TensorFlow学习-基于CNN实现手写数字识别

• 一、网络结构
• 二、模型训练
• 三、模型评估
• 四、源码

一、网络结构

使用2个卷积层，2个池化层, 2个全连接层组成网络

输入→ 卷积→ ReLU→max polling→ 卷积→ ReLU→max polling→ FC→输出

1. 输入
   一个4维的tensor: [batch_size, image_width, image_height, channels], 分别代表梯度下降处理的批量数据大小，图片宽度，图片高度和图片的channel个数（彩色图片channel数为3[Red, Green, Blue]，单色图片channel数为1）

   # Input Layer # Reshape X to 4-D tensor: [batch_size, width, height, channels] # MNIST images are 28x28 pixels, and have one color channel input_layer  =  tf.reshape(features, [ - 1 ,  28 ,  28 ,  1 ])

   
   

2. 卷积层#1：
   采用32(channel)个5*5的过滤器(kernel)对原始输入图像做卷积(局部感知)， 另外对输入矩阵加了zero padding以保持卷积输出宽高和输入一致，并用ReLU作为激活函数引入非线性特性

   # Convolutional Layer #1   # Computes 32 features using a 5x5 filter with ReLU activation.   # Padding is added to preserve width and height.   # Input Tensor Shape: [batch_size, 28, 28, 1]   # Output Tensor Shape: [batch_size, 28, 28, 32]   conv1  =  tf.layers.conv2d(       inputs = input_layer,       filters = 32 ,       kernel_size = [ 5 ,  5 ],       padding = "same" ,       activation = tf.nn.relu)

   
   

3. 池化层#1
   采用2*2的过滤器(stride=2)对卷积层#1的输出做最大值下采样(max polling)， 降低了数据纬度，并避免过拟合

   # Pooling Layer #1 # First max pooling layer with a 2x2 filter and stride of 2 # Input Tensor Shape: [batch_size, 28, 28, 32] # Output Tensor Shape: [batch_size, 14, 14, 32] pool1  =  tf.layers.max_pooling2d(inputs = conv1, pool_size = [ 2 ,  2 ], strides = 2 )

4. 卷积层#2
   采用64个5*5的过滤器(kernel)对池化层#1的输出做卷积， 并用ReLU作为激活函数
   

   # Convolutional Layer #2 # Computes 64 features using a 5x5 filter. # Padding is added to preserve width and height. # Input Tensor Shape: [batch_size, 14, 14, 32] # Output Tensor Shape: [batch_size, 14, 14, 64] conv2  =  tf.layers.conv2d(      inputs = pool1,      filters = 64 ,      kernel_size = [ 5 ,  5 ],      padding = "same" ,      activation = tf.nn.relu)

   
   

5. 池化层#2
   采用2*2的过滤器(stride=2)对卷积层#2的输出做最大值下采样(max polling)
   

   # Pooling Layer #2 # Second max pooling layer with a 2x2 filter and stride of 2 # Input Tensor Shape: [batch_size, 14, 14, 64] # Output Tensor Shape: [batch_size, 7, 7, 64] pool2  =  tf.layers.max_pooling2d(inputs = conv2, pool_size = [ 2 ,  2 ], strides = 2 )

   
   

6. 全连接层#1
   首先把池化层#2的输出打平(flatten)成二维[batch_size, 7*7*64]矩阵，然后和1024个神经元做全连接，同时指定dropout=0.4(随机保留60%的数据做训练，避免过拟合)

   # Flatten tensor into a batch of vectors # Input Tensor Shape: [batch_size, 7, 7, 64] # Output Tensor Shape: [batch_size, 7 * 7 * 64] pool2_flat  =  tf.reshape(pool2, [ - 1 ,  7  *  7  *  64 ])   # Dense Layer # Densely connected layer with 1024 neurons # Input Tensor Shape: [batch_size, 7 * 7 * 64] # Output Tensor Shape: [batch_size, 1024] dense  =  tf.layers.dense(inputs = pool2_flat, units = 1024 , activation = tf.nn.relu)   # Add dropout operation; 0.6 probability that element will be kept dropout  =  tf.layers.dropout(      inputs = dense, rate = 0.4 , training = mode  = =  learn.ModeKeys.TRAIN)

   
   

7. 输出
   10个神经元，依次代表0-9

   # Logits layer # Input Tensor Shape: [batch_size, 1024] # Output Tensor Shape: [batch_size, 10] logits  =  tf.layers.dense(inputs = dropout, units = 10 )

   
   

二、模型训练

1. 对label做one-hot encoding

   # tf.one_hot接受两个参数： # indices代表one-hot encoding后，值为1的位置(其余为0) # depth代表目标值的个数(以手写数字识别为例，目标值为0-9， 所以depth=10) onehot_labels  =  tf.one_hot(indices = tf.cast(labels, tf.int32), depth = 10 )

2. 计算交叉熵损失:

   loss  =  tf.losses.softmax_cross_entropy(onehot_labels = onehot_labels, logits = logits)

3.  配置训练操作， 学习率=0.001，优化方法采用随机梯度下降：

   train_op  =  tf.contrib.layers.optimize_loss(          loss = loss,          global_step = tf.contrib.framework.get_global_step(),          learning_rate = 0.001 ,          optimizer = "SGD" )

4. 模型预测

   # Generate Predictions # classes： 预测的分类,取值0-9 # probabilities: classed对应的可能性， 经过softmax激活函数处理   predictions  =  {       "classes" : tf.argmax(           input = logits, axis = 1 ),       "probabilities" : tf.nn.softmax(           logits, name = "softmax_tensor" )   }

5. 创建评估器(Estimator)，返回一个分类器，能做训练和评估

   # Create the Estimator # 这里的cnn_model_fn几乎就是上面全部代码的一个wrap, 详见：https://www.tensorflow.org/tutorials/layers#building_the_cnn_mnist_classifier mnist_classifier  =  learn.Estimator(        model_fn = cnn_model_fn, model_dir = "/tmp/mnist_convnet_model" )

6. 训练：

   # Train the model mnist_classifier.fit(      x = train_data,      y = train_labels,      batch_size = 100 ,      steps = 20000 ,      monitors = [logging_hook])

三、模型评估

1. 配置评估metric并做评估

   # Configure the accuracy metric for evaluation    metrics  =  {        "accuracy" :            learn.MetricSpec(                metric_fn = tf.metrics.accuracy, prediction_key = "classes" ),    }      # Evaluate the model and print results    eval_results  =  mnist_classifier.evaluate(        x = eval_data, y = eval_labels, metrics = metrics)

四、源码

完整代码： https://www.github.com/tensorflow/tensorflow/blob/r1.1/tensorflow/examples/tutorials/layers/cnn_mnist.py

原文地址： https://www.tensorflow.org/tutorials/layers