python中 x[:,0]和x[:,1] 理解與shuffle VS permutation 實際是關於梯度優化的問題


x[m,n]是通過numpy庫引用數組或矩陣中的某一段數據集的一種寫法,

m代表第m維,n代表m維中取第幾段特征數據。

通常用法:

x[:,n]或者x[n,:]

1.x[:,n]表示在全部數組(維)中取第n個數據,直觀來說,x[:,n]就是取所有集合的第n個數據, 

1 import numpy as np
2  
3 X = np.array([[0,1],[2,3],[4,5],[6,7],[8,9],[10,11],[12,13],[14,15],[16,17],[18,19]])
4 print X[:,0]

X[:,0]輸出結果是:

 

 

1 import numpy as np
2  
3 X = np.array([[0,1],[2,3],[4,5],[6,7],[8,9],[10,11],[12,13],[14,15],[16,17],[18,19]])
4 print X[:,1]

 

2. x[n,:]表示在n個數組(維)中取全部數據,直觀來說,x[n,:]就是取第n集合的所有數據, 

X[1,:]即取第一維中下標為1的元素的所有值,輸出結果:

 

 

 

X[:,  m:n],即取所有數據的第m到n-1列數據,含左不含右

例:輸出X數組中所有行第1到2列數據

X = np.array([[0,1,2],[3,4,5],[6,7,8],[9,10,11],[12,13,14],[15,16,17],[18,19,20]])
print X[:,1:3]

 

以下 numpy.random.shuffle() 簡稱 shuffle,numpy.random.permutation() 簡稱 permutation。

shuffle 的參數只能是 array_like,而 permutation 除了 array_like 還可以是 int 類型,如果是 int 類型,那就隨機打亂 numpy.arange(int)。
shuffle 返回 None,這點尤其要注意,也就是說沒有返回值,而 permutation 則返回打亂后的 array。

def permutation(self, object x):
    '''這里都是幫助文檔,我就省略了'''
    if isinstance(x, (int, long, np.integer)):
        arr = np.arange(x)
    else:
        arr = np.array(x)
    self.shuffle(arr)
    return arr

所以permutation內部調用的也是shuffle

1.獲取mini-batch:

def random_mini_batches(X,Y,mini_batch_size=64,seed=0):

 第一,打亂順序

permutation = list(np.random.permutation(m)) #它會返回一個長度為m的隨機數組,且里面的數是0到m-1
shuffled_X = X[:,permutation]   #將每一列的數據按permutation的順序來重新排列。
shuffled_Y = Y[:,permutation].reshape((1,m))

X - 輸入數據,維度為(輸入節點數量,樣本的數量) Y - 對應的是X的標簽,【1 | 0】(藍|紅),維度為(1,樣本的數量)

這一步就相當於對X的維度=1隨機取,等於被打亂了(因為那個位置對應的都是樣本)

第二.分割

 num_complete_minibatches = math.floor(m / mini_batch_size) #把你的訓練集分割成多少份,請注意,如果值是99.99,那么返回值是99,剩下的0.99會被舍棄
    for k in range(0,num_complete_minibatches):
        mini_batch_X = shuffled_X[:,k * mini_batch_size:(k+1)*mini_batch_size]
        mini_batch_Y = shuffled_Y[:,k * mini_batch_size:(k+1)*mini_batch_size]

 

類似於

a = np.array([[1,2,3,4,5,6,7,8,9],
                      [9,8,7,6,5,4,3,2,1],
                      [1,2,3,4,5,6,7,8,9]])
k=1
mini_batch_size=3
print(a[:,1*3:(1+1)*3]) #從第4列到第6列
'''
[[4 5 6]
[6 5 4]
[4 5 6]]
'''
k=2
print(a[:,2*3:(2+1)*3]) #從第7列到第9列
'''
[[7 8 9]
[3 2 1]
[7 8 9]]
'''

包含動量的梯度下降:

這里的V還是指的dw與db的值

def initialize_velocity(parameters):
    """
    初始化速度,velocity是一個字典:
        - keys: "dW1", "db1", ..., "dWL", "dbL" 
        - values:與相應的梯度/參數維度相同的值為零的矩陣。
    參數:
        parameters - 一個字典,包含了以下參數:
            parameters["W" + str(l)] = Wl
            parameters["b" + str(l)] = bl
    返回:
        v - 一個字典變量,包含了以下參數:
            v["dW" + str(l)] = dWl的速度
            v["db" + str(l)] = dbl的速度
    
    """
    L = len(parameters) // 2 #神經網絡的層數
    v = {}
    
    for l in range(L):
        v["dW" + str(l + 1)] = np.zeros_like(parameters["W" + str(l + 1)])
        v["db" + str(l + 1)] = np.zeros_like(parameters["b" + str(l + 1)])
    
    return v

 

 

