[人工智能]TensorFlow # 调试中间结果，打印神经网络中间层

TensorFlow直接print Tensor拿到的是结构不是真实值。

暂时没有想到好方法，如果读者有更好的方法也请留言给博主，共同学习，谢谢。

对于线性程序：

import tensorflow as tf
import tensorflow.contrib.layers as layers
import numpy as np
import os

os.environ["TF_CPP_MIN_LOG_LEVEL"]='3' # 去掉TF warnings
obs = np.array([-3.45687799e+01, -7.04237729e+02,  5.00000000e+02, -1.22464680e-13, 0.00000000e+00,  0.00000000e+00]) # 4N+2M
datain = tf.placeholder(tf.float32, shape=(1,6), name="in")
out1 = layers.fully_connected(datain, num_outputs=64, activation_fn=tf.nn.relu)
out2 = layers.fully_connected(out1, num_outputs=64, activation_fn=tf.nn.relu)
out3 = layers.fully_connected(out2, num_outputs=64, activation_fn=tf.nn.relu)
out4 = layers.fully_connected(out3, num_outputs=1, activation_fn=tf.nn.tanh)
 
with tf.Session() as sess:
    init_op = tf.global_variables_initializer()
    sess.run(init_op)
    prt0 = sess.run(datain, feed_dict={datain: [obs]})
    print("prt0:",prt0)
    prt1 = sess.run(out1, feed_dict={datain: [obs]})
    print("prt1:",prt1)
    prt2 = sess.run(out2, feed_dict={datain: [obs]})
    print("prt2:",prt2)
    prt3 = sess.run(out3, feed_dict={datain: [obs]})
    print("prt3:",prt3)
    prt4 = sess.run(out4, feed_dict={datain: [obs]})
    print("prt4:",prt4)

对于复杂程序，比如我们封装了业务代码，并不是像原本这样的线性结构：model作为函数变量又传进了一个函数…对update方法也有相关的封装处理，我们可以用类似上述的方法，新建一个config文件，用于跨文件共享变量

def model(input, num_outputs, scope, reuse=False, num_units=64):
    with tf.variable_scope(scope, reuse=reuse):
        out = input
        print(out)
        out = layers.fully_connected(out, num_outputs=num_units, activation_fn=tf.nn.relu)
        out = layers.fully_connected(out, num_outputs=num_units, activation_fn=tf.nn.relu)
        out = layers.fully_connected(out, num_outputs=num_outputs, activation_fn=tf.nn.relu)
        out = tf.nn.softmax(out)
        return out

在config.py中:

class Config:
    flag = False
    p_func = None # model
    p_input_ph = None # placeholder

flag 标记是否开始训练，后俩为某一tensorflow网络的输入和输出
通过在对应封装代码里：

        Config.p_input_ph = obs_ph
        Config.p_func = p

然后在主循环中：

if Config.p_func is not None:
	sess = U.get_session() 
    print("[debug]:p:",sess.run(Config.p_func,feed_dict={Config.p_input_ph: [obs]}))

def get_session():
    """Returns recently made Tensorflow session"""
    return tf.get_default_session()