将 TensorBoard 2 中的 2 个图与 TensorFlow 2 合并

Question

我想使用 Tensorflow 和 tensorboard V2 在同一个图上合并精度和召回率。我为以前的版本找到了许多示例，但在我的情况下它们都不起作用。

我创建了一个 Keras 回调来计算精度和召回率，然后我调用一个 tensorflow 摘要将它们记录在同一个记录器中。我可以在 Tensorboard 中将它们可视化，但在 2 个单独的图中。

Class ClassificationReport(Callback):
    def __init__(self, data_generator, steps, label_names, log_directory):
        """
        Instantiator
        :param data_generator: the data generator that produces the input data
        :param steps: int, batch size
        :param data_type, string, 'training', 'validation' or 'test', used a prefix in the logs
        :param log_directory: pathlib2 path to the TensorBoard log directory

        """

        self.data_generator = data_generator
        self.steps = steps
        self.data_type = data_type
        self.logger = tensorflow.summary.create_file_writer(str(log_directory / self.data_type))

        # names of the scalar to consider in the sklearn classification report
        self._scalar_names = ['precision', 'recall']

    def on_epoch_end(self, epoch, logs={}):
        """
        log the precision and recall

        :param epoch: int, number of epochs
        :param logs: the Keras dictionary where the metrics are stored
        """

        y_true = numpy.zeros(self.steps)
        y_predicted = numpy.zeros(self.steps)

       ...Here I fetch y_true and y_predicted with the data_generator

        # The current report is calculated by SciKit-Learn
        current_report = classification_report(y_true, y_predicted, output_dict=True)

        with self.logger.as_default():
            for scalar_name in self._scalar_names:
                tensorflow.summary.scalar(
                    name="{} / macro average / {}".format(self.data_type, scalar_name),
                    data=current_report['macro avg'][scalar_name],
                    step=epoch)

        return super().on_epoch_end(epoch, logs)

据我了解 Tensorboard 2 的逻辑，似乎不可能在同一个图上绘制 2 个标量摘要......现阶段欢迎任何建议。

Answer 1

使用具有相同标量摘要名称的两个不同编写器。

import numpy as np
import tensorflow as tf

logger1 = tf.summary.create_file_writer('logs/scalar/precision')
logger2 = tf.summary.create_file_writer('logs/scalar/recall')

precision = np.random.uniform(size=10)
recall = np.random.uniform(size=10)

for i in range(10):
    with logger1.as_default():
        tf.summary.scalar(name='precision-recall', data=precision[i], step=i)
    with logger2.as_default():
        tf.summary.scalar(name='precision-recall', data=recall[i], step=i)

tensorboard --logdir 日志/标量

来自这个答案，适用于 tf2：https://stackoverflow.com/a/38718948/5597718