keras 手把手入门#1-MNIST手写数字识别 深度学习实战闪电入门
本文属转载，仅用作个人学习。
人们已经教会计算机自动找出那些重要的特征和属性， 那么下一步我们改教会计算机什么？ — David 9
用深度学习框架跑过实际问题的朋友一定有这样的感觉: 太神奇了， 它竟然能自己学习重要的特征 ! 下一步我们改教会计算机什么？莫非是教会他们寻找新的未知特征？
对于卷积神经网络cnn， 其中每个卷积核就是一个cnn习得的特征， 详见David 9之前的关于cnn博客。
今天我们的主角是keras，其简洁性和易用性简直出乎David 9我的预期。大家都知道keras是在TensorFlow上又包装了一层，向简洁易用的深度学习又迈出了坚实的一步。
所以，今天就来带大家写keras中的Hello World ， 做一个手写数字识别的cnn。回顾cnn架构：
我们要处理的是这样的灰度像素图：
识别0-9的手写数字，60来行代码就能搞定，简单易懂，全程无尿点。我们先来看单机版跑完的结果：
- (python3_env) [email protected]:~$ python keras_mnist_cnn.py
- Using TensorFlow backend.
- x_train shape: (60000, 28, 28, 1)
- 60000 train samples
- 10000 test samples
- Train on 60000 samples, validate on 10000 samples
- Epoch 1/12
- 60000/60000 [==============================] - 446s - loss: 0.3450 - acc: 0.8933 - val_loss: 0.0865 - val_acc: 0.9735
- Epoch 2/12
- 60000/60000 [==============================] - 415s - loss: 0.1246 - acc: 0.9627 - val_loss: 0.0565 - val_acc: 0.9819
- Epoch 3/12
- 60000/60000 [==============================] - 389s - loss: 0.0957 - acc: 0.9717 - val_loss: 0.0493 - val_acc: 0.9842
- Epoch 4/12
- 60000/60000 [==============================] - 385s - loss: 0.0805 - acc: 0.9761 - val_loss: 0.0413 - val_acc: 0.9865
- Epoch 5/12
- 60000/60000 [==============================] - 366s - loss: 0.0708 - acc: 0.9788 - val_loss: 0.0361 - val_acc: 0.9874
- Epoch 6/12
- 60000/60000 [==============================] - 368s - loss: 0.0619 - acc: 0.9824 - val_loss: 0.0355 - val_acc: 0.9875
- Epoch 7/12
- 60000/60000 [==============================] - 363s - loss: 0.0585 - acc: 0.9826 - val_loss: 0.0345 - val_acc: 0.9875
- Epoch 8/12
- 60000/60000 [==============================] - 502s - loss: 0.0527 - acc: 0.9837 - val_loss: 0.0339 - val_acc: 0.9883
- Epoch 9/12
- 60000/60000 [==============================] - 444s - loss: 0.0501 - acc: 0.9852 - val_loss: 0.0309 - val_acc: 0.9891
- Epoch 10/12
- 60000/60000 [==============================] - 358s - loss: 0.0472 - acc: 0.9861 - val_loss: 0.0308 - val_acc: 0.9897
- Epoch 11/12
- 60000/60000 [==============================] - 358s - loss: 0.0441 - acc: 0.9872 - val_loss: 0.0330 - val_acc: 0.9893
- Epoch 12/12
- 60000/60000 [==============================] - 381s - loss: 0.0431 - acc: 0.9871 - val_loss: 0.0304 - val_acc: 0.9901
- Test loss: 0.0303807387119
- Test accuracy: 0.9901
所以我们跑的是keras_mnist_cnn.py。最后达到99%的预测准确率。首先来解释一下输出：
测试样本格式是28*28像素的1通道，灰度图，数量为60000个样本。
测试集是10000个样本。
一次epoch是一次完整迭代（所有样本都训练过），这里我们用了12次迭代，最后一次迭代就可以收敛到99.01%预测准确率了。
loss是训练集损失值. acc是训练集准确率。val_loss是测试集上的损失值，val_acc是测试集上的准确率。
接下来我们看代码：
- from __future__ import print_function
- import keras
- from keras.datasets import mnist
- from keras.models import Sequential
- from keras.layers import Dense, Dropout, Flatten
- from keras.layers import Conv2D, MaxPooling2D
- from keras import backend as K
一开始我们导入一些基本库，包括：
- mnist数据源
- Sequential类，可以封装各种神经网络层，包括Dense全连接层，Dropout层，Cov2D 卷积层，等等。
-
我们都知道keras支持两个后端TensorFlow和Theano，可以在
$HOME/.keras/keras.json中配置。
接下来，我们准备训练集和测试集，以及一些重要参数：
- # batch_size 太小会导致训练慢，过拟合等问题，太大会导致欠拟合。所以要适当选择
- batch_size = 128
- # 0-9手写数字一个有10个类别
- num_classes = 10
- # 12次完整迭代，差不多够了
- epochs = 12
- # 输入的图片是28*28像素的灰度图
- img_rows, img_cols = 28, 28
- # 训练集，测试集收集非常方便
- (x_train, y_train), (x_test, y_test) = mnist.load_data()
- # keras输入数据有两种格式，一种是通道数放在前面，一种是通道数放在后面，
- # 其实就是格式差别而已
- if K.image_data_format() == 'channels_first':
- x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
- x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
- input_shape = (1, img_rows, img_cols)
- else:
- x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
- x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
- input_shape = (img_rows, img_cols, 1)
- # 把数据变成float32更精确
- x_train = x_train.astype('float32')
- x_test = x_test.astype('float32')
- x_train /= 255
- x_test /= 255
- print('x_train shape:', x_train.shape)
