TensorFlow拾遗(一) 打印网络结构与变量

　　使用tensorflow搭建网络之后，如果可视化一下网络的结构与变量，会对网络结构有一个更直观的了解。

　　另外，这种方式也可以获得网络输出节点名称，便于pb文件的生成。

　　在许多源码中都会包含这一操作，只不过大多可能并没有打印出来

variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # type：list

import tensorflow as tf
import os
def txt_save(data, output_file):
    file = open(output_file, 'a')
    for i in data:
        s = str(i) + '\n'
        file.write(s)
    file.close()


def network_param(input_checkpoint, output_file=None):
    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=True)
    with tf.Session() as sess:
        saver.restore(sess, input_checkpoint)
        variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
        for i in variable:
            print(i)     # 打印
        txt_save(variables, output_file)  # 保存txt   二选一

if __name__ == '__main__':
    checkpoint_path = 'ckpt/model.ckpt'
    output_file = 'network_param.txt'
    if not os.path.exists(output_file):
        network_param(checkpoint_path, output_file)

　　获得的txt文件部分如下所示，详细的txt文件猛戳这里：

<tf.Variable 'Layer1/conv2d/kernel:0' shape=(7, 7, 3, 64) dtype=float32_ref>
<tf.Variable 'Layer1/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer1/batch_normalization/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/conv2d/kernel:0' shape=(1, 1, 64, 64) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/conv2d_1/kernel:0' shape=(3, 3, 64, 64) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization_1/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization_1/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/conv2d_2/kernel:0' shape=(1, 1, 64, 256) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization_2/gamma:0' shape=(256,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/batch_normalization_2/beta:0' shape=(256,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/Downsample/conv2d/kernel:0' shape=(1, 1, 64, 256) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/Downsample/batch_normalization/gamma:0' shape=(256,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_0/Downsample/batch_normalization/beta:0' shape=(256,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/conv2d/kernel:0' shape=(1, 1, 256, 64) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/batch_normalization/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/batch_normalization/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/conv2d_1/kernel:0' shape=(3, 3, 64, 64) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/batch_normalization_1/gamma:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/batch_normalization_1/beta:0' shape=(64,) dtype=float32_ref>
<tf.Variable 'Layer2/Block_1/conv2d_2/kernel:0' shape=(1, 1, 64, 256) dtype=float32_ref>

　　其实，在打印网络结构与变量获得的结果中，也可以获得输出节点的名称，如下所示：

Score/Head/conv2d_1/BiasAdd 
BBox/Head/conv2d_1/BiasAdd