【作业四】林轩田机器学习技法 + 机器学习公开新课学习个人体会
这次作业的coding任务量比较大,总的来说需要实现neural network, knn, kmeans三种模型。
Q11~Q14为Neural Network的题目,我用单线程实现的,运行的时间比较长,因此把这几道题的正确答案记录如下:
Q11: 6
Q12: 0.001
Q13: 0.01
Q14: 0.02 ≤ Eout ≤ 0.04
其中Q11和Q14的答案比较明显,Q12和Q13有两个答案比较接近(参考了讨论区的内容,最终也调出来了)
neural network的代码实现思路如下:
1)实现权重矩阵W初始化(def init_W(nnet_struct, w_range))
2)实现计算每一轮神经元输出的函数,即bp算法中的forward过程(def forward_process(x, y, W))
3)实现计算每一轮output error对于每个神经元输入score的导数,即bp算法中的backward过程(def backward_process(x, y, neuron_output, W))
4)利用梯度下降方法,更新各层权重矩阵W的函数(def update_W_withGD(x, neuron_output, gradient, W, ita))
其中最难的是步骤3),要想实现矩阵化编程,需要对神经网络的每层结构熟练,同时对于你使用的编程语言的矩阵化操作要非常熟悉;自己在这个方面比较欠缺,还得是熟能生巧。
>>自己第一次写NNet的算法,从单隐层(隐层个数2)开始调试的:按照模块1)2)3)4)的顺序,各个模块调试;循序渐进的调试速度比较慢,但模块质量高一些,后面的联合调试就省事一些。
>>如果是特别复杂的网络,如何对这种gradient的算法进行调试呢?因为gradient各个点的gradient几乎是不可能都算到的,在网上查了gradient checking方法:http://ufldl.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization
>>NNet的调参真的很重要,就Q14来说,即使是hidden units的总个数一样,如果每层的个数不同,最后的结果也是有差别的(我第一次比较粗心,把NNet的结构按照 3 8 1这样了,发现结果没有 8 3 1这样好),后面多搜搜调参相关的资料积累一下。
代码如下(没有把调试的代码删掉,可以记录调试的经过,同时也防止以后犯类似的错误),确实乱了一些,请看官包涵了:
#encoding=utf8 import sys import numpy as np import math from random import * ## # read data from local file # return with numpy array def read_input_data(path): x = [] y = [] for line in open(path).readlines(): if line.strip()=='': continue items = line.strip().split(' ') tmp_x = [] for i in range(0,len(items)-1): tmp_x.append(float(items[i])) x.append(tmp_x) y.append(float(items[-1])) return np.array(x),np.array(y) ## # initialize weight matrix # input neural network structure & initilizing uniform value range (both low and high) # each layer's bias need to be added # return with inialized W def init_W(nnet_struct, w_range): W = [] for i in range(1,len(nnet_struct)): tmp_w = np.random.uniform(w_range['low'], w_range['high'], (nnet_struct[i-1]+1,nnet_struct[i]) ) W.append(tmp_w) return W ## # randomly pick sample from raw data for Stochastic Gradient Descent # T indicates the iterative numbers # return with data for each SGD iteration def pick_SGD_data(x, y, T): sgd_x = np.zeros((T,x.shape[1])) sgd_y = np.zeros(T) for i in range(T): index = randint(0, x.shape[0]-1) sgd_x[i] = x[index] sgd_y[i] = y[index] return sgd_x, sgd_y ## # forward process # calculate each neuron's output def forward_process(x, y, W): ret = [] #print W[0].shape #print W[1].shape pre_x = np.hstack((1,x)) for i in range(len(W)): pre_x = np.tanh(np.dot(pre_x, W[i])) ret.append(pre_x) pre_x = np.hstack((1,pre_x)) return ret ## # backward process # calcultae the gradient of error and each neuron's input score def backward_process(x, y, neuron_output, W): ret = [] L = len(neuron_output) # print neuron_output[0].shape, neuron_output[1].shape # Output layer score = np.dot( np.hstack((1, neuron_output[L-2])), W[L-1]) # print score # print score.shape gradient = np.array( [-2 * (y-neuron_output[L-1][0]) * tanh_gradient(score)] ) # print gradient # print gradient.shape ret.insert(0, gradient) # Hidden layer for i in range(L-2,-1,-1): if i==0: score = np.dot(np.hstack((1, x)),W[i]) # print score.shape # print gradient.shape # print W[1][1:].transpose().shape # print score gradient = np.dot(gradient, W[1][1:].transpose()) * tanh_gradient(score) # print gradient # print gradient.shapeq ret.insert(0, gradient) else: score = np.dot(np.hstack((1,neuron_output[i-1])),W[i]) # print score.shape # print gradient.shape # print W[i+1][1:].transpose().shape # print "......" gradient = np.dot(gradient , W[i+1][1:].transpose()) * tanh_gradient(score) # print gradient.shape # print "======" ret.insert(0, gradient) return ret # give a numpy array # boardcast tanh gradient to each element def tanh_gradient(s): ret = np.zeros(s.shape) for i in range(s.shape[0]): ret[i] = 4.000001 / (math.exp(2*s[i])+math.exp(-2*s[i])+2) return ret ## # update W with Gradient Descent def update_W_withGD(x, neuron_output, gradient, W, ita): ret = [] L = len(W) # print "L:"+str(L) # print neuron_output[0].shape, neuron_output[1].shape # print gradient[0].shape, gradient[1].shape # print W[0].shape, W[1].shape # print np.hstack((1,x)).transpose().shape # print gradient[0].shape ret.append( W[0] - ita * np.array([np.hstack((1,x))]).transpose() * gradient[0] ) for i in range(1, L, 1): ret.append( W[i] - ita * np.array([np.hstack((1,neuron_output[i-1]))]).transpose() * gradient[i] ) # print len(ret) return ret ## # calculate Eout def calculate_E(W, path): x,y = read_input_data(path) error_count = 0 for i in range(x.shape[0]): if predict(x[i],y[i],W): error_count += 1 return 1.000001*error_count/x.shape[0] def predict(x, y, W): y_predict = x for i in range(0, len(W), 1): y_predict = np.tanh( np.dot( np.hstack((1,y_predict)), W[i] ) ) y_predict = 1 if y_predict>0 else -1 return y_predict!=y ## # Q11 def Q11(x,y): R = 20 # repeat time Ms = { 6, 16 } # hidden units M_lowests = {} for M in Ms: M_lowests[M] = 0 for r in range(R): T = 50000 ita = 0.1 min_M = -1 E_min = float("inf") for M in Ms: sgd_x, sgd_y = pick_SGD_data(x, y, T) nnet_struct = [ x.shape[1], M, 1 ] # print nnet_struct w_range = {} w_range['low'] = -0.1 w_range['high'] = 0.1 W = init_W(nnet_struct, w_range) # for i in range(len(W)): # print W[i] # print sgd_x,sgd_y for t in range(T): neuron_output = forward_process(sgd_x[t], sgd_y[t], W) # print sgd_x[t],sgd_y[t] # print W # print neuron_output error_neuronInputScore_gradient = backward_process(sgd_x[t], sgd_y[t], neuron_output, W) # print error_neuronInputScore_gradient W = update_W_withGD(sgd_x[t], neuron_output, error_neuronInputScore_gradient, W, ita) E = calculate_E(W,"test.dat") # print str(r)+":::"+str(M)+":"+str(E) M_lowests[M] += E for k,v in M_lowests.items(): print str(k)+":"+str(v) ## # Q12 def Q12(x,y): ita = 0.1 M = 3 nnet_struct = [ x.shape[1], M, 1 ] Rs = { 0.001, 0.1 } R_lowests = {} for R in Rs: R_lowests[R] = 0 N = 40 T = 30000 for i in range(N): for R in Rs: sgd_x, sgd_y = pick_SGD_data(x, y, T) w_range = {} w_range['low'] = -1*R w_range['high'] = R W = init_W(nnet_struct, w_range) for t in range(T): neuron_output = forward_process(sgd_x[t], sgd_y[t], W) error_neuronInputScore_gradient = backward_process(sgd_x[t], sgd_y[t], neuron_output, W) W = update_W_withGD(sgd_x[t], neuron_output, error_neuronInputScore_gradient, W, ita) E = calculate_E(W, "test.dat") print str(R)+":"+str(E) R_lowests[R] += E for k,v in R_lowests.items(): print str(k)+":"+str(v) ## # Q13 def Q13(x,y): M = 3 nnet_struct = [ x.shape[1], M, 1 ] itas = {0.001,0.01,0.1} ita_lowests = {} for ita in itas: ita_lowests[ita] = 0 N = 20 T = 20000 for i in range(N): for ita in itas: sgd_x, sgd_y = pick_SGD_data(x, y, T) w_range = {} w_range['low'] = -0.1 w_range['high'] = 0.1 W = init_W(nnet_struct, w_range) for t in range(T): neuron_output = forward_process(sgd_x[t], sgd_y[t], W) error_neuronInputScore_gradient = backward_process(sgd_x[t], sgd_y[t], neuron_output, W) W = update_W_withGD(sgd_x[t], neuron_output, error_neuronInputScore_gradient, W, ita) E = calculate_E(W, "test.dat") print str(ita)+":"+str(E) ita_lowests[ita] += E for k,v in ita_lowests.items(): print str(k)+":"+str(v) ## # Q14 def Q14(x,y): T = 50000 ita = 0.01 E_total = 0 R = 10 for i in range(R): nnet_struct = [ x.shape[1], 8, 3, 1 ] w_range = {} w_range['low'] = -0.1 w_range['high'] = 0.1 W = init_W(nnet_struct, w_range) sgd_x, sgd_y = pick_SGD_data(x, y, T) for t in range(T): neuron_output = forward_process(sgd_x[t], sgd_y[t], W) error_neuronInputScore_gradient = backward_process(sgd_x[t], sgd_y[t], neuron_output, W) W = update_W_withGD(sgd_x[t], neuron_output, error_neuronInputScore_gradient, W, ita) E = calculate_E(W, "test.dat") print E E_total += E print E_total*1.0/R def main(): x,y = read_input_data("train.dat") # print x.shape, y.shape # Q11(x, y) # Q12(x, y) # Q13(x, y) Q14(x, y) if __name__ == '__main__': main()
Q15~Q18是KNN算法相关的,各道题几乎秒出结果,这里不记录答案了:
KNN的核心,也就是KNN函数了:
1)给定K个邻居数,返回这个点属于哪一类,代码尽量写的可配置一些
2)numpy有个argsort函数,可以根据数组的value大小,对下标index进行排序;并返回排序后的index;利用好这个特性,代码很简洁
3)如果是其他的语言,应该实现一个类似numpy.argsort的模块,代码整体上清晰不少能
KNN的代码如下:
#encoding=utf8 import sys import numpy as np import math from random import * ## # read data from local file # return with numpy array def read_input_data(path): x = [] y = [] for line in open(path).readlines(): if line.strip()=='': continue items = line.strip().split(' ') tmp_x = [] for i in range(0,len(items)-1): tmp_x.append(float(items[i])) x.append(tmp_x) y.append(float(items[-1])) return np.array(x),np.array(y) ## # KNN ( for binary classification ) # input all labeled data & test sample # return with label def KNN(k, x, y, test_x): distance = np.sum((x-test_x)*(x-test_x), axis=1) order = np.argsort(distance) ret = 0 for i in range(k): ret += y[order[i]] return 1 if ret>0 else -1 ## # Q15 calculate Ein def calculate_Ein(x, y): error_count = 0 k = 5 for i in range(x.shape[0]-1): # tmp_x = np.vstack( ( x[0:i],x[(i+1):(x.shape[0]-1)] ) ) # tmp_y = np.hstack( ( y[0:i],y[(i+1):(x.shape[0]-1)] ) ) ret = KNN( k, x, y, x[i]) if y[i]!=ret: error_count += 1 return 1.0*error_count/x.shape[0] ## # Q16 calculate Eout def calculate_Eout(x, y, path): test_x, test_y = read_input_data(path) error_count = 0 k = 1 for i in range(test_x.shape[0]): ret = KNN (k, x, y, test_x[i]) if test_y[i]!=ret: error_count += 1 return 1.0*error_count/test_x.shape[0] def main(): x,y = read_input_data("knn_train.dat") print calculate_Ein(x,y) print calculate_Eout(x,y, "knn_test.dat") if __name__ == '__main__': main()
