• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2020 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ============================================================================
15"""evaluation metric."""
16
17from mindspore.communication.management import GlobalComm
18from mindspore.ops import operations as P
19import mindspore.nn as nn
20import mindspore.common.dtype as mstype
21
22
23class ClassifyCorrectCell(nn.Cell):
24    r"""
25    Cell that returns correct count of the prediction in classification network.
26    This Cell accepts a network as arguments.
27    It returns orrect count of the prediction to calculate the metrics.
28
29    Args:
30        network (Cell): The network Cell.
31
32    Inputs:
33        - **data** (Tensor) - Tensor of shape :math:`(N, \ldots)`.
34        - **label** (Tensor) - Tensor of shape :math:`(N, \ldots)`.
35
36    Outputs:
37        Tuple, containing a scalar correct count of the prediction
38
39    Examples:
40        >>> # For a defined network Net without loss function
41        >>> net = Net()
42        >>> eval_net = nn.ClassifyCorrectCell(net)
43    """
44
45    def __init__(self, network):
46        super(ClassifyCorrectCell, self).__init__(auto_prefix=False)
47        self._network = network
48        self.argmax = P.Argmax()
49        self.equal = P.Equal()
50        self.cast = P.Cast()
51        self.reduce_sum = P.ReduceSum()
52        self.allreduce = P.AllReduce(P.ReduceOp.SUM, GlobalComm.WORLD_COMM_GROUP)
53
54    def construct(self, data, label):
55        outputs = self._network(data)
56        y_pred = self.argmax(outputs)
57        y_pred = self.cast(y_pred, mstype.int32)
58        y_correct = self.equal(y_pred, label)
59        y_correct = self.cast(y_correct, mstype.float32)
60        y_correct = self.reduce_sum(y_correct)
61        total_correct = self.allreduce(y_correct)
62        return (total_correct,)
63
64
65class DistAccuracy(nn.Metric):
66    r"""
67    Calculates the accuracy for classification data in distributed mode.
68    The accuracy class creates two local variables, correct number and total number that are used to compute the
69    frequency with which predictions matches labels. This frequency is ultimately returned as the accuracy: an
70    idempotent operation that simply divides correct number by total number.
71
72    .. math::
73
74        \text{accuracy} =\frac{\text{true_positive} + \text{true_negative}}
75
76        {\text{true_positive} + \text{true_negative} + \text{false_positive} + \text{false_negative}}
77
78    Args:
79        eval_type (str): Metric to calculate the accuracy over a dataset, for classification (single-label).
80
81    Examples:
82        >>> y_correct = Tensor(np.array([20]))
83        >>> metric = nn.DistAccuracy(batch_size=3, device_num=8)
84        >>> metric.clear()
85        >>> metric.update(y_correct)
86        >>> accuracy = metric.eval()
87    """
88
89    def __init__(self, batch_size, device_num):
90        super(DistAccuracy, self).__init__()
91        self.clear()
92        self.batch_size = batch_size
93        self.device_num = device_num
94
95    def clear(self):
96        """Clears the internal evaluation result."""
97        self._correct_num = 0
98        self._total_num = 0
99
100    def update(self, *inputs):
101        """
102        Updates the internal evaluation result :math:`y_{pred}` and :math:`y`.
103
104        Args:
105            inputs: Input `y_correct`. `y_correct` is a `scalar Tensor`.
106                `y_correct` is the right prediction count that gathered from all devices
107                it's a scalar in float type
108
109        Raises:
110            ValueError: If the number of the input is not 1.
111        """
112
113        if len(inputs) != 1:
114            raise ValueError('Distribute accuracy needs 1 input (y_correct), but got {}'.format(len(inputs)))
115        y_correct = self._convert_data(inputs[0])
116        self._correct_num += y_correct
117        self._total_num += self.batch_size * self.device_num
118
119    def eval(self):
120        """
121        Computes the accuracy.
122
123        Returns:
124            Float, the computed result.
125
126        Raises:
127            RuntimeError: If the sample size is 0.
128        """
129
130        if self._total_num == 0:
131            raise RuntimeError('Accuracy can not be calculated, because the number of samples is 0.')
132        return self._correct_num / self._total_num
133