学习进度笔记

学习进度笔记14

卷集神经网络学习

import tensorflow as tf

 

slim = tf.contrib.slim

 

 

def lenet(images, num_classes=10, is_training=False,

          dropout_keep_prob=0.5,

          prediction_fn=slim.softmax,

          scope='LeNet'):

  end_points = {}

  with tf.variable_scope(scope, 'LeNet', [images]):

    net = end_points['conv1'] = slim.conv2d(images, 32, [5, 5], scope='conv1')

    net = end_points['pool1'] = slim.max_pool2d(net, [2, 2], 2, scope='pool1')

    net = end_points['conv2'] = slim.conv2d(net, 64, [5, 5], scope='conv2')

    net = end_points['pool2'] = slim.max_pool2d(net, [2, 2], 2, scope='pool2')

    net = slim.flatten(net)

    end_points['Flatten'] = net

 

    net = end_points['fc3'] = slim.fully_connected(net, 1024, scope='fc3')

    if not num_classes:

      return net, end_points

    net = end_points['dropout3'] = slim.dropout(

        net, dropout_keep_prob, is_training=is_training, scope='dropout3')

    logits = end_points['Logits'] = slim.fully_connected(

        net, num_classes, activation_fn=None, scope='fc4')

 

  end_points['Predictions'] = prediction_fn(logits, scope='Predictions')

 

  return logits, end_points

lenet.default_image_size = 28

 

 

def lenet_arg_scope(weight_decay=0.0):

  """Defines the default lenet argument scope.

  Args:

    weight_decay: The weight decay to use for regularizing the model.

  Returns:

    An `arg_scope` to use for the inception v3 model.

  """

  with slim.arg_scope(

      [slim.conv2d, slim.fully_connected],

      weights_regularizer=slim.l2_regularizer(weight_decay),

      weights_initializer=tf.truncated_normal_initializer(stddev=0.1),

      activation_fn=tf.nn.relu) as sc:

    return sc