Python 在tensorflow中从头创建评估指标

我正在将tensorflow估计器API与tensorboard一起使用，我正在尝试创建一个自定义的评估指标，它不在tf.compat.v1.metrics中，比如“准确性”。 我的评估模型是：

def _model_fn_eval(self, mode, features, labels, endpoints, logits,
                       use_logits):
        """This is the EVAL part of model_fn."""
        if mode != tf.estimator.ModeKeys.EVAL:
            return None
        if use_logits:
            eval_predictions = logits
        else:
            eval_predictions = endpoints['Predictions']
        variant_type = features['variant_type']
        eval_metrics = (
            eval_metric_fn, [labels, eval_predictions, variant_type])
        if not self.use_tpu:
            for name, value in eval_metrics[0](*eval_metrics[1]).items():
                tf.compat.v1.summary.scalar(tensor=value, name=name)
        return eval_metrics

此函数使用函数eval_metric_fn：

def eval_metric_fn(labels, predictions, variant_types):
        """Calculate eval metrics from Tensors, on CPU host.
    
        Args:
          labels: the ground-truth labels for the examples.
          predictions: the predicted labels for the examples.
          variant_types: variant types (int64 of EncodedVariantType.value) as a tensor
            of (batch_size,) or None. The types of these variants. If None, no type
            specific evals will be performed.
    
        Returns:
          A dictionary of string name to metric.
        """
        predicted_classes = tf.argmax(input=predictions, axis=1)
    
        metrics = {}
    
        metrics['accuracy'] = tf.compat.v1.metrics.accuracy(labels=labels,
                                                            predictions=predicted_classes)
        return metrics

我想要实现的度量记录了被估计器分类为XX类的样本的百分比

我所尝试的： 天真地，我尝试将这部分代码添加到

eval\u metric\u fn

：

# a function to measure the percentage of predictions with each label
    per_label_predictions = lambda label: tf.math.reduce_mean(
        tf.cast(
            (tf.equal(tf.cast(predicted_classes, tf.int64), label)),
            tf.float32))
    predictions_per_label = {f'percentage_{l}_predictions':
                                 per_label_predictions(l)
                             for l in range(NUM_CLASSES)}

将此字典附加到

度量

dict后，应在度量dict中添加一个（键，值），值为浮点值

  File "/home/yonatan/anaconda3/envs/yonatan_env/lib/python3.8/site-packages/tensorflow_estimator/python/estimator/model_fn.py", line 475, in _validate_eval_metric_ops
    raise TypeError(
TypeError: Values of eval_metric_ops must be (metric_value, update_op) tuples, given: Tensor("Mean_10:0", shape=(), dtype=float32) for key: percentage_0_predictions

Process finished with exit code 1

我现在明白了度量不仅仅是一个简单的浮点，还有我应该实现的

update\u ops

。然而，我似乎找不到这方面的好例子。TF的例子使用了一个预构建的评估指标（这里：），而类似的问题对我来说没有任何意义，因为它假设了我猜之前的一些水含量（这里：）

如果需要更多信息，请告诉我