自定义指标

    有时你会遇到特定任务的Loss计算方式在框架既有的Loss接口中不存在，或算法不符合自己的需求，那么期望能够自己来进行Loss的自定义。这里介绍如何进行Loss的自定义操作，首先来看下面的代码：

    2.     def __init__(self):
    3.        super(SoftmaxWithCrossEntropy, self).__init__()
    5.     def forward(self, input, label):
    6.        loss = F.softmax_with_cross_entropy(input,
    7.                                            label,
    8.                                            return_softmax=False,
    9.                                            axis=1)
    10.        return paddle.mean(loss)

    和Loss一样，你也可以来通过框架实现自定义的评估方法，具体的实现如下：

    1. from paddle.metric import Metric
    3. class Precision(Metric):
    4.     """
    5.     Precision (also called positive predictive value) is the fraction of
    6.     relevant instances among the retrieved instances. Refer to
    7.     https://en.wikipedia.org/wiki/Evaluation_of_binary_classifiers
    8.     Noted that this class manages the precision score only for binary
    9.     classification task.
    11.     ......
    12.     """
    14.     def __init__(self, name='precision', *args, **kwargs):
    15.         super(Precision, self).__init__(*args, **kwargs)
    16.         self.tp = 0  # true positive
    17.         self.fp = 0  # false positive
    18.         self._name = name
    20.     def update(self, preds, labels):
    21.         """
    22.         Update the states based on the current mini-batch prediction results.
    23.             preds (numpy.ndarray): The prediction result, usually the output
    24.                of two-class sigmoid function. It should be a vector (column
    26.            labels (numpy.ndarray): The ground truth (labels),
    27.                the shape should keep the same as preds.
    28.                The data type is 'int32' or 'int64'.
    29.         """
    30.         if isinstance(preds, paddle.Tensor):
    31.             preds = preds.numpy()
    32.         elif not _is_numpy_(preds):
    33.             raise ValueError("The 'preds' must be a numpy ndarray or Tensor.")
    34.         if isinstance(labels, paddle.Tensor):
    35.             labels = labels.numpy()
    36.         elif not _is_numpy_(labels):
    37.             raise ValueError("The 'labels' must be a numpy ndarray or Tensor.")
    39.         sample_num = labels.shape[0]
    40.         preds = np.floor(preds + 0.5).astype("int32")
    42.         for i in range(sample_num):
    43.             pred = preds[i]
    44.             label = labels[i]
    45.             if pred == 1:
    46.                 if pred == label:
    47.                     self.tp += 1
    48.                 else:
    49.                     self.fp += 1
    51.     def reset(self):
    52.         """
    53.         Resets all of the metric state.
    54.         """
    55.         self.tp = 0
    56.         self.fp = 0
    58.         """
    61.         Returns:
    62.            A scaler float: results of the calculated precision.
    63.         """
    64.         ap = self.tp + self.fp
    65.         return float(self.tp) / ap if ap != 0 else .0
    67.     def name(self):
    68.         """
    69.         Returns metric name
    70.         """
    71.         return self._name

    fit接口的callback参数支持传入一个`` Callback``类实例，用来在每轮训练和每个`` batch``训练前后进行调用，可以通过`` callback``收集到训练过程中的一些数据和参数，或者实现一些自定义操作。

    1. class ModelCheckpoint(Callback):
    2.     def __init__(self, save_freq=1, save_dir=None):
    3.         self.save_freq = save_freq
    4.         self.save_dir = save_dir
    6.     def on_epoch_begin(self, epoch=None, logs=None):
    7.         self.epoch = epoch
    9.     def _is_save(self):
    10.         return self.model and self.save_dir and ParallelEnv().local_rank == 0
    12.     def on_epoch_end(self, epoch, logs=None):
    13.         if self._is_save() and self.epoch % self.save_freq == 0:
    14.             path = '{}/{}'.format(self.save_dir, epoch)
    15.             print('save checkpoint at {}'.format(os.path.abspath(path)))
    16.             self.model.save(path)
    18.     def on_train_end(self, logs=None):
    19.         if self._is_save():
    20.             path = '{}/final'.format(self.save_dir)
    21.             print('save checkpoint at {}'.format(os.path.abspath(path)))