Q19~Q20是Kmeans算法相关的,运行代码也很快可以得出结果,不记录答案了:
Kmeans的算法实现思路非常清晰:
1)实现初始化随机选各类中心点的功能(题目中是随机选原始数据的点,如果是其他的选点方法,单独拎出来一个模块,不影响其他模块)
2)实现每次更新各个数据点类别的功能(def update_category(x, K, centers))
3)固定各个点的类别,更新各个类别的center点坐标(def update_centers(x, y, K))
模块实现上,得益于numpy的矩阵计算操作函数。(应该掌握一套自己的矩阵计算操作代码,这样可以随时拿起来二次开发)
代码如下:
#encoding=utf8 import sys import numpy as np import math from random import * ## # read data from local file # return with numpy array def read_input_data(path): x = [] for line in open(path).readlines(): if line.strip()=='': continue items = line.strip().split(' ') tmp_x = [] for i in range(0,len(items)): tmp_x.append(float(items[i])) x.append(tmp_x) return np.array(x) ## # input all data and category K # return K category centers def Kmeans(x, K): T = 50 E_total = 0 for t in range(T): centers = init_centers(x, K) y = np.zeros(x.shape[0]) R = 50 for r in range(R): y = update_category(x, K, centers) centers = update_centers(x, y, K) E = calculate_Ein(x, y, centers) print E E_total += E return E_total*1.0/T def init_centers(x, K): ret = [] order = range(x.shape[0]) np.random.shuffle(order) for i in range(K): ret.append(x[order[i]]) return np.array(ret) def update_category(x, K, centers): y = [] for i in range(x.shape[0]): category = -1 distance = float("inf") for k in range(K): d = np.sum((x[i] - centers[k])*(x[i] - centers[k]),axis=0) if d < distance: distance = d category = k y.append(category) return np.array(y) def update_centers(x, y, K): centers = [] for k in range(K): # print "np.sum(x[np.where(y==k)],axis=0)" # print np.sum(x[np.where(y==k)],axis=0).shape center = np.sum(x[np.where(y==k)],axis=0)*1.0/np.array(np.where(y==k)).shape[1] centers.append(center) return np.array(centers) def calculate_Ein(x, y, centers): # print centers[0].shape error_total = 0 for i in range(x.shape[0]): error_total += np.sum((x[i]-centers[y[i]])*(x[i]-centers[y[i]]),axis=0) return 1.0*error_total/x.shape[0] def main(): x = read_input_data("kmeans_train.dat") # print x.shape print Kmeans(x,2) if __name__ == '__main__': main()
==========================================================================
完成了这次作业后,终于跟完了《机器学习基石+机器学习技法》32次课,8次coding作业。
个人上完这门课后,主要有三点收获:
1)通过coding的作业题目,实现了一些主流机器学习算法(Perceptron、AdaBoost-stump、Linear Regression、Logistic Regression、Decision Tree、Neural Network、KNN、Kmeans);以前都是用算法包,对各个算法的理解不如实现过一遍来得深和细。
2)以前对各个算法的理解就是会用(其实也不能说太会用),上完课程后,对每个模型的Motivation有了一定的掌握:模型为什么要这么设计?Regularizer为什么要这么设计?模型的利弊有哪些?以及模型的一些比较直观的数学原理推导。
3)以前看待各个机器学习算法,都是孤立的看待每个算法(这个算法是解决啥的,那个算法是解决啥的),没有成体系地把各个算法拎起来。台大这门课在整个授课环节中,都贯穿了非常强的体系的观念,这里举两个例子:
a. Linear Network与Factorization有啥联系(15讲)
b. Decision Tree与AdaBoost有啥关系(8、9讲)
c. Linear Regression与Neural Network有啥关系(12讲)
在看这门课之前,是绝对不会把上面的每组中两个模型联系起来看待的;但这门课确实给了比较深的motivation,非常强的全局主线。
最后,谈一点个人上公开课的体会:
1)只听一遍:走马观花,学到的东西微乎其微
2)听课,写作业:实践者的态度去学,学到的东西比只听课要多了去了
3)听课,写作业,写听课blog:实践者+研究者的态度去学;“最好的学就是教”,在写blog的过程中,会强迫自己把当时很多不清晰的point都搞清楚,要不然真的写不出来
4)循环进行3):温故知新的道理大家都懂,就看有没有时间吧
Sign 就写到这了.....