TF-slim快速搭建DNN

参考：
1、https://github.com/tensorflow/models/blob/master/research/slim/nets/cifarnet.py
2、https://github.com/tensorflow/models/blob/master/research/slim

---

模型：slim搭建的DNN
数据：mnist
通过直接对TF-slim快速搭建cnn 修改，即可完成DNN的搭建

---

#!/usr/bin/env python3
# -*- coding: UTF-8 -*-
"""使用slim API快速搭建cnn，
使用的数据集 mnist
参考：
1、https://github.com/tensorflow/models/blob/master/research/slim/nets/cifarnet.py
2、http://blog.csdn.net/wc781708249/article/details/78414028
"""


from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf
import tensorflow.contrib.slim as slim
# slim = tf.contrib.slim
from tensorflow.examples.tutorials.mnist import input_data
import argparse
import sys


# 搭建dnn 多层感知神经层
def cifarnet(images,num_classes=10,is_training=False,
             dropout_keep_prob=0.5,
             prediction_fn=slim.softmax,
             scope='CifarNet'):
    """
    Creates a variant of the CifarNet model.
    :param images: 输入图像数据 形状[n,h*w*c]
    :param num_classes: 类别数
    :param is_training: 是否训练 模型训练设置为True，测试、推理设置为False
    :param dropout_keep_prob: droupout保持率
    :param prediction_fn: 输出层的激活函数
    :param scope: 节点名
    :return: 
        net：2D Tensor ,logits （pre-softmax激活）如果num_classes
            是非零整数，或者如果num_classes为0或None输入到逻辑层           
        end_points：从网络组件到相应的字典激活。
    """

    end_points = {}
    # 如果使用mnist  images的shape为[n,28*28*1]
    with tf.variable_scope(scope, 'CifarNet', [images]): # 其中[images]为传入的数据
        net = slim.fully_connected(images, 384, scope='fc1') # 全连接层，输出节点384 ；shape [n,384]
        end_points['fc1'] = net
        net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
                           scope='dropout3')  # droupout 层 ；shape [n,384]
        net = slim.fully_connected(net, 192, scope='fc2') # # 全连接层，输出节点192 ；shape [n,192]
        end_points['fc2'] = net
        if not num_classes:  # num_classes为0、Nnoe，返回的不是最终的输出层，
            return net, end_points
        logits = slim.fully_connected(net, num_classes,
                                      biases_initializer=tf.zeros_initializer(),
                                      weights_initializer=tf.truncated_normal_initializer(1 / 192.0),
                                      weights_regularizer=None,
                                      activation_fn=None,
                                      scope='logits') # 输出层 ；shape [n,num_classes]

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

    return logits, end_points


cifarnet.default_image_size = 28  # 这里使用mnist数据  如果使用cifar 改成32

def cifarnet_arg_scope(weight_decay=0.004):
  """Defines the default cifarnet argument scope.
  设置各层的一些参数
  Args:
    weight_decay: 用于规范模型的权重衰减。.
  Returns:
    用于inception v3模型的`arg_scope`。
  """
  with slim.arg_scope(
      [slim.conv2d], # 设置卷积层一些参数，
      weights_initializer=tf.truncated_normal_initializer(stddev=5e-2),# 权重初始化
      activation_fn=tf.nn.relu): # 设置激活函数
    with slim.arg_scope(
        [slim.fully_connected], # 设置全连接层的一些参数
        biases_initializer=tf.constant_initializer(0.1), # biases 初始化为0.1
        weights_initializer=tf.truncated_normal_initializer(0.04), # 权重初始化
        weights_regularizer=slim.l2_regularizer(weight_decay), # L2正则化
        activation_fn=tf.nn.relu) as sc:
      return sc


