Source code for torcheval.metrics.classification.recall
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
# pyre-ignore-all-errors[16]: Undefined attribute of metric states.
from typing import Iterable, Optional, TypeVar
import torch
from torcheval.metrics.functional.classification.recall import (
_binary_recall_compute,
_binary_recall_update,
_recall_compute,
_recall_param_check,
_recall_update,
)
from torcheval.metrics.metric import Metric
TRecall = TypeVar("TRecall")
TBinaryRecall = TypeVar("TBinaryRecall")
[docs]class BinaryRecall(Metric[torch.Tensor]):
"""
Compute the recall score for binary classification tasks, which is calculated as the ratio of the true positives and the sum of
true positives and false negatives.
Its functional version is :func:`torcheval.metrics.functional.binary_recall`.
We cast NaNs to 0 when classes have zero instances in the ground-truth labels
(when TP + FN = 0).
See also :class:`MulticlassRecall <MulticlassRecall>`
Args:
threshold (float, default 0.5): Threshold for converting input into predicted labels for each sample.
``torch.where(input < threshold, 0, 1)`` will be applied to the ``input``.
Examples::
>>> import torch
>>> from torcheval.metrics.classification import BinaryRecall
>>> metric = BinaryRecall()
>>> input = torch.tensor([0, 0, 1, 1])
>>> target = torch.tensor([0, 1, 1, 1])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.6667) # 2 / 3
>>> metric = BinaryRecall()
>>> input = torch.tensor([0, 0.2, 0.4, 0.7])
>>> target = torch.tensor([1, 0, 1, 1])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.3333) # 1 / 3
>>> metric = BinaryRecall(threshold=0.4)
>>> input = torch.tensor([0, 0.2, 0.4, 0.7])
>>> target = torch.tensor([1, 0, 1, 1])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.5000) # 1 / 2
"""
[docs] def __init__(
self: TBinaryRecall,
*,
threshold: float = 0.5,
device: Optional[torch.device] = None,
) -> None:
super().__init__(device=device)
self.threshold = threshold
self._add_state("num_tp", torch.tensor(0.0, device=self.device))
self._add_state("num_true_labels", torch.tensor(0.0, device=self.device))
@torch.inference_mode()
# pyre-ignore[14]: inconsistent override on *_:Any, **__:Any
def update(
self: TBinaryRecall, input: torch.Tensor, target: torch.Tensor
) -> TBinaryRecall:
"""
Update states with the ground truth labels and predictions.
Args:
input (Tensor): Tensor of the predicted labels/logits/probabilities, with shape of (n_sample, ).
``torch.where(input ‹ threshold, 0, 1)`` will be used to convert input into predicted labels
target (Tensor): Tensor of ground truth labels with shape of (n_sample, ).
"""
input = input.to(self.device)
target = target.to(self.device)
num_tp, num_true_labels = _binary_recall_update(input, target, self.threshold)
self.num_tp += num_tp
self.num_true_labels += num_true_labels
return self
@torch.inference_mode()
def compute(self: TBinaryRecall) -> torch.Tensor:
"""
Return the recall score.
NaN is returned if no calls to ``update()`` are made before ``compute()`` is called.
"""
return _binary_recall_compute(self.num_tp, self.num_true_labels)
@torch.inference_mode()
def merge_state(
self: TBinaryRecall, metrics: Iterable[TBinaryRecall]
) -> TBinaryRecall:
for metric in metrics:
self.num_tp += metric.num_tp.to(self.device)
self.num_true_labels += metric.num_true_labels.to(self.device)
return self
[docs]class MulticlassRecall(Metric[torch.Tensor]):
"""
Compute the recall score, the ratio of the true positives and the sum of
true positives and false negatives.
Its functional version is :func:`torcheval.metrics.functional.multiclass_recall`.
We cast NaNs to 0 when classes have zero instances in the ground-truth labels
(when TP + FN = 0).
See also :class:`BinaryRecall <BinaryRecall>`
Args:
num_classes (int):
Number of classes.
average (str, Optional):
- ``'micro'`` [default]:
Calculate the metrics globally.
