tensorflow深度学习之注意（四）

import tensorflow as tf

这个方法就是深度学习的神经网络，代码的核心部分。

def inference(images, batch_size, n_classes):

    with tf.variable_scope('conv1') as scope:

        weights = tf.get_variable('weights',

                                  shape = [3,3,3, 16],

                                  dtype = tf.float32,

                                  initializer=tf.truncated_normal_initializer(stddev=0.1,dtype=tf.float32))

        biases = tf.get_variable('biases',

                                shape=[16],

                                dtype=tf.float32,

                                initializer=tf.constant_initializer(0.1))

        conv = tf.nn.conv2d(images, weights, strides=[1,1,1,1], padding='SAME')

        pre_activation = tf.nn.bias_add(conv, biases)

        conv1 = tf.nn.relu(pre_activation, name= scope.name)

    #pool1 and norm1 

    with tf.variable_scope('pooling1_lrn') as scope:

        pool1 = tf.nn.max_pool(conv1, ksize=[1,3,3,1],strides=[1,2,2,1],

                              padding='SAME', name='pooling1')

        norm1 = tf.nn.lrn(pool1, depth_radius=4, bias=1.0, alpha=0.001/9.0,

                          beta=0.75,name='norm1')

    #conv2

    with tf.variable_scope('conv2') as scope:

        weights = tf.get_variable('weights',

                                  shape=[3,3,16,16],

                                  dtype=tf.float32,

                                  initializer=tf.truncated_normal_initializer(stddev=0.1,dtype=tf.float32))

        biases = tf.get_variable('biases',

                                shape=[16],

                                dtype=tf.float32,

                                initializer=tf.constant_initializer(0.1))

        conv = tf.nn.conv2d(norm1, weights, strides=[1,1,1,1],padding='SAME')

        pre_activation = tf.nn.bias_add(conv, biases)

        conv2 = tf.nn.relu(pre_activation, name='conv2')

    #pool2 and norm2

    with tf.variable_scope('pooling2_lrn') as scope:

        norm2 = tf.nn.lrn(conv2, depth_radius=4, bias=1.0, alpha=0.001/9.0,

                          beta=0.75,name='norm2')

        pool2 = tf.nn.max_pool(norm2, ksize=[1,3,3,1], strides=[1,1,1,1],

                              padding='SAME',name='pooling2')

    #local3

    with tf.variable_scope('local3') as scope:

        reshape = tf.reshape(pool2, shape=[batch_size, -1])

        dim = reshape.get_shape()[1].value

        weights = tf.get_variable('weights',

                                  shape=[dim,128],

                                  dtype=tf.float32,

                                  initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

        biases = tf.get_variable('biases',

                                shape=[128],

                                dtype=tf.float32,

                                initializer=tf.constant_initializer(0.1))

        local3 = tf.nn.relu(tf.matmul(reshape, weights) + biases, name=scope.name)   

    #local4

    with tf.variable_scope('local4') as scope:

        weights = tf.get_variable('weights',

                                  shape=[128,128],

                                  dtype=tf.float32,

                                  initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

        biases = tf.get_variable('biases',

                                shape=[128],

                                dtype=tf.float32,

                                initializer=tf.constant_initializer(0.1))

        local4 = tf.nn.relu(tf.matmul(local3, weights) + biases, name='local4')

    # softmax

    with tf.variable_scope('softmax_linear') as scope:

        weights = tf.get_variable('softmax_linear',

                                  shape=[128, n_classes],

                                  dtype=tf.float32,

                                  initializer=tf.truncated_normal_initializer(stddev=0.005,dtype=tf.float32))

        biases = tf.get_variable('biases',

                                shape=[n_classes],

                                dtype=tf.float32,

                                initializer=tf.constant_initializer(0.1))

        softmax_linear = tf.add(tf.matmul(local4, weights), biases, name='softmax_linear')

    return softmax_linear

#%%

def losses(logits, labels):

    with tf.variable_scope('loss') as scope:

        cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits\

                        (logits=logits, labels=labels, name='xentropy_per_example')

        loss = tf.reduce_mean(cross_entropy, name='loss')

        tf.summary.scalar(scope.name+'/loss', loss)

    return loss

def trainning(loss, learning_rate):

    with tf.name_scope('optimizer'):

        optimizer = tf.train.AdamOptimizer(learning_rate= learning_rate)

        global_step = tf.Variable(0, name='global_step', trainable=False)

        train_op = optimizer.minimize(loss, global_step= global_step)

    return train_op

def evaluation(logits, labels):

  with tf.variable_scope('accuracy') as scope:

      correct = tf.nn.in_top_k(logits, labels, 1)

      correct = tf.cast(correct, tf.float16)

      accuracy = tf.reduce_mean(correct)

      tf.summary.scalar(scope.name+'/accuracy', accuracy)

  return accuracy