' default_metric_name = None )
Image classification evaluator.
This image classification evaluator can currently be loaded from evaluator() using the default task name
image-classification.
Methods in this class assume a data format compatible with the ImageClassificationPipeline.
( input_column: str = 'image' *args **kwargs )
Parameters
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
image-classification. 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.
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.
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.
str or FeatureExtractionMixin, optional, defaults to None) —
Argument can be used to overwrite a default feature extractor 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.
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.float, defaults to 0.95) —
The confidence_level value passed to bootstrap if "bootstrap" strategy is chosen.
int, defaults to 9999) —
The n_resamples value passed to bootstrap if "bootstrap" strategy is chosen.
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. IfNone is provided it will be inferred and
CUDA:0 used if available, CPU otherwise.
int, optional, defaults to None) —
The random_state value passed to bootstrap if "bootstrap" strategy is chosen. Useful for
debugging.
str, defaults to "image") —
the name of the column containing the images as PIL ImageFile in the dataset specified by data.
str, defaults to "label") —
the name of the column containing the labels in the dataset specified by data.
Dict[str, Number], optional, defaults to None) —
We want to map class labels defined by the model in the pipeline to values consistent with those
defined in the label_column of the data dataset.
Compute the metric for a given pipeline and dataset combination.
Examples:
>>> from evaluate import evaluator
>>> from datasets import load_dataset
>>> task_evaluator = evaluator("image-classification")
>>> data = load_dataset("beans", split="test[:40]")
>>> results = task_evaluator.compute(
>>> model_or_pipeline="nateraw/vit-base-beans",
>>> data=data,
>>> label_column="labels",
>>> metric="accuracy",
>>> label_mapping={'angular_leaf_spot': 0, 'bean_rust': 1, 'healthy': 2},
>>> strategy="bootstrap"
>>> )( task = 'question-answering' default_metric_name = None )
Question answering evaluator. This evaluator handles extractive question answering, where the answer to the question is extracted from a context.
This question answering evaluator can currently be loaded from evaluator() using the default task name
question-answering.
Methods in this class assume a data format compatible with the
QuestionAnsweringPipeline.
( model_or_pipeline: typing.Union[str, ForwardRef('Pipeline'), typing.Callable, ForwardRef('PreTrainedModel'), ForwardRef('TFPreTrainedModel')] = None data: typing.Union[str, datasets.arrow_dataset.Dataset] = None metric: typing.Union[str, evaluate.module.EvaluationModule] = None tokenizer: typing.Union[str, ForwardRef('PreTrainedTokenizer'), NoneType] = None strategy: typing.Literal['simple', 'bootstrap'] = 'simple' confidence_level: float = 0.95 n_resamples: int = 9999 device: int = None random_state: typing.Optional[int] = None question_column: str = 'question' context_column: str = 'context' id_column: str = 'id' label_column: str = 'answers' squad_v2_format: typing.Optional[bool] = None )
Parameters
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
question-answering). 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.
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.
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.
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.
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.float, defaults to 0.95) —
The confidence_level value passed to bootstrap if "bootstrap" strategy is chosen.
int, defaults to 9999) —
The n_resamples value passed to bootstrap if "bootstrap" strategy is chosen.
int, defaults to None) —
Device ordinal for CPU/GPU support of the pipeline. Setting this to -1 will leverage CPU, a positive
will run the model on the associated CUDA device ID. IfNone is provided it will be inferred and
CUDA:0 used if available, CPU otherwise.
int, optional, defaults to None) —
The random_state value passed to bootstrap if "bootstrap" strategy is chosen. Useful for
debugging.
str, defaults to "question") —
the name of the column containing the question in the dataset specified by data.
str, defaults to "context") —
the name of the column containing the context in the dataset specified by data.
str, defaults to "id") —
the name of the column cointaing the identification field of the question and answer pair in the
dataset specified by data.
str, defaults to "answers") —
the name of the column containing the answers in the dataset specified by data.
bool, optional, defaults to None) —
whether the dataset follows the format of squad_v2 dataset where a question may have no answer in the context. If this parameter is not provided,
the format will be automatically inferred.
Compute the metric for a given pipeline and dataset combination.
Examples:
>>> from evaluate import evaluator
>>> from datasets import load_dataset
>>> task_evaluator = evaluator("question-answering")
>>> data = load_dataset("squad", split="validation[:2]")
>>> results = task_evaluator.compute(
>>> model_or_pipeline="sshleifer/tiny-distilbert-base-cased-distilled-squad",
>>> data=data,
>>> metric="squad",
>>> )Datasets where the answer may be missing in the context are supported, for example SQuAD v2 dataset. In this case, it is safer to pass squad_v2_format=True to
the compute() call.
>>> from evaluate import evaluator
>>> from datasets import load_dataset
>>> task_evaluator = evaluator("question-answering")
>>> data = load_dataset("squad_v2", split="validation[:2]")
>>> results = task_evaluator.compute(
>>> model_or_pipeline="mrm8488/bert-tiny-finetuned-squadv2",
>>> data=data,
>>> metric="squad_v2",
>>> squad_v2_format=True,
>>> )( task = 'text-classification' default_metric_name = None )
Text classification evaluator.
