使用Tensorboard展示Tensorflow模型的图结构

1 将ckpt的meta文件转换成tensorboard需要的log文件

参考代码如下:

import tensorflow as tf


def write_ckpt_meta_to_tensorboard(ckpt_meta_file, log_dir):
    g = tf.Graph()

    with g.as_default() as g:
        tf.train.import_meta_graph(ckpt_meta_file)

    with tf.Session(graph=g) as sess:
        file_writer = tf.summary.FileWriter(logdir=log_dir, graph=g)
        file_writer.close()


if __name__ == "__main__":
    meta_file = "/tmp/model.ckpt-806.meta"
    log_dir = "/tmp/tensorboard_log_806"
    write_ckpt_meta_to_tensorboard(meta_file, log_dir)

运行上面的代码,会产生log文件tensorboard_log_806.

2.启动Tensorboard服务

接着运行下面的命令启动tensorboard服务,前提是安装了tensorboard库.运行下面的pip命令进行安装

# 安装tensorboard
pip install tensorboard


# 启动tensorboard服务
tensorboard --logdir=/tmp/tensorboard_log_806

3.使用浏览器访问Tensorboard的Web

打开浏览器输入: localhost:6006, 即可查看图模型