def inception_resnet_v2_arg_scope(weight_decay=0.00004,
                                  batch_norm_decay=0.9997,
                                  batch_norm_epsilon=0.001,
                                  activation_fn=tf.nn.relu):
  """Returns the scope with the default parameters for inception_resnet_v2.
  使用了batch_norm，相对来说要比cifarnet_arg_scope效果更佳，推荐使用该方式进行各层参数配置
  Args:
    weight_decay: the weight decay for weights variables.
    batch_norm_decay: decay for the moving average of batch_norm momentums.
    batch_norm_epsilon: small float added to variance to avoid dividing by zero.
    activation_fn: Activation function for conv2d.
  Returns:
    a arg_scope with the parameters needed for inception_resnet_v2.
  """
  # Set weight_decay for weights in conv2d and fully_connected layers.
  with slim.arg_scope([slim.conv2d, slim.fully_connected],
                      weights_regularizer=slim.l2_regularizer(weight_decay),
                      biases_regularizer=slim.l2_regularizer(weight_decay)):

    batch_norm_params = {
        'decay': batch_norm_decay,
        'epsilon': batch_norm_epsilon,
        'fused': None,  # Use fused batch norm if possible.
    }
    # Set activation_fn and parameters for batch_norm.
    with slim.arg_scope([slim.conv2d], activation_fn=activation_fn,
                        normalizer_fn=slim.batch_norm,
                        normalizer_params=batch_norm_params) as scope:
      return scope


class Conv_model(object):
    # def __init__(self, X, Y, weights, biases, learning_rate, keep):
    def __init__(self, Y, learning_rate):
        # super(Conv_model, self).__init__(X,Y,w,b,learning_rate)  # 返回父类的对象
        # 或者 model.Model.__init__(self,X,Y,w,b,learning_rate)
        # self.X = X
        self.Y = Y
        # self.weights = weights
        # self.biases = biases
        self.learning_rate = learning_rate
        # self.keep = keep

    '''
    def conv2d(self, x, W, b, strides=1):
        # Conv2D wrapper, with bias and relu activation
        x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
        x = tf.nn.bias_add(x, b)  # strides中间两个为1 表示x,y方向都不间隔取样
        return tf.nn.relu(x)

    def maxpool2d(self, x, k=2):
        # MaxPool2D wrapper
        return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
                              padding='SAME')  # strides中间两个为2 表示x,y方向都间隔1个取样

    def inference(self, name='conv', activation='softmax'):  # 重写inference函数
        with tf.name_scope(name):
            conv1 = self.conv2d(self.X, self.weights['wc1'], self.biases['bc1'])
            conv1 = self.maxpool2d(conv1, k=2)  # shape [N,1,1,32]
            conv1 = tf.nn.lrn(conv1, depth_radius=5, bias=2.0, alpha=1e-3, beta=0.75)
            conv1 = tf.nn.dropout(conv1, self.keep)

            fc1 = tf.reshape(conv1, [-1, self.weights['wd1'].get_shape().as_list()[0]])
            fc1 = tf.add(tf.matmul(fc1, self.weights['wd1']), self.biases['bd1'])
            fc1 = tf.nn.relu(fc1)
            fc1 = tf.nn.dropout(fc1, self.keep)

            y = tf.add(tf.matmul(fc1, self.weights['out']), self.biases['out'])

        if activation == 'softmax':
            y = tf.nn.softmax(y)
        return y
    '''

    def loss(self, pred_value, MSE_error=False, one_hot=True):
        if MSE_error:
            return tf.reduce_mean(tf.reduce_sum(
                tf.square(pred_value - self.Y), reduction_indices=[1]))
        else:
            if one_hot:
                return tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
                    labels=self.Y, logits=pred_value))
            else:
                return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
                    labels=tf.cast(self.Y, tf.int32), logits=pred_value))

    def evaluate(self, pred_value, one_hot=True):
        if one_hot:
            correct_prediction = tf.equal(tf.argmax(pred_value, 1), tf.argmax(self.Y, 1))
            # correct_prediction = tf.nn.in_top_k(pred_value, Y, 1)
        else:
            correct_prediction = tf.equal(tf.argmax(pred_value, 1), tf.cast(self.Y, tf.int64))
        return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

