Keras的使用方法

一种high-level的python API。

cv-tricks教程篇
Keras 可以作为tensorflow和Theano的前端接口。
The power of being able to run the same code with different back-end is a great reason for choosing Keras.

Install

sudo pip install h5py  #Used for save the networksudo pip install numpy scipy pillowsudo pip install keras

装好后，可以到配置文件~/.keras/keras.json里指定后台运行在Tensorflow还是Theano上。

{    "epsilon": 1e-07,    "floatx": "float32",    "image_data_format": "channels_last",    "backend": "tensorflow"}

Tensorflow要用channels_last, 而Theano要用channels_first, 因为：

Note that, the value of image_data_format is “channels_last”, which is the correct value for Tensorflow. In Tensorflow, images are stored as Tensors/arrays of shape [height, width, channels] while in Theano the order is different [channels, height, width]. So, if you don’t have this parameter set correctly, your intermediate results will be very strange. For Theano, this value will be “channels_first”.

Fundamentals of Keras

两种方式可以创建模型：
1. Sequential models: 创建简单模型
2. Functional API: 创建复杂模型

Keras Sequential Model

使用这一种方法创建模型，首先需要创建一个顺序模型：

from keras.models import Sequentialmodel = Sequential()

定义网络结构的相关方法：Dense(fully connected layer),Activation,Conv2D,MaxPooling2D,Flatten,Dropout

from keras.layers import Dense, Activation,Conv2D,MaxPooling2D,Flatten,Dropoutmodel.add(Conv2D(64, (3, 3), activation='relu'))#This adds a Convolutional layer with 64 filters of size 3 * 3 to the graphmodel.add(MaxPooling2D(pool_size=(2, 2)))#MaxPooling layer:  Specify the type of layer and specify the pool size and you are done.model.add(Dense(256, activation='relu'))#Fully connected layer: It’s called Dense in Keras. Just specify the number of outputs and you are done.model.add(Dropout(0.5))model.add(Flatten())

网络的第一层需要读入数据，因此第一层网络需要额外指明数据的shape：

model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(224, 224, 3)))# Here the shape is 224 * 224 * 3.

网络的损失函数和优化方式。使用model.compile()函数。
For example, in this line below we are asking the network to use the ‘rmsprop’ optimizer to change weights in such a way that the loss ‘binary_crossentropy’ is minimized at each iteration.

model.compile(loss='binary_crossentropy', optimizer='rmsprop')

If you want to specify stochastic gradient descent and you want to choose proper initialization and other hyperparameters:

from keras.optimizers import SGD.............sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)model.compile(loss='categorical_crossentropy', optimizer=sgd)

使用fit函数数据输入。

model.fit(x_train, y_train, batch_size=32, epochs=10,validation_data=(x_val, y_val))# epochs: maximum numbers of epochs you want training to go on.

最后是使用evaluate函数进行模型测试。

score = model.evaluate(x_test, y_test, batch_size=32)

Sequence Model解决一个数值回归问题

import kerasfrom keras.models import Sequentialfrom keras.layers import Denseimport numpy as nptrX = np.linspace(-1, 1, 101)trY = 3 * trX + np.random.randn(*trX.shape) * 0.33model = Sequential()#model.add(Dense(input_dim=1, output_dim=1, init='uniform', activation='linear'))model.add(Dense(input_dim=1, units=1, kernel_initializer='uniform', activation='linear'))model.compile(optimizer='sgd', loss='mse')weights = model.layers[0].get_weights()w_init = weights[0][0][0]b_init = weights[1][0]print('Linear regression model is initialized with weight w: %.2f, b: %.2f' % (w_init, b_init))# Linear regression model is initialized with weight w: -0.04, b: 0.00model.fit(trX, trY, epochs=200, verbose=1)weights = model.layers[0].get_weights()w_final = weights[0][0][0]b_final = weights[1][0]print('Linear regression model is trained to have final weight w: %.2f, b: %.2f' % (w_final, b_final))# Linear regression model is trained to have final weight w: 2.98, b: 0.09

如果不使用Keras的话，tensorflow的代码应该这样写：

import tensorflow as tfimport numpy as nptrainX = np.linspace(-1, 1, 101)trainY = 3 * trainX + np.random.randn(*trainX.shape) * 0.33X = tf.placeholder("float")Y = tf.placeholder("float")w = tf.Variable(0.0, name="weights")init = tf.global_variables_initializer()print( "Your Tensorflow version is "+ str(tf.__version__) +".") print("If you Tensorflow version is < 0.11, you will face error in tf.multiply function. Check code comment"  )y_model = tf.multiply(X, w)# This was tf.mul for older versions cost = (tf.pow(Y-y_model, 2))train_op = tf.train.GradientDescentOptimizer(0.01).minimize(cost)with tf.Session() as sess:    sess.run(init)    for i in range(100):        for (x, y) in zip(trainX, trainY):            sess.run(train_op, feed_dict={X: x, Y: y})    print(sess.run(w))

简单对比一下可以发现：创建placeholder, session和变量及其初始化上面都可以省略，确实方便了一些。如果是创建卷积神经网络，则会方便更多。

保存网络和加载

这里使用的数据形式是HDF5 Binary format。保存方法：

model.save_weights("my_model.h5")

加载方法:

model.load_weights('my_model_weights.h5')