1 # -*- coding: utf-8 -*- 2 """ 3 Created on Wed Apr 22 17:39:19 2015 4 5 @author: 90Zeng 6 """ 7 8 import numpy 9 import theano 10 import theano.tensor as T 11 import matplotlib.pyplot as plt 12 rng = numpy.random 13 N = 400 # 400个样本 14 feats = 784 # 每个样本的维度 15 D = (rng.randn(N, feats), rng.randint(size=N, low=0, high=2)) 16 training_steps = 10000 17 18 # Declare Theano symbolic variables 19 x = T.dmatrix("x") 20 y = T.dvector("y") 21 22 # 随机初始化权重 23 w = theano.shared(rng.randn(feats), name="w") 24 # 偏置初始化为 0 25 b = theano.shared(0.0, name="b") 26 print "Initial model:" 27 print w.get_value(), b.get_value() 28 29 # Construct Theano expression graph 30 p_1 = 1 / (1 + T.exp(-T.dot(x, w) - b)) # Probability that target = 1 31 prediction = p_1 > 0.5 # The prediction thresholded 32 xent = -y * T.log(p_1) - (1-y) * T.log(1-p_1) # Cross-entropy loss function 33 lost_avg = xent.mean() 34 cost = xent.mean() + 0.01 * (w ** 2).sum()# The cost to minimize 35 gw, gb = T.grad(cost, [w, b]) # Compute the gradient of the cost 36 # (we shall return to this in a 37 # following section of this tutorial) 38 39 # Compile 40 train = theano.function( 41 inputs=[x,y], 42 outputs=[prediction, lost_avg], 43 updates=((w, w - 0.1 * gw),(b, b - 0.1 * gb)), 44 ) 45 predict=theano.function( 46 inputs=[x], 47 outputs=prediction, 48 ) 49 50 # Train 51 err = [] 52 for i in range(training_steps): 53 pred, er = train(D[0], D[1]) 54 err.append(er) 55 56 print "Final model:" 57 print w.get_value(), b.get_value() 58 print "target values for D:", D[1] 59 print "prediction on D:", predict(D[0]) 60 61 # 画出损失函数图 62 x = range(1000) 63 plt.plot(x,err[0:1000])
损失函数随着迭代次数变化，运行结果：