    def train(self, cross_entropy):
        global_step = tf.Variable(0, trainable=False)
        return tf.train.GradientDescentOptimizer(self.learning_rate).minimize(cross_entropy,
                                                                              global_step=global_step)

class Inputs(object):
    def __init__(self,file_path,batch_size,one_hot=True):
        self.file_path=file_path
        self.batch_size=batch_size
        self.mnist=input_data.read_data_sets(self.file_path, one_hot=one_hot)
    def inputs(self):
        batch_xs, batch_ys = self.mnist.train.next_batch(self.batch_size)
        return batch_xs, batch_ys
    def test_inputs(self):
        return self.mnist.test.images,self.mnist.test.labels

FLAGS=None

def train():
    # 开始搭建cnn

    input_model = Inputs(FLAGS.data_dir, FLAGS.batch_size, one_hot=FLAGS.one_hot)

    with tf.name_scope('input'):
        x = tf.placeholder(tf.float32, [None, 28*28*1],'x')
        y_ = tf.placeholder(tf.float32, [None,10],'y_')
        keep=tf.placeholder(tf.float32)
        is_training= tf.placeholder(tf.bool, name='MODE')

    # image_shaped_input = tf.reshape(x, [-1, 28, 28, 1])

    # with slim.arg_scope(cifarnet_arg_scope()):
    with slim.arg_scope(inception_resnet_v2_arg_scope()):
        y, _ = cifarnet(images=x,num_classes=10,is_training=is_training,dropout_keep_prob=keep)
    model=Conv_model(y_,FLAGS.learning_rate)
    cross_entropy = model.loss(y, MSE_error=False, one_hot=FLAGS.one_hot)
    train_op = model.train(cross_entropy)
    accuracy = model.evaluate(y, one_hot=FLAGS.one_hot)
    init = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())

    with tf.Session() as sess:
        sess.run(init)
        for step in range(FLAGS.num_steps):
            batch_xs, batch_ys = input_model.inputs()
            train_op.run({x: batch_xs, y_: batch_ys,keep:0.7,is_training:True})

            if step % FLAGS.disp_step == 0:
                acc=accuracy.eval({x: batch_xs, y_: batch_ys,keep:1.,is_training:False})
                print("step", step, 'acc', acc,
                      'loss', cross_entropy.eval({x: batch_xs, y_: batch_ys,keep:1.,is_training:False}))
        # test acc
        test_x, test_y = input_model.test_inputs()
        acc = accuracy.eval({x: test_x, y_: test_y,keep:1.,is_training:False})
        print('test acc', acc)

def main(_):
    # if tf.gfile.Exists(FLAGS.log_dir):
    #     tf.gfile.DeleteRecursively(FLAGS.log_dir)
    # if not tf.gfile.Exists(FLAGS.log_dir):
    #     tf.gfile.MakeDirs(FLAGS.log_dir)
    train()

if __name__=="__main__":
    # 设置必要参数
    parser = argparse.ArgumentParser()
    parser.add_argument('--num_steps', type=int, default=1000,
                        help = 'Number of steps to run trainer.')
    parser.add_argument('--disp_step', type=int, default=100,
                        help='Number of steps to display.')
    parser.add_argument('--learning_rate', type=float, default=0.01,
                        help='Learning rate.')
    parser.add_argument('--batch_size', type=int, default=128,
                        help='Number of mini training samples.')
    parser.add_argument('--one_hot', type=bool, default=True,
                        help='One-Hot Encoding.')
    parser.add_argument('--data_dir', type=str, default='./MNIST_data',
            help = 'Directory for storing input data')
    parser.add_argument('--log_dir', type=str, default='./log_dir',
                        help='Summaries log directory')
    FLAGS, unparsed = parser.parse_known_args()
    tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

结果：