Tensorflow实现卷积神经网络模型

import tensorflow as tf
from  tensorflow.examples.tutorials.mnist import input_data

##神经网络参数初始化
def init_NN():
    mnist = input_data.read_data_sets('data/', one_hot=True)
    ##定义神经网络参数 输入->输出
    n_input = 784  ##输入层
    n_cnn_1 = 64  ##卷积层1  channel
    n_cnn_2 = 128  ##卷积层2 channel
    n_link_all = 1024  ##全链接层
    n_classes = 10  ##输出层

    stddev = 0.1  ##标准差

    x = tf.placeholder(tf.float32, [None, n_input])
    y = tf.placeholder(tf.float32, [None, n_classes])
    keep_ratio = tf.placeholder(tf.float32)

    ##初始化每层神经网络的权重
    weigths = {
        'wc1': tf.Variable(tf.random_normal([3, 3, 1, n_cnn_1], stddev=stddev)),  ##卷积窗口大小3*3
        'wc2': tf.Variable(tf.random_normal([3, 3, n_cnn_1, n_cnn_2], stddev=stddev)),
        ## (28+2-3)/1+1=28  两次池化28/2/2=7
        'wd1': tf.Variable(tf.random_normal([7 * 7 * n_cnn_2, n_link_all], stddev=stddev)),
        'wd2': tf.Variable(tf.random_normal([n_link_all, n_classes], stddev=stddev))
    }
    ##初始化每层神经网络的偏移量
    biases = {
        'bc1': tf.Variable(tf.random_normal([n_cnn_1], stddev=stddev)),
        'bc2': tf.Variable(tf.random_normal([n_cnn_2], stddev=stddev)),
        'bd1': tf.Variable(tf.random_normal([n_link_all], stddev=stddev)),
        'bd2': tf.Variable(tf.random_normal([n_classes], stddev=stddev))
    }
    ##init = {'mnist':mnist,'x': x, 'y': y, 'keep_ratio': keep_ratio, 'weigths': weigths, 'biases': biases}
    return mnist,x,y,keep_ratio,weigths,biases


##前向传播
def conv_basic(_x, _w, _b, _keep_ratio):
    ##reshape input
    _input_r = tf.reshape(_x, [-1, 28, 28, 1])
    ##CONV 1
    _conv1 = tf.nn.conv2d(_input_r, _w['wc1'], strides=[1, 1, 1, 1], padding='SAME')
    ##relu
    _conv1 = tf.nn.relu(tf.nn.bias_add(_conv1, _b['bc1']))
    ##pool1
    _pool1 = tf.nn.max_pool(_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    ##dropout
    _pool1_drop = tf.nn.dropout(_pool1, _keep_ratio)

    ##CONV 2
    _conv2 = tf.nn.conv2d(_pool1_drop, _w['wc2'], strides=[1, 1, 1, 1], padding='SAME')
    ##relu
    _conv2 = tf.nn.relu(tf.nn.bias_add(_conv2, _b['bc2']))
    ##pool2
    _pool2 = tf.nn.max_pool(_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    ##dropout
    _pool2_drop = tf.nn.dropout(_pool2, _keep_ratio)

    ##VECTORIZE
    _densel = tf.reshape(_pool2_drop, [-1, _w['wd1'].get_shape().as_list()[0]])
    ##Fully connected layer1
    _fc1 = tf.nn.relu(tf.add(tf.matmul(_densel, _w['wd1']), _b['bd1']))
    ##dropout
    _fc1_drop = tf.nn.dropout(_fc1, _keep_ratio)

    ##Fully connected layer2
    _out = tf.add(tf.matmul(_fc1_drop, _w['wd2']), _b['bd2'])

    ##output dic
    out = {'input_r': _input_r, 'conv1': _conv1, 'pool1': _pool1, 'pool1_drop': _pool1_drop,
           'conv2': _conv2, 'pool2': _pool2, 'pool2_drop': _pool2_drop, 'densel': _densel,
           'fc1': _fc1, 'fc1_drop': _fc1_drop, 'out': _out
           }
    return out


##反向传播
def back_propagation(_pred, _y):
    ##学习率
    learning_rate = 0.01
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(_pred, _y))
    optm = tf.train.AdamOptimizer(learning_rate).minimize(cost)
    _corr = tf.equal(tf.arg_max(_pred, 1), tf.arg_max(_y, 1))
    accu = tf.reduce_mean(tf.cast(_corr, tf.float32))
    return cost, optm, accu


##开始训练
def start_train(_cost, _optm, _accu, _mnist,_x,_y,_keep_ratio):
    training_epochs = 30
    batch_size = 100
    display_step = 1
    ##saver
    saver_step=1
    saver=tf.train.Saver(max_to_keep=3)

    do_train=1   ##训练 测试标志位

    sess = tf.Session()
    sess.run(tf.global_variables_initializer())

    if do_train:
        for epoch in range(1, training_epochs + 1):
            avg_cost = 0
            total_batch = int(_mnist.train.num_examples / batch_size)
            for i in range(total_batch):
                batch_xs, batch_ys = _mnist.train.next_batch(batch_size)
                sess.run(_optm, feed_dict={_x: batch_xs, _y: batch_ys, keep_ratio: 0.7})
                avg_cost += sess.run(_cost, feed_dict={_x: batch_xs,_y: batch_ys, _keep_ratio: 1.0})
            avg_cost = avg_cost / total_batch
            if epoch % display_step == 0:
                print("Epoch:%03d/%03d cos:%.9f" % (epoch, training_epochs, avg_cost))
                train_acc = sess.run(_accu, feed_dict={_x: batch_xs,_y: batch_ys, _keep_ratio: 1.0})
                print("train_acc:%.3f" % (train_acc))
            if epoch%saver_step==0:
                saver.save(sess,'data/nets/cnn_mnist_basic.ckpt-'+str(epoch))
    else:
        saver.restore(sess,'data/nets/cnn_mnist_basic.ckpt-'+str(training_epochs-2))
        test_acc=sess.run(accu,feed_dict={_x:mnist.test.images,_y:mnist.test.labels,_keep_ratio:1.0})
        print("test_acc:%.3f" % (test_acc))
    print('FINISH')
    

##28 98.4/29 98/30 98.3/ 第28代的训练模型的测试结果要好一些

if __name__ == '__main__':    

    mnist, x, y, keep_ratio, weigths, biases = init_NN()    

   ##前向传播公式 

   pred = conv_basic(x,weigths,biases,keep_ratio)['out']    

   ##反向传播公式    

   cost, optm, accu = back_propagation(pred,y)    

   ##训练数据    

   start_train(cost, optm, accu, mnist,x,y,keep_ratio)