使用残差网络(residual network)分类mnist image

require: tensorflow version >= 0.8.0

'''Created on Nov 4, 2016'''import numpy as npimport tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_dataprint 'may be download data...'mnist = input_data.read_data_sets("your_mnist_dir/", one_hot=True)print 'read data finished'is_training=Truedef tf_variable(shape, name=None):    return tf.Variable(tf.truncated_normal(shape, stddev=0.1), name=name)def dense_connect(x, shape):    w = tf_variable(shape)    b = tf.Variable(tf.zeros([shape[1]]))    return tf.matmul(x, w) + bdef batch_norm(inputs, is_training,is_conv_out=True,decay = 0.999):    scale = tf.Variable(tf.ones([inputs.get_shape()[-1]]))    beta = tf.Variable(tf.zeros([inputs.get_shape()[-1]]))    pop_mean = tf.Variable(tf.zeros([inputs.get_shape()[-1]]), trainable=False)    pop_var = tf.Variable(tf.ones([inputs.get_shape()[-1]]), trainable=False)    if is_training:        if is_conv_out:            batch_mean, batch_var = tf.nn.moments(inputs,[0,1,2])        else:            batch_mean, batch_var = tf.nn.moments(inputs,[0])           train_mean = tf.assign(pop_mean,                               pop_mean * decay + batch_mean * (1 - decay))        train_var = tf.assign(pop_var,                              pop_var * decay + batch_var * (1 - decay))        with tf.control_dependencies([train_mean, train_var]):            return tf.nn.batch_normalization(inputs,                batch_mean, batch_var, beta, scale, 0.001)    else:        return tf.nn.batch_normalization(inputs,            pop_mean, pop_var, beta, scale, 0.001)def conv2d_with_batch_norm(x, filter_shape, stride):    filter_ = tf_variable(filter_shape)    conv = tf.nn.conv2d(x, filter=filter_, strides=[1, stride, stride, 1], padding="SAME")    normed=batch_norm(conv, is_training)    return  tf.nn.relu(normed)def conv2d(x, filter_shape, stride):    out_channels = filter_shape[3]    conv = tf.nn.conv2d(x, filter=tf_variable(filter_shape), strides=[1, stride, stride, 1], padding="SAME")    bias = tf.Variable(tf.zeros([out_channels]), name="bias")    return tf.nn.relu(tf.nn.bias_add(conv,bias))def max_pool(x):    return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')def residual_block(x, out_channels, down_sample, projection=False):    in_channels = x.get_shape().as_list()[3]    if down_sample:        x = max_pool(x)    output = conv2d_with_batch_norm(x, [3, 3, in_channels, out_channels], 1)    output = conv2d_with_batch_norm(output, [3, 3, out_channels, out_channels], 1)    if in_channels != out_channels:        if projection:            # projection shortcut            input_ = conv2d(x, [1, 1, in_channels, out_channels], 2)        else:            # zero-padding            input_ = tf.pad(x, [[0,0], [0,0], [0,0], [0, out_channels - in_channels]])    else:        input_ = x    return output + input_def residual_group(name,x,num_block,out_channels):    assert num_block>=1,'num_block must greater than 1'    with tf.variable_scope('%s_head'%name):        output = residual_block(x, out_channels, True)    for i in xrange (num_block-1):        with tf.variable_scope('%s_%d' % (name,i+1)):            output = residual_block(output,out_channels, False)    return outputdef residual_net(inpt):    with tf.variable_scope('conv1'):        output = conv2d(inpt, [3, 3, 1, 16], 1)    output=residual_group('conv2', x=output,num_block=2,out_channels=16)    output=residual_group('conv3', x=output,num_block=2,out_channels=32)    #output=residual_group('conv4', x=output,num_block=2,out_channels=64)    with tf.variable_scope('fc'):        output=max_pool(output)        shape=output.get_shape().as_list()        i_shape=shape[1]*shape[2]*shape[3]        output=tf.reshape(output,[-1,i_shape])        return dense_connect(output, [i_shape, 10])def train_network(batch_size = 120,training_iters=800,learning_rate=0.001):    x = tf.placeholder("float", [None, 28, 28, 1])#[batch_size,width,height,channels]    y = tf.placeholder("float", [None, 10])#[batch_size,num_classes]    pred = residual_net(x)    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)    correct_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))    accuracytr = tf.reduce_mean(tf.cast(correct_pred, tf.float32))    accuracyte = tf.reduce_mean(tf.cast(correct_pred, tf.float32))    tf.scalar_summary('cost', cost)    tf.scalar_summary('train accuracy', accuracytr)        tf.scalar_summary('test accuracy', accuracyte)    merged = tf.merge_all_summaries()      init = tf.initialize_all_variables()    print 'start training...'    with tf.Session() as sess:        sess.run(init)        swriter = tf.train.SummaryWriter("your_summary_dir/", sess.graph)          step = 1        while step< training_iters:            batch_xs, batch_ys = mnist.train.next_batch(batch_size)            #print np.shape(batch_xs),np.shape(batch_ys)            batch_xs=np.reshape(batch_xs,[np.shape(batch_xs)[0],28,28,1])            sess.run(optimizer, feed_dict={x: batch_xs, y: batch_ys})            if step % 10 == 0:                summary,acc = sess.run([merged,accuracytr], feed_dict={x: batch_xs, y: batch_ys})                swriter.add_summary(summary,step)                  summary,loss = sess.run([merged,cost], feed_dict={x: batch_xs, y: batch_ys})                swriter.add_summary(summary,step)                  batch_test=mnist.test.images[:256]                summary,ta=sess.run([merged,accuracyte], feed_dict={x: np.reshape(batch_test,[np.shape(batch_test)[0],28,28,1]), y: mnist.test.labels[:256]})                swriter.add_summary(summary,step)                  print "%s,loss:%s, train accuracy:%s, test accuray:%s"%(step,"{:.6f}".format(loss),"{:.6f}".format(acc),"{:.6f}".format(ta))            step += 1    print "train finished"if __name__ == '__main__':    train_network()