12 学习笔记 例子

下面来看几个例子，来了解一下Keras的便捷之处。不需要具体去研究代码的意思，只需要看一下这个实现过程。用编程的装饰模式把各个组件模块化，然后可以自己随意的拼装。首先介绍一个基于Keras做的手写MNIST识别的代码，剩下的就看一下实现过程即可。

0用Keras实现MNIST识别。

from keras.models import Sequential   
from keras.layers.core import Dense, Dropout,Activation   
from keras.optimizers import SGD   
from keras.datasets import mnist   
   
import numpy as np 
   
model = Sequential()   
model.add(Dense(500, input_shape=(784,), init='glorot_uniform')) # 输入层，28*28=784   
model.add(Activation('tanh')) # 激活函数是tanh   
model.add(Dropout(0.5)) # 采用50%的dropout  
   
model.add(Dense(500, init='glorot_uniform')) # 隐层节点500个   
model.add(Activation('tanh'))   
model.add(Dropout(0.5))  
   
# 输出结果是10个类别，所以维度是10  
model.add(Dense(10))  
model.add(Activation('softmax')) # 最后一层用softmax  
   
# 设定学习率（lr）等参数   
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9,nesterov=True)    
# 使用交叉熵作为loss函数，就是熟知的log损失函数  
model.compile(loss='categorical_crossentropy',  
optimizer=sgd, class_mode='categorical')  

# 也可以现行下载，然后加载本地文件
path = r"https://s3.amazonaws.com/img-datasets/mnist.npz"
f = np.load(path)
X_train = f['x_train']
Y_train = f['y_train']
X_test = f['x_test']
Y_test = f['y_test']
f.close()

# 使用Keras自带的mnist工具读取数据（第一次需要联网）  
#(X_train, y_train), (X_test, y_test) = mnist.load_data()  
# 由于输入数据维度是(num, 28, 28)，这里需要把后面的维度直接拼起来变成784维  
X_train = X_train.reshape(X_train.shape[0],X_train.shape[1]* X_train.shape[2])  
X_test = X_test.reshape(X_test.shape[0], X_test.shape[1]* X_test.shape[2])   
# 这里需要把index转换成一个one hot的矩阵  
Y_train = (np.arange(10) == Y_train[:,None]).astype(int)   
Y_test = (np.arange(10) == Y_test[:,None]).astype(int)  
   
# 开始训练，这里参数比较多。batch_size就是batch_size，nb_epoch就是最多迭代的次数， shuffle就是是否把数据随机打乱之后再进行训练  
# verbose是屏显模式，官方这么说的：verbose: 0 forno logging to stdout, 1 for progress bar logging, 2 for one log line per epoch.  
# 就是说0是不屏显，1是显示一个进度条，2是每个epoch都显示一行数据  
# show_accuracy就是显示每次迭代后的正确率  
# validation_split就是拿出百分之多少用来做交叉验证  
model.fit(X_train, Y_train, batch_size=200, nb_epoch=20,shuffle=True, verbose=1, validation_split=0.3)   
print('test set')
model.evaluate(X_test, Y_test, batch_size=200,verbose=1) 

Epoch 86/100
42000/42000 [==============================] - 3s - loss: 0.3688 - val_loss: 0.2104
Epoch 87/100
42000/42000 [==============================] - 3s - loss: 0.3673 - val_loss: 0.2193
Epoch 88/100
42000/42000 [==============================] - 3s - loss: 0.3767 - val_loss: 0.2185
Epoch 89/100
42000/42000 [==============================] - 3s - loss: 0.3667 - val_loss: 0.2092
Epoch 90/100
42000/42000 [==============================] - 3s - loss: 0.3526 - val_loss: 0.2109
Epoch 91/100
42000/42000 [==============================] - 3s - loss: 0.3544 - val_loss: 0.2103
Epoch 92/100
42000/42000 [==============================] - 3s - loss: 0.3686 - val_loss: 0.2115
Epoch 93/100
42000/42000 [==============================] - 3s - loss: 0.3647 - val_loss: 0.2057
Epoch 94/100
42000/42000 [==============================] - 3s - loss: 0.3591 - val_loss: 0.2043
Epoch 95/100
42000/42000 [==============================] - 3s - loss: 0.3516 - val_loss: 0.2032
Epoch 96/100
42000/42000 [==============================] - 3s - loss: 0.3487 - val_loss: 0.2042
Epoch 97/100
42000/42000 [==============================] - 3s - loss: 0.3451 - val_loss: 0.2053
Epoch 98/100
42000/42000 [==============================] - 3s - loss: 0.3450 - val_loss: 0.2061
Epoch 99/100
42000/42000 [==============================] - 3s - loss: 0.3491 - val_loss: 0.2033
Epoch 100/100
42000/42000 [==============================] - 3s - loss: 0.3400 - val_loss: 0.2069
test set
 8800/10000 [=========================>....] - ETA: 0sOut[5]: 0.19544119497761131