# Copyright 2020-2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Metric base class."""
from __future__ import absolute_import
from abc import ABCMeta, abstractmethod
import functools
import numpy as np
from mindspore.common.tensor import Tensor
_eval_types = {'classification', 'multilabel'}
[文档]class Metric(metaclass=ABCMeta):
"""
Base class of metric, which is used to evaluate metrics.
The `clear`, `update`, and `eval` should be called when evaluating metric, and they should be overridden by
subclasse. `update` will accumulate intermediate results in the evaluation process, `eval` will evaluate the final
result, and `clear` will reinitialize the intermediate results.
Never use this class directly, but instantiate one of its subclasses instead, for examples,
:class:`mindspore.train.MAE`, :class:`mindspore.train.Recall` etc.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> import numpy as np
>>> import mindspore as ms
>>>
>>> class MyMAE(ms.train.Metric):
... def __init__(self):
... super(MyMAE, self).__init__()
... self.clear()
...
... def clear(self):
... self._abs_error_sum = 0
... self._samples_num = 0
...
... def update(self, *inputs):
... y_pred = inputs[0].asnumpy()
... y = inputs[1].asnumpy()
... abs_error_sum = np.abs(y - y_pred)
... self._abs_error_sum += abs_error_sum.sum()
... self._samples_num += y.shape[0]
...
... def eval(self):
... return self._abs_error_sum / self._samples_num
>>>
>>> x = ms.Tensor(np.array([[0.1, 0.2, 0.6, 0.9], [0.1, 0.2, 0.6, 0.9]]), ms.float32)
>>> y = ms.Tensor(np.array([[0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.1, 0.1]]), ms.float32)
>>> y2 = ms.Tensor(np.array([[0.1, 0.25, 0.7, 0.9], [0.1, 0.25, 0.7, 0.9]]), ms.float32)
>>> metric = MyMAE().set_indexes([0, 2])
>>> metric.clear()
>>> # indexes is [0, 2], using x as logits, y2 as label.
>>> metric.update(x, y, y2)
>>> accuracy = metric.eval()
>>> print(accuracy)
1.399999976158142
>>> print(metric.indexes)
[0, 2]
"""
def __init__(self):
self._indexes = None
def _convert_data(self, data):
"""
Convert data type to numpy array.
Args:
data (Object): Input data.
Returns:
Ndarray, data with `np.ndarray` type.
"""
if isinstance(data, Tensor):
data = data.asnumpy()
elif isinstance(data, list):
data = np.array(data)
elif isinstance(data, np.ndarray):
pass
else:
raise TypeError(f"For 'Metric' and its derived classes, the input data type must be tensor, list or "
f"numpy.ndarray, but got {type(data)}.")
return data
@property
def indexes(self):
"""Get the current indexes value. The default value is None and can be changed by `set_indexes`.
"""
return getattr(self, '_indexes', None)
[文档] def set_indexes(self, indexes):
"""
This interface is to rearrange the inputs of `update`.
Given (label0, label1, logits), set the `indexes` to [2, 1] then the (logits, label1) will be the actually
inputs of `update`.
Note:
When customize a metric, decorate the `update` function with the decorator
:func:`mindspore.train.rearrange_inputs` for the `indexes` to take effect.
Args:
indexes (List(int)): The order of logits and labels to be rearranged.
Outputs:
:class:`Metric`, its original Class instance.
Raises:
ValueError: If the type of input 'indexes' is not a list or its elements are not all int.
"""
if not isinstance(indexes, list) or not all(isinstance(i, int) for i in indexes):
raise ValueError("For 'set_indexes', the argument 'indexes' must be a list and all its elements must "
"be int, please check whether it is correct.")
self._indexes = indexes
return self
def __call__(self, *inputs):
"""
Evaluate input data once.
Args:
inputs (tuple): The first item is a predict array, the second item is a target array.
Returns:
Float, compute result.
"""
self.clear()
self.update(*inputs)
return self.eval()
[文档] @abstractmethod
def clear(self):
"""
An interface describes the behavior of clearing the internal evaluation result.
Note:
All subclasses must override this interface.
Tutorial Examples:
- `Evaluation Metrics - Customized Metrics
<https://mindspore.cn/tutorials/en/r2.3.1/advanced/model/metric.html#customized-metrics>`_
"""
raise NotImplementedError('Must define clear function to use this base class')
[文档] @abstractmethod
def eval(self):
"""
An interface describes the behavior of computing the evaluation result.
Note:
All subclasses must override this interface.
Tutorial Examples:
- `Evaluation Metrics - Customized Metrics
<https://mindspore.cn/tutorials/en/r2.3.1/advanced/model/metric.html#customized-metrics>`_
"""
raise NotImplementedError('Must define eval function to use this base class')
[文档] @abstractmethod
def update(self, *inputs):
"""
An interface describes the behavior of updating the internal evaluation result.
