x = tf.placeholder(tf.float32, [None, 784]) 

x_image = tf.reshape(x, [-1,28,28,1])    
W_conv1 = weight_variable([5, 5, 1, 32]) # 第一个卷积层的32个卷积核  
b_conv1 = bias_variable([32])  
# 第一个卷积层：  
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool(h_conv1)  # 第一个池化层    

from PIL import Image
import numpy as np
#from mnist_try001 import W_conv1

img1 = (W_conv1.eval()) # 将张量转换为numpy数组
W_conv1_10 = img1[:,:,:,9] 

W_conv1_10 = np.asmatrix(W_conv1_10) # 将数组转换为矩阵格式
W_conv1_10_visual = Image.fromarray(W_conv1_10 * 255.0 / W_conv1_10.max()) # 像素值归一化，Image.fromarray方法的输入范围是[0~255]
W_conv1_10_visual.show()

import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data

x = tf.placeholder(tf.float32, [None, 784])
mnist = input_data.read_data_sets('/TensorflowCode/MNIST_data', one_hot=True)

h_conv2 = tf.nn.relu(tf.nn.conv2d(h_pool1, W_conv2, strides=[1, 1, 1, 1], padding='SAME') + b_conv2) #14*14*64
# 可视化第二层输出的图像
input_image = mnist.train.images[100] # 输入一幅指定图像，mnist.train.images[100]尺寸为[784,],即1维：[1,784]
conv2 = sess.run(h_conv2, feed_dict={x:input_image}) # [64, 14, 14 ,1] 若前面网络中加入了dropout,这里的feed_dict中不要忘记加上keep_prob: 0.5
conv2 = sess.run(tf.reshape(conv2 , [64, 1, 14, 14]))
conv2 = np.sum(conv2,axis = 0) # 对中间层图像各通道求和，作为输出图像
h_conv1 = np.asmatrix(h_conv1) # 将conv2数组转换成矩阵格式
h_conv1 = Image.fromarray(h_conv1 * 255.0 / h_conv1.max()) # 矩阵数值归一化
h_conv1.show() # 输出14*14的灰度图像

# 输出第一层的32个卷积核（5×5*32）
import matplotlib.pyplot as plt

input_image = mnist.train.images[100]
W_conv1 = sess.run(W_conv1, feed_dict={x:input_image})   
W_conv1 = sess.run(tf.reshape(conv1_16, [32, 1, 5, 5]))
fig1,ax1 = plt.subplots(nrows=1, ncols=32, figsize = (32,1))
for i in range(32):
  ax1[i].imshow( W_conv1[i][0])           
plt.title('W_conv1 32×5×5')
plt.show()