anaconda+theano+keras手写字符识别新版 - 邗影

anaconda+theano+keras手写字符识别新版

标题介绍运行环境了win7

看网上好多keras识别minist 但是一般由于版本问题，无法直接用，，，这里还要特别感谢keras中文文档作者（三当家SCP）。教程整的非常好。还有就是最好你在安装anaconda 之前把原来安装过的PY卸载掉，要不然安装mingw的时候会出问题，，，安装就不详细介绍了网上有很多种----大致流程——anaconda-mingw-theano(注意环境变量，系统变量啥的)-keras。

下边附上一个可用程序哈，亲测可用。。。并附上数据，数据来源于网络，见文章底部，你就不用运行的时候在下载数据，这样很容易出错的。非GPU环境 win7 64bit

好的：其实重点是新的语法的问题。。。。。。。。。。。。。。。。。。。更新地方已标红。。。。。。。

 1 from __future__ import absolute_import
 2 from __future__ import print_function
 3 import numpy as np
 4 np.random.seed(1337)  # for reproducibility
 5 import cPickle as pickle  
 6 
 7 from keras.models import Sequential
 8 from keras.layers.core import Dense, Dropout, Activation
 9 from keras.optimizers import SGD, Adam, RMSprop
10 from keras.utils import np_utils
11 
12 \'\'\'
13     Train a simple deep NN on the MNIST dataset.
14     Get to 98.30% test accuracy after 20 epochs (there is *a lot* of margin for parameter tuning).
15     2 seconds per epoch on a GRID K520 GPU.
16 \'\'\'
17 
18 batch_size = 128
19 nb_classes = 10
20 nb_epoch = 10
21 
22 
23 def read_data(data_file):  
24     import gzip  
25     f = gzip.open(data_file, "rb")  
26     train, val, test = pickle.load(f)  
27     f.close()  
28     train_x = train[0]  
29     train_y = train[1]  
30     test_x = test[0]  
31     test_y = test[1]  
32     return train_x, train_y, test_x, test_y  
33     
34     
35 # the data, shuffled and split between tran and test sets
36 #(X_train, y_train), (X_test, y_test) = mnist.load_data()
37 train_x, train_y, test_x, test_y = read_data("C:\Users\PC\.spyder2\mnist.pkl.gz")  
38 X_train = train_x
39 X_test = test_x
40 X_train = X_train.astype("float32")
41 X_test = X_test.astype("float32")
42 X_train /= 255
43 X_test /= 255
44 print(X_train.shape[0], \'train samples\')
45 print(X_test.shape[0], \'test samples\')
46 
47 # convert class vectors to binary class matrices
48 Y_train = np_utils.to_categorical(train_y, nb_classes)
49 Y_test = np_utils.to_categorical(test_y, nb_classes)
50 
51 model = Sequential()
52 model.add(Dense(input_dim=784, output_dim=128))
53 model.add(Activation(\'relu\'))
54 model.add(Dropout(0.2))
55 model.add(Dense(output_dim=128))
56 model.add(Activation(\'relu\'))
57 model.add(Dropout(0.2))
58 model.add(Dense(output_dim=10))
59 model.add(Activation(\'softmax\'))
60 
61 rms = RMSprop()
62 model.compile(loss=\'categorical_crossentropy\', optimizer=rms,metrics=[\'accuracy\'])
63 
64 model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=nb_epoch)
65 score = model.evaluate(X_test, Y_test, batch_size=batch_size)
66 print(\'Test score:\', score[0])
67 print(\'Test accuracy:\', score[1])

数据在这里：http://www.cnblogs.com/xueliangliu/archive/2013/04/03/2997437.html。。。由于没法上传有15兆，我看这个大牛有这个数据，而且还附带了安装及原理等，自己看吧。。。。。。。