Note:
All subclasses must override this interface.
Args:
inputs: A variable-length input argument list, usually are the logits and the corresponding labels.
Tutorial Examples:
- `Evaluation Metrics - Customized Metrics
<https://mindspore.cn/tutorials/en/r2.3.1/advanced/model/metric.html#customized-metrics>`_
"""
raise NotImplementedError('Must define update function to use this base class')
class EvaluationBase(Metric):
"""
Base class of evaluation.
Note:
Please refer to the definition of class `Accuracy`.
Args:
eval_type (str): Type of evaluation must be in {'classification', 'multilabel'}.
Raises:
TypeError: If the input type is not classification or multilabel.
"""
def __init__(self, eval_type):
super(EvaluationBase, self).__init__()
if eval_type not in _eval_types:
raise TypeError("The argument 'eval_type' must be in {}, but got {}".format(_eval_types, eval_type))
self._type = eval_type
def _check_shape(self, y_pred, y):
"""
Checks the shapes of y_pred and y.
Args:
y_pred (Tensor): Predict array.
y (Tensor): Target array.
"""
if self._type == 'classification':
if y_pred.ndim != y.ndim + 1:
raise ValueError("In classification case, the dimension of y_pred (predicted value) should equal to "
"the dimension of y (true value) add 1, but got y_pred dimension: {} and y "
"dimension: {}.".format(y_pred.ndim, y.ndim))
if y.shape != (y_pred.shape[0],) + y_pred.shape[2:]:
raise ValueError("In classification case, y_pred (predicted value) shape and y (true value) shape "
"can not match, y shape should be equal to y_pred shape that the value at index 1 "
"is deleted. Such as y_pred shape (1, 2, 3), then y shape should be (1, 3). "
"But got y_pred shape {} and y shape {}".format(y_pred.shape, y.shape))
else:
if y_pred.ndim != y.ndim:
raise ValueError("In {} case, the dimension of y_pred (predicted value) should equal to the dimension"
" of y (true value), but got y_pred dimension: {} and y dimension: {}."
.format(self._type, y_pred.ndim, y.ndim))
if y_pred.shape != y.shape:
raise ValueError("In {} case, the shape of y_pred (predicted value) should equal to the shape of y "
"(true value), but got y_pred shape: {} and y shape: {}."
.format(self._type, y_pred.shape, y.shape))
def _check_value(self, y_pred, y):
"""
Checks the values of y_pred and y.
Args:
y_pred (Tensor): Predict array.
y (Tensor): Target array.
"""
if self._type != 'classification' and not (np.equal(y_pred ** 2, y_pred).all() and np.equal(y ** 2, y).all()):
raise ValueError("In multilabel case, all elements in y_pred (predicted value) and y (true value) should "
"be 0 or 1.Please check whether your inputs y_pred and y are correct.")
def clear(self):
"""
A interface describes the behavior of clearing the internal evaluation result.
Note:
All subclasses must override this interface.
"""
raise NotImplementedError
def update(self, *inputs):
"""
A interface describes the behavior of updating the internal evaluation result.
Note:
All subclasses must override this interface.
Args:
inputs: The first item is a predicted array and the second item is a target array.
"""
raise NotImplementedError
def eval(self):
"""
A interface describes the behavior of computing the evaluation result.
Note:
All subclasses must override this interface.
"""
raise NotImplementedError
def _check_onehot_data(data):
"""
Whether input data is one-hot encoding.
Args:
data (numpy.array): Input data.
Returns:
bool, return true, if input data is one-hot encoding.
"""
if data.ndim > 1 and np.equal(data ** 2, data).all():
shp = (data.shape[0],) + data.shape[2:]
if np.equal(np.ones(shp), data.sum(axis=1)).all():
return True
return False
def _binary_clf_curve(preds, target, sample_weights=None, pos_label=1):
"""Calculate True Positives and False Positives per binary classification threshold."""
if sample_weights is not None and not isinstance(sample_weights, np.ndarray):
sample_weights = np.array(sample_weights)
if preds.ndim > target.ndim:
preds = preds[:, 0]
desc_score_indices = np.argsort(-preds)
preds = preds[desc_score_indices]
target = target[desc_score_indices]
if sample_weights is not None:
weight = sample_weights[desc_score_indices]
else:
weight = 1.
distinct_value_indices = np.where(preds[1:] - preds[:-1])[0]
threshold_idxs = np.pad(distinct_value_indices, (0, 1), constant_values=target.shape[0] - 1)
target = np.array(target == pos_label).astype(np.int64)
tps = np.cumsum(target * weight, axis=0)[threshold_idxs]
if sample_weights is not None:
fps = np.cumsum((1 - target) * weight, axis=0)[threshold_idxs]
else:
fps = 1 + threshold_idxs - tps
return fps, tps, preds[threshold_idxs]