- ``'macro'``:
Calculate metrics for each class separately, and return their unweighted mean.
Classes with 0 true and predicted instances are ignored.
- ``'weighted'``"
Calculate metrics for each class separately, and return their weighted sum.
Weights are defined as the proportion of occurrences of each class in "target".
Classes with 0 true and predicted instances are ignored.
- ``None``:
Calculate the metric for each class separately, and return
the metric for every class.
Examples::
>>> import torch
>>> from torcheval.metrics.classification import MulticlassRecall
>>> metric = MulticlassRecall(num_classes=4)
>>> input = torch.tensor([0, 2, 1, 3])
>>> target = torch.tensor([0, 1, 2, 3])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.5000)
>>> metric = MulticlassRecall(average=None, num_classes=4)
>>> input = torch.tensor([0, 2, 1, 3])
>>> target = torch.tensor([0, 1, 2, 3])
>>> metric.update(input, target)
>>> metric.compute()
tensor([1., 0., 0., 1.])
>>> metric = MulticlassRecall(average="macro", num_classes=2)
>>> input = torch.tensor([0, 0, 1, 1, 1])
>>> target = torch.tensor([0, 0, 0, 0, 1])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.5000)
>>> metric = MulticlassRecall(num_classes=4)
>>> input = torch.tensor([[0.9, 0.1, 0, 0], [0.1, 0.2, 0.4, 0.3], [0, 1.0, 0, 0], [0, 0, 0.2, 0.8]])
>>> target = torch.tensor([0, 1, 2, 3])
>>> metric.update(input, target)
>>> metric.compute()
tensor(0.5000)
"""
[docs] def __init__(
self: TRecall,
*,
num_classes: Optional[int] = None,
average: Optional[str] = "micro",
device: Optional[torch.device] = None,
) -> None:
super().__init__(device=device)
_recall_param_check(num_classes, average)
self.num_classes = num_classes
self.average = average
if average == "micro":
self._add_state("num_tp", torch.tensor(0.0, device=self.device))
self._add_state("num_labels", torch.tensor(0.0, device=self.device))
self._add_state(
"num_predictions",
torch.tensor(0.0, device=self.device),
)
else:
assert isinstance(
num_classes, int
), f"`num_classes` must be an integer, but got {num_classes}."
self._add_state(
"num_tp",
torch.zeros(num_classes, device=self.device),
)
self._add_state(
"num_labels",
torch.zeros(num_classes, device=self.device),
)
self._add_state(
"num_predictions",
torch.zeros(num_classes, device=self.device),
)
@torch.inference_mode()
# pyre-ignore[14]: inconsistent override on *_:Any, **__:Any
def update(self: TRecall, input: torch.Tensor, target: torch.Tensor) -> TRecall:
"""
Update states with the ground truth labels and predictions.
Args:
input (Tensor): Tensor of label predictions.
It could be the predicted labels, with shape of (n_sample, ).
It could also be probabilities or logits with shape of (n_sample, n_class).
``torch.argmax`` will be used to convert input into predicted labels.
target (Tensor): Tensor of ground truth labels with shape of (n_sample, ).
"""
input = input.to(self.device)
target = target.to(self.device)
num_tp, num_labels, num_predictions = _recall_update(
input, target, self.num_classes, self.average
)
self.num_tp += num_tp
self.num_labels += num_labels
self.num_predictions += num_predictions
return self
@torch.inference_mode()
def compute(self: TRecall) -> torch.Tensor:
"""
Return the recall score.
NaN is returned if no calls to ``update()`` are made before ``compute()`` is called.
"""
return _recall_compute(
self.num_tp, self.num_labels, self.num_predictions, self.average
)
@torch.inference_mode()
def merge_state(self: TRecall, metrics: Iterable[TRecall]) -> TRecall:
for metric in metrics:
self.num_tp += metric.num_tp.to(self.device)
self.num_labels += metric.num_labels.to(self.device)
self.num_predictions += metric.num_predictions.to(self.device)
return self
