tensorflow 卷积cnn

卷积神经网络的结构图

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import input_data

mnist = input_data.read_data_sets('data/', one_hot=True)
trainimg   = mnist.train.images
trainlabel = mnist.train.labels
testimg    = mnist.test.images
testlabel  = mnist.test.labels
print ("MNIST ready")

n_input  = 784
n_output = 10
weights  = {
        'wc1': tf.Variable(tf.random_normal([3, 3, 1, 64], stddev=0.1)),#第一个3是h，第二个是w，第三个是deep，第四个是out特征图
        'wc2': tf.Variable(tf.random_normal([3, 3, 64, 128], stddev=0.1)),
        'wd1': tf.Variable(tf.random_normal([7*7*128, 1024], stddev=0.1)),#其中的7*7*128是怎么算出来了如下：

其中pooling层由2*2减半为1*1，covn层不变，所以的出来第一个输出是28*28（由上面的式子计算出来），第一个pooling后是14*14，第二个pooling后是7*7*128
        'wd2': tf.Variable(tf.random_normal([1024, n_output], stddev=0.1))
    }
biases   = {
        'bc1': tf.Variable(tf.random_normal([64], stddev=0.1)),
        'bc2': tf.Variable(tf.random_normal([128], stddev=0.1)),
        'bd1': tf.Variable(tf.random_normal([1024], stddev=0.1)),
        'bd2': tf.Variable(tf.random_normal([n_output], stddev=0.1))
    }

def conv_basic(_input, _w, _b, _keepratio):
        # INPUT
        _input_r = tf.reshape(_input, shape=[-1, 28, 28, 1])#第一个数是batch，第二是图像，第三是w，第四维是deep，如果确定了其他三维，剩下的一维计算机可以自己确定，-1就是让计算机自己确定
        # CONV LAYER 
        _conv1 = tf.nn.conv2d(_input_r, _w['wc1'], strides=[1, 1, 1, 1], padding='SAME')#strides在官方定义中是一个一维具有四个元素的张量，其规定前后必须为1，所以我们可以改的是中间两个数，中间两个数分别代表了水平滑动和垂直滑动步长值。第一个是batch，最后一个是in_channel。。padding是same就是边界补0。vald是不填充。
        #_mean, _var = tf.nn.moments(_conv1, [0, 1, 2])
        #_conv1 = tf.nn.batch_normalization(_conv1, _mean, _var, 0, 1, 0.0001)
        _conv1 = tf.nn.relu(tf.nn.bias_add(_conv1, _b['bc1']))
        _pool1 = tf.nn.max_pool(_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        _pool_dr1 = tf.nn.dropout(_pool1, _keepratio)#随机杀死一些神经节点
        # CONV LAYER 2
        _conv2 = tf.nn.conv2d(_pool_dr1, _w['wc2'], strides=[1, 1, 1, 1], padding='SAME')
        #_mean, _var = tf.nn.moments(_conv2, [0, 1, 2])
        #_conv2 = tf.nn.batch_normalization(_conv2, _mean, _var, 0, 1, 0.0001)
        _conv2 = tf.nn.relu(tf.nn.bias_add(_conv2, _b['bc2']))
        _pool2 = tf.nn.max_pool(_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        _pool_dr2 = tf.nn.dropout(_pool2, _keepratio)
        # VECTORIZE
        _dense1 = tf.reshape(_pool_dr2, [-1, _w['wd1'].get_shape().as_list()[0]])#全连接层，_w['wd1'].get_shape().as_list()[0]是对上的进行了先计算shape在转化为list
        # FULLY CONNECTED LAYER 1
        _fc1 = tf.nn.relu(tf.add(tf.matmul(_dense1, _w['wd1']), _b['bd1']))#激活函数relu
        _fc_dr1 = tf.nn.dropout(_fc1, _keepratio)
        # FULLY CONNECTED LAYER 2
        _out = tf.add(tf.matmul(_fc_dr1, _w['wd2']), _b['bd2'])
        # RETURN
        out = { 'input_r': _input_r, 'conv1': _conv1, 'pool1': _pool1, 'pool1_dr1': _pool_dr1,
            'conv2': _conv2, 'pool2': _pool2, 'pool_dr2': _pool_dr2, 'dense1': _dense1,
            'fc1': _fc1, 'fc_dr1': _fc_dr1, 'out': _out
        }
        return out
print ("CNN READY")

x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_output])
keepratio = tf.placeholder(tf.float32)

# FUNCTIONS

_pred = conv_basic(x, weights, biases, keepratio)['out']
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(_pred, y))
optm = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
_corr = tf.equal(tf.argmax(_pred,1), tf.argmax(y,1)) 
accr = tf.reduce_mean(tf.cast(_corr, tf.float32)) 
init = tf.global_variables_initializer()
    
# SAVER
save_step = 1
saver = tf.train.Saver(max_to_keep=3) 

print ("GRAPH READY")

do_train = 0
sess = tf.Session()
sess.run(init)

training_epochs = 15
batch_size      = 16
display_step    = 1
if do_train == 1:
    for epoch in range(training_epochs):
        avg_cost = 0.
        #total_batch = int(mnist.train.num_examples/batch_size)
        total_batch = 10
        # Loop over all batches
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            # Fit training using batch data
            sess.run(optm, feed_dict={x: batch_xs, y: batch_ys, keepratio:0.7})
            # Compute average loss
            avg_cost += sess.run(cost, feed_dict={x: batch_xs, y: batch_ys, keepratio:1.})/total_batch

        # Display logs per epoch step
        if epoch % display_step == 0: 
            print ("Epoch: %03d/%03d cost: %.9f" % (epoch, training_epochs, avg_cost))
            train_acc = sess.run(accr, feed_dict={x: batch_xs, y: batch_ys, keepratio:1.})
            print (" Training accuracy: %.3f" % (train_acc))
            #test_acc = sess.run(accr, feed_dict={x: testimg, y: testlabel, keepratio:1.})
            #print (" Test accuracy: %.3f" % (test_acc))
            
        # Save Net
        if epoch % save_step == 0:
            saver.save(sess, "save/nets/cnn_mnist_basic.ckpt-" + str(epoch))

    print ("OPTIMIZATION FINISHED")

if do_train == 0:
    epoch = training_epochs-1
    saver.restore(sess, "save/nets/cnn_mnist_basic.ckpt-" + str(epoch))
    
    test_acc = sess.run(accr, feed_dict={x: testimg, y: testlabel, keepratio:1.})
    print (" TEST ACCURACY: %.3f" % (test_acc))