- print(x_train.shape[0], 'train samples')
- print(x_test.shape[0], 'test samples')
- # 把类别0-9变成2进制，方便训练
- y_train = keras.utils.np_utils.to_categorical(y_train, num_classes)
- y_test = keras.utils.np_utils.to_categorical(y_test, num_classes)
然后，是令人兴奋而且简洁得令人吃鲸的训练构造cnn和训练过程：
- # 牛逼的Sequential类可以让我们灵活地插入不同的神经网络层
- model = Sequential()
- # 加上一个2D卷积层， 32个输出（也就是卷积通道），**函数选用relu，
- # 卷积核的窗口选用3*3像素窗口
- model.add(Conv2D(32,
- activation='relu',
- input_shape=input_shape,
- nb_row=3,
- nb_col=3))
- # 64个通道的卷积层
- model.add(Conv2D(64, activation='relu',
- nb_row=3,
- nb_col=3))
- # 池化层是2*2像素的
- model.add(MaxPooling2D(pool_size=(2, 2)))
- # 对于池化层的输出，采用0.35概率的Dropout
- model.add(Dropout(0.35))
- # 展平所有像素，比如[28*28] -> [784]
- model.add(Flatten())
- # 对所有像素使用全连接层，输出为128，**函数选用relu
- model.add(Dense(128, activation='relu'))
- # 对输入采用0.5概率的Dropout
- model.add(Dropout(0.5))
- # 对刚才Dropout的输出采用softmax**函数，得到最后结果0-9
- model.add(Dense(num_classes, activation='softmax'))
- # 模型我们使用交叉熵损失函数，最优化方法选用Adadelta
- model.compile(loss=keras.metrics.categorical_crossentropy,
- optimizer=keras.optimizers.Adadelta(),
- metrics=['accuracy'])
- # 令人兴奋的训练过程
- model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs,
- verbose=1, validation_data=(x_test, y_test))
完整地训练完毕之后，可以计算一下预测准确率：
- score = model.evaluate(x_test, y_test, verbose=0)
- print('Test loss:', score[0])
- print('Test accuracy:', score[1])
整个cnn的MNIST手写数字识别就训练完毕了，是不是非常简单，如果觉得注释还不够详尽，请自己试试源码感受下，或者看看我们文章的底部参考文献。
keras_mnist_cnn.py 完整源码：
- '''Trains a simple convnet on the MNIST dataset.
- Gets to 99.25% test accuracy after 12 epochs
- (there is still a lot of margin for parameter tuning).
- 16 seconds per epoch on a GRID K520 GPU.
- '''
- from __future__ import print_function
- import keras
- from keras.datasets import mnist
- from keras.models import Sequential
- from keras.layers import Dense, Dropout, Flatten
- from keras.layers import Conv2D, MaxPooling2D
- from keras import backend as K
- batch_size = 128
- num_classes = 10
- epochs = 12
- # input image dimensions
- img_rows, img_cols = 28, 28
- # the data, shuffled and split between train and test sets
- (x_train, y_train), (x_test, y_test) = mnist.load_data()
- if K.image_data_format() == 'channels_first':
- x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
- x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
- input_shape = (1, img_rows, img_cols)
- else:
- x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
- x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
- input_shape = (img_rows, img_cols, 1)
- x_train = x_train.astype('float32')
- x_test = x_test.astype('float32')
- x_train /= 255
- x_test /= 255
- print('x_train shape:', x_train.shape)
- print(x_train.shape[0], 'train samples')
- print(x_test.shape[0], 'test samples')
- # convert class vectors to binary class matrices
- y_train = keras.utils.np_utils.to_categorical(y_train, num_classes)
- y_test = keras.utils.np_utils.to_categorical(y_test, num_classes)
- model = Sequential()
- model.add(Conv2D(32,
- activation='relu',
- input_shape=input_shape,
- nb_row=3,
- nb_col=3))
- model.add(Conv2D(64, activation='relu',
- nb_row=3,
- nb_col=3))
- # model.add(MaxPooling2D(pool_size=(2, 2)))
- model.add(MaxPooling2D(pool_size=(2, 2)))
- model.add(Dropout(0.35))
- model.add(Flatten())
- model.add(Dense(128, activation='relu'))
- model.add(Dropout(0.5))
- model.add(Dense(num_classes, activation='softmax'))
- model.compile(loss=keras.metrics.categorical_crossentropy,
- optimizer=keras.optimizers.Adadelta(),
- metrics=['accuracy'])
- model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs,
- verbose=1, validation_data=(x_test, y_test))
- score = model.evaluate(x_test, y_test, verbose=0)
- print('Test loss:', score[0])
- print('Test accuracy:', score[1])
参考文献：
本文章属于“David 9的博客”原创，如需转载，请联系微信: david9ml，或邮箱：[email protected]
或直接扫二维码:
