Theano(1) 安装(Window64)及教程

一、Theano的安装

   环境：Windows7 64位 + Python2.7  
   依赖：Numpy、Scipy包
   正文 : 
   1>Theano的编译安装需要上面提到的两个包，还需要g++的编译器，不幸的是一般我们只有GCC。解决办法是安装Minwg，注意要64位的，否则不兼容。
   下载地址：https://sourceforge.net/projects/mingw-w64/
   按提示一步步安装就可以了，但要添加环境变量到Path(到bin目录一级就可以了)，使用gcc -v命令查看是否安装成功
   2>之前已经安装了pip，所以可以直接使用pip安装：

 pip install theano

import theano 就会进行初次编译

二、教程翻译

官方文档：http://deeplearning.net/software/theano/tutorial/
译文1：http://www.gumpcs.com/index.php/archives/576
译文2：http://blog.csdn.net/walegahaha/article/details/50884627
译文3：http://www.cnblogs.com/charleshuang/p/3648804.html
译文4：http://www.cnblogs.com/charleshuang/p/3651843.html
还有某个大大的全部翻译版：
http://www.cnblogs.com/shouhuxianjian/category/699462.html
有时间再写心得。。。

softmax 的粗略实现，依然使用文本分类任务：

#coding=utf-8
from numpy import *
import theano
import theano.tensor as T
from sklearn.datasets import load_files
from sklearn.feature_extraction.text import CountVectorizer
from sklearn import metrics
import warnings
warnings.filterwarnings("ignore")

def calculate_result(actual,pred):
    m_precision = metrics.precision_score(actual,pred)
    m_recall = metrics.recall_score(actual,pred)
    m_acc = metrics.accuracy_score(actual,pred)
    print 'predict info:'
    print 'accuracy:{0:.3f}'.format(m_acc)
    print 'precision:{0:.3f}'.format(m_precision)
    print 'recall:{0:0.3f}'.format(m_recall)
    print 'f1-score:{0:.3f}'.format(metrics.f1_score(actual,pred))

if __name__ == '__main__':
#load datasets
    doc_train = load_files('training')
    doc_test = load_files('test')
#Bool型特征（one-hot）
    count_vec = CountVectorizer(binary=True,decode_error='replace')
    doc_train_bool = count_vec.fit_transform(doc_train.data)
    doc_test_bool = count_vec.transform(doc_test.data)
    m, n = shape(doc_train_bool)
#generate a dataset: D = (input_values, target_class)
    D = (doc_train_bool.toarray(), doc_train.target)
    training_steps = 100

# Declare Theano symbolic variables
    x = T.matrix("x")
    y = T.ivector("y")

    w = theano.shared(zeros([n,10]), name="w")
# initialize the bias term
    b = theano.shared(zeros(10), name="b")

    print("Initial model:")
    #print(w.get_value())

# Construct Theano expression graph
    p = T.nnet.softmax(T.dot(x,w) + b)# Probability that target = i   
    prediction = T.argmax(p,1)    # The prediction thresholded
    xent = T.log(p[T.arange(y.shape[0]), y])
    cost = -T.mean(xent)  #The cost to minimize
    gw, gb = T.grad(cost, [w, b]) # Compute the gradient of the cost

# Compile
    train = theano.function(
        inputs=[x,y],
        outputs=[prediction,xent,cost],
        updates=((w, w - 2 * gw), (b, b - 2 * gb)))
    #predict = theano.function(inputs=[x], outputs=prediction)
# Train
    for i in range(training_steps):
        print i,
        pred,xent,cost = train(D[0], D[1])
        print cost
#Test
    D1 = (doc_test_bool.toarray(), doc_test.target)
    test = theano.function(inputs=[x], outputs=prediction)
    calculate_result(doc_test.target,test(D1[0]))

结果：

predict info:
accuracy:0.879
precision:0.873
recall:0.879
f1-score:0.872