实现GoogLenet中Inception结构，基于tensorflow

图片是经过inception结构后concat后的特征图拼接的结果
row方向上0到7是11conv后的结果，8到15是11conv + 33conv后的结果， 16到22是11conv+5*5conv后的结果，很显然右侧的特征抽象程度更高，左边的特征更加底层一些
colum方向是不同的输入图像

测试后发现，相比于lenet在mnist上的正确率并没有什么提升，（实际是低了一些）自己感觉可能是由于手写数字的任务太简单了，网络太复杂过拟合或者是根本就用不到那么多的特征，而且lenet正确率本身就很高了已经

import sys
sys.path.append(r'E:\anaconda\Lib\site-packages')
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import cv2

sess = tf.InteractiveSession()
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
W = tf.Variable(tf.zeros([784,10]))
b = tf.Variable(tf.zeros([10]))

#网络参数
training_step = 10000
batch_size = 100
learing_rate = 1e-4

def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)

def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')


#inference

conv1in = 1
conv1out = 24
conv2out = 48
x_image = tf.reshape(x, [-1,28,28,1])

W_conv1 = weight_variable([3, 3, conv1in, conv1out])
b_conv1 = bias_variable([conv1out])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)


W_conv2 = weight_variable([3, 3, conv1out, conv2out])
b_conv2 = bias_variable([conv2out])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

#inception
#1
conv3out = 12
W_conv3 = weight_variable([1, 1,conv2out , conv3out])
b_conv3 = bias_variable([conv3out])
h_conv3 = conv2d(h_pool2, W_conv3) + b_conv3
#2
conv4out = 12
W_conv4 = weight_variable([1, 1, conv2out, conv4out])
b_conv4 = bias_variable([conv4out])
h_conv4 = conv2d(h_pool2, W_conv4) + b_conv4
conv5out = 12
W_conv5 = weight_variable([3, 3, conv4out, conv5out])
b_conv5 = bias_variable([conv5out])
h_conv5 = conv2d(h_conv4, W_conv5) + b_conv5
#3
conv6out = 12
W_conv6 = weight_variable([1, 1, conv2out, conv6out])
b_conv6 = bias_variable([conv6out])
h_conv6 = conv2d(h_pool2, W_conv6) + b_conv6
conv7out = 6
W_conv7 = weight_variable([3, 3, conv6out, conv7out])
b_conv7 = bias_variable([conv7out])
h_conv7 = conv2d(h_conv6, W_conv7) + b_conv7

#concat
concat_out = tf.concat([h_conv3,h_conv5,h_conv7],3)   #7 *7 *21
concat_relu = tf.nn.relu(concat_out)
concat_pool = max_pool_2x2(concat_relu)               #4*4*21

concat_out_depth = conv3out + conv5out + conv7out
# FC

h_pool_flat = tf.reshape(concat_pool, [-1, 4*4*concat_out_depth])
fc1size = 150
W_fc1 = weight_variable([4 * 4 * concat_out_depth, fc1size])
b_fc1 = bias_variable([fc1size])
W_fc2 = weight_variable([fc1size, 10])
b_fc2 = bias_variable([10])

keep_prob = tf.placeholder(tf.float32)

h_fc1 = tf.nn.relu(tf.matmul(h_pool_flat, W_fc1) + b_fc1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

y_out = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_out))


train_step = tf.train.AdamOptimizer(learing_rate).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_out,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

saver = tf.train.Saver() 

sess.run(tf.global_variables_initializer())

tra_accuracy = []
test_acc = []

cross_entropy_lst = []

'''
画准确率曲线
'''
def draw_pic(k):   
    plt.plot(np.linspace(0, (k+1)*200, len(tra_accuracy)),tra_accuracy,'b-',label='train acc')
    plt.plot(np.linspace(0, k*200, len(test_acc)),test_acc,'k-.',label='text acc')
    plt.legend(loc='lower right')
    plt.show()
    
'''
保存特征图，将特征图的深度方向的图在row方向上展开
'''

def save_conv1_feature_map(k):
    feature_map = np.zeros((28*10,28*conv1out))
    for j in range(10):        
        aa = sess.run(h_conv1,feed_dict={x: batch[0], keep_prob: 1})[j,:,:,:]    #shape  (28, 28, 16)   
        for i in range(conv1out):
            feature_map[j*28:(j+1)*28 , i*28:(i+1)*28] = aa[:,:,i] * 200
        cv2.imwrite(r'C:\Users\Administrator\Desktop\tensorflow_mnist\conv1_feature_map\{}con1.jpg'.format(k),feature_map)

def save_conv2_feature_map(k):
    feature_map = np.zeros((14*10,14*conv2out))
    for j in range(10):        
        aa = sess.run(h_conv2,feed_dict={x: batch[0], keep_prob: 1})[j,:,:,:]    #shape  (28, 28, 16)   
        for i in range(conv2out):
            feature_map[j*14:(j+1)*14 , i*14:(i+1)*14] = aa[:,:,i] * 200
        cv2.imwrite(r'C:\Users\Administrator\Desktop\tensorflow_mnist\conv2_feature_map\{}conv2.jpg'.format(k),feature_map)


def save_pool1_feature_map(k):
    feature_map = np.zeros((14*10,14*conv1out))
    for j in range(10):        
        aa = sess.run(h_pool1,feed_dict={x: batch[0], keep_prob: 1})[j,:,:,:]    #shape  (28, 28, 16)   
        for i in range(conv1out):
            feature_map[j*14:(j+1)*14 , i*14:(i+1)*14] = aa[:,:,i] * 200
        cv2.imwrite(r'C:\Users\Administrator\Desktop\tensorflow_mnist\pool1_feature_map\{}pool1.jpg'.format(k),feature_map)

def save_concat_out_feature_map(k):
    feature_map = np.zeros((7*10,7*concat_out_depth))
    for j in range(10):        
        aa = sess.run(concat_out,feed_dict={x: batch[0], keep_prob: 1})[j,:,:,:]    #shape  (28, 28, 16)   
        for i in range(concat_out_depth):
            feature_map[j*7:(j+1)*7 , i*7:(i+1)*7] = aa[:,:,i] * 200
        cv2.imwrite(r'C:\Users\Administrator\Desktop\tensorflow_mnist\concat_out_feature_map\{}concat_out.jpg'.format(k),feature_map)
        
   
k = 0
j = 0
for i in range(training_step):
  batch = mnist.train.next_batch(batch_size)
  if i%200 == 0:
    k += 1
    train_accuracy = accuracy.eval(feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
    test_accuracy = accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})
    tra_accuracy.append(train_accuracy)
    test_acc.append(test_accuracy)
    loss = cross_entropy.eval(feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
    print("step %d, training accuracy %g ,loss %g,test accuracy %g"%(i, train_accuracy,loss, test_accuracy))
    '''
    每200步保存一次feature map
    '''
    if i%200 == 0:
        draw_pic(k)
        save_conv1_feature_map(i)
        save_conv2_feature_map(i)
#        save_pool1_feature_map(i)
        save_concat_out_feature_map(i)
        j += 1
        
  train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})  
  if i%200 == 0:     
    saver.save(sess, r'C:\Users\Administrator\Desktop\tensorflow_mnist\model\model{}.ckpt'.format(i))