This text classification evaluator can currently be loaded from evaluator() using the default task name
text-classification or with a "sentiment-analysis" alias.
Methods in this class assume a data format compatible with the TextClassificationPipeline - a single textual
feature as input and a categorical label as output.
( *args **kwargs )
Parameters
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.
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.
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.
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.
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.float, defaults to 0.95) —
The confidence_level value passed to bootstrap if "bootstrap" strategy is chosen.
int, defaults to 9999) —
The n_resamples value passed to bootstrap if "bootstrap" strategy is chosen.
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. IfNone is provided it will be inferred and
CUDA:0 used if available, CPU otherwise.
int, optional, defaults to None) —
The random_state value passed to bootstrap if "bootstrap" strategy is chosen. Useful for
debugging.
str, defaults to "text") —
the name of the column containing the text feature in the dataset specified by data.
str, defaults to "label") —
the name of the column containing the labels in the dataset specified by data.
Dict[str, Number], optional, defaults to None) —
We want to map class labels defined by the model in the pipeline to values consistent with those
defined in the label_column of the data dataset.
Compute the metric for a given pipeline and dataset combination.
Examples:
>>> from evaluate import evaluator
>>> from datasets import load_dataset
>>> task_evaluator = evaluator("text-classification")
>>> data = load_dataset("imdb", split="test[:2]")
>>> results = task_evaluator.compute(
>>> model_or_pipeline="huggingface/prunebert-base-uncased-6-finepruned-w-distil-mnli",
>>> data=data,
>>> metric="accuracy",
>>> label_mapping={"LABEL_0": 0.0, "LABEL_1": 1.0},
>>> strategy="bootstrap",
>>> n_resamples=10,
>>> random_state=0
>>> )( task = 'token-classification' default_metric_name = None )
Token classification evaluator.
This token classification evaluator can currently be loaded from evaluator() using the default task name
token-classification.
Methods in this class assume a data format compatible with the TokenClassificationPipeline.
( model_or_pipeline: typing.Union[str, ForwardRef('Pipeline'), typing.Callable, ForwardRef('PreTrainedModel'), ForwardRef('TFPreTrainedModel')] = None data: typing.Union[str, datasets.arrow_dataset.Dataset] = None metric: typing.Union[str, ForwardRef('EvaluationModule')] = None tokenizer: typing.Union[str, ForwardRef('PreTrainedTokenizer'), NoneType] = None strategy: typing.Literal['simple', 'bootstrap'] = 'simple' confidence_level: float = 0.95 n_resamples: int = 9999 device: typing.Optional[int] = None random_state: typing.Optional[int] = None input_column: str = 'tokens' label_column: str = 'ner_tags' join_by: typing.Optional[str] = ' ' )
Parameters
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
token-classification). 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.
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.
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.
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.
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.float, defaults to 0.95) —
The confidence_level value passed to bootstrap if "bootstrap" strategy is chosen.
int, defaults to 9999) —
The n_resamples value passed to bootstrap if "bootstrap" strategy is chosen.
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. IfNone is provided it will be inferred and
CUDA:0 used if available, CPU otherwise.
int, optional, defaults to None) —
The random_state value passed to bootstrap if "bootstrap" strategy is chosen. Useful for
debugging.
str, defaults to "tokens") —
the name of the column containing the tokens feature in the dataset specified by data.
str, defaults to "label") —
the name of the column containing the labels in the dataset specified by data.
str, optional, defaults to " ") —
This evaluator supports dataset whose input column is a list of words. This parameter specifies how to join
words to generate a string input. This is especially useful for languages that do not separate words by a space.
Compute the metric for a given pipeline and dataset combination.
The dataset input and label columns are expected to be formatted as a list of words and a list of labels respectively, following conll2003 dataset. Datasets whose inputs are single strings, and labels are a list of offset are not supported.
Examples:
>>> from evaluate import evaluator
>>> from datasets import load_dataset
>>> task_evaluator = evaluator("token-classification")
>>> data = load_dataset("conll2003", split="validation[:2]")
>>> results = task_evaluator.compute(
>>> model_or_pipeline="elastic/distilbert-base-uncased-finetuned-conll03-english",
>>> data=data,
>>> metric="seqeval",
>>> )For example, the following dataset format is accepted by the evaluator:
dataset = Dataset.from_dict(
mapping={
"tokens": [["New", "York", "is", "a", "city", "and", "Felix", "a", "person", "."]],
"ner_tags": [[1, 2, 0, 0, 0, 0, 3, 0, 0, 0]],
},
features=Features({
"tokens": Sequence(feature=Value(dtype="string")),
"ner_tags": Sequence(feature=ClassLabel(names=["O", "B-LOC", "I-LOC", "B-PER", "I-PER"])),
}),
)For example, the following dataset format is not accepted by the evaluator:
dataset = Dataset.from_dict(
mapping={
"tokens": [["New York is a city and Felix a person."]],
"starts": [[0, 23]],
"ends": [[7, 27]],
"ner_tags": [["LOC", "PER"]],
},
features=Features({
"tokens": Value(dtype="string"),
"starts": Sequence(feature=Value(dtype="int32")),
"ends": Sequence(feature=Value(dtype="int32")),
"ner_tags": Sequence(feature=Value(dtype="string")),
}),
)