如何使用图形在Tensorflow中创建模型

Question

我试图理解以下模型是如何在tensorflow中建立的。我更习惯于看到使用Tensorflow.kera.Sequential（）制作的多层感知器。如果有人能解释模型是如何创建的，或者如何了解更多关于其体系结构的信息，比如model.summary（），我会非常感激。谢谢

资料来源：https://github.com/github/CodeSearchNet/blob/master/src/models/model.py

该类的完整定义可以在上面的链接中找到

def make_model(self, is_train: bool):
        with self.__sess.graph.as_default():
            random.seed(self.hyperparameters['seed'])
            np.random.seed(self.hyperparameters['seed'])
            tf.set_random_seed(self.hyperparameters['seed'])

            self._make_model(is_train=is_train)
            self._make_loss()
            if is_train:
                self._make_training_step()
                self.__summary_writer = tf.summary.FileWriter(self.__tensorboard_dir, self.__sess.graph)

def _make_model(self, is_train: bool) -> None:
        """
        Create the actual model.
        Note: This has to create self.ops['code_representations'] and self.ops['query_representations'],
        tensors of the same shape and rank 2.
        """
        self.__placeholders['dropout_keep_rate'] = tf.placeholder(tf.float32,
                                                                  shape=(),
                                                                  name='dropout_keep_rate')
        self.__placeholders['sample_loss_weights'] = \
            tf.placeholder_with_default(input=np.ones(shape=[self.hyperparameters['batch_size']],
                                                      dtype=np.float32),
                                        shape=[self.hyperparameters['batch_size']],
                                        name='sample_loss_weights')

        with tf.variable_scope("code_encoder"):
            language_encoders = []
            for (language, language_metadata) in sorted(self.__per_code_language_metadata.items(), key=lambda kv: kv[0]):
                with tf.variable_scope(language):
                    self.__code_encoders[language] = self.__code_encoder_type(label="code",
                                                                              hyperparameters=self.hyperparameters,
                                                                              metadata=language_metadata)
                    language_encoders.append(self.__code_encoders[language].make_model(is_train=is_train))
            self.ops['code_representations'] = tf.concat(language_encoders, axis=0)
        with tf.variable_scope("query_encoder"):
            self.__query_encoder = self.__query_encoder_type(label="query",
                                                             hyperparameters=self.hyperparameters,
                                                             metadata=self.__query_metadata)
            self.ops['query_representations'] = self.__query_encoder.make_model(is_train=is_train)

        code_representation_size = next(iter(self.__code_encoders.values())).output_representation_size
        query_representation_size = self.__query_encoder.output_representation_size
        assert code_representation_size == query_representation_size, \
            f'Representations produced for code ({code_representation_size}) and query ({query_representation_size}) cannot differ!'

Answer 1

如果你想得到模型架构，你可以简单地使用tensorboard。正如你在这行看到的

self.__summary_writer = tf.summary.FileWriter(self.__tensorboard_dir, self.__sess.graph)

它将会话图写入self.__tensorboard_dir位置的一个文件中。您所需要的只是启动tensorboard并通过给定的url访问它

要启动tensorboard，请打开终端并使用此命令

tensorboard  logdir="<file path (url of self.__tensorboard_dir)>"

这将启动服务器并显示指向tensorboard的URL。在tensorboard中，您有图形选项卡，该选项卡将显示整个体系结构