 

def update_parameters_with_momentun(parameters,grads,v,beta,learning_rate):
    """
    使用動量更新參數
    參數:
        parameters - 一個字典類型的變量,包含了以下字段:
            parameters["W" + str(l)] = Wl
            parameters["b" + str(l)] = bl
        grads - 一個包含梯度值的字典變量,具有以下字段:
            grads["dW" + str(l)] = dWl
            grads["db" + str(l)] = dbl
        v - 包含當前速度的字典變量,具有以下字段:
            v["dW" + str(l)] = ...
            v["db" + str(l)] = ...
        beta - 超參數,動量,實數
        learning_rate - 學習率,實數
    返回:
        parameters - 更新后的參數字典
        v - 包含了更新后的速度變量
    """
    L = len(parameters) // 2 
    for l in range(L):
        #計算速度
        v["dW" + str(l + 1)] = beta * v["dW" + str(l + 1)] + (1 - beta) * grads["dW" + str(l + 1)]
        v["db" + str(l + 1)] = beta * v["db" + str(l + 1)] + (1 - beta) * grads["db" + str(l + 1)]
        
        #更新參數
        parameters["W" + str(l + 1)] = parameters["W" + str(l + 1)] - learning_rate * v["dW" + str(l + 1)]
        parameters["b" + str(l + 1)] = parameters["b" + str(l + 1)] - learning_rate * v["db" + str(l + 1)]
    
    return parameters,v

Adam算法

 

 

 

def initialize_adam(parameters):
    """
    初始化v和s,它們都是字典類型的變量,都包含了以下字段:
        - keys: "dW1", "db1", ..., "dWL", "dbL" 
        - values:與對應的梯度/參數相同維度的值為零的numpy矩陣
    
    參數:
        parameters - 包含了以下參數的字典變量:
            parameters["W" + str(l)] = Wl
            parameters["b" + str(l)] = bl
    返回:
        v - 包含梯度的指數加權平均值,字段如下:
            v["dW" + str(l)] = ...
            v["db" + str(l)] = ...
        s - 包含平方梯度的指數加權平均值,字段如下:
            s["dW" + str(l)] = ...
            s["db" + str(l)] = ...
    
    """
    
    L = len(parameters) // 2
    v = {}
    s = {}
    
    for l in range(L):
        v["dW" + str(l + 1)] = np.zeros_like(parameters["W" + str(l + 1)])
        v["db" + str(l + 1)] = np.zeros_like(parameters["b" + str(l + 1)])
        
        s["dW" + str(l + 1)] = np.zeros_like(parameters["W" + str(l + 1)])
        s["db" + str(l + 1)] = np.zeros_like(parameters["b" + str(l + 1)])
    
    return (v,s)
np.arange(24).reshape(4,6)

 

 

def update_parameters_with_adam(parameters,grads,v,s,t,learning_rate=0.01,beta1=0.9,beta2=0.999,epsilon=1e-8):
    """
    使用Adam更新參數
    
    參數:
        parameters - 包含了以下字段的字典:
            parameters['W' + str(l)] = Wl
            parameters['b' + str(l)] = bl
        grads - 包含了梯度值的字典,有以下key值:
            grads['dW' + str(l)] = dWl
            grads['db' + str(l)] = dbl
        v - Adam的變量,第一個梯度的移動平均值,是一個字典類型的變量
        s - Adam的變量,平方梯度的移動平均值,是一個字典類型的變量
        t - 當前迭代的次數
        learning_rate - 學習率
        beta1 - 動量,超參數,用於第一階段,使得曲線的Y值不從0開始(參見天氣數據的那個圖)
        beta2 - RMSprop的一個參數,超參數
        epsilon - 防止除零操作(分母為0)
    
    返回:
        parameters - 更新后的參數
        v - 第一個梯度的移動平均值,是一個字典類型的變量
        s - 平方梯度的移動平均值,是一個字典類型的變量
    """
    L = len(parameters) // 2
    v_corrected = {} #偏差修正后的值
    s_corrected = {} #偏差修正后的值
    
    for l in range(L):
        #梯度的移動平均值,輸入:"v , grads , beta1",輸出:" v "
        v["dW" + str(l + 1)] = beta1 * v["dW" + str(l + 1)] + (1 - beta1) * grads["dW" + str(l + 1)]
        v["db" + str(l + 1)] = beta1 * v["db" + str(l + 1)] + (1 - beta1) * grads["db" + str(l + 1)]
        
        #計算第一階段的偏差修正后的估計值,輸入"v , beta1 , t" , 輸出:"v_corrected"
        v_corrected["dW" + str(l + 1)] = v["dW" + str(l + 1)] / (1 - np.power(beta1,t))
        v_corrected["db" + str(l + 1)] = v["db" + str(l + 1)] / (1 - np.power(beta1,t))
    
        #計算平方梯度的移動平均值,輸入:"s, grads , beta2",輸出:"s"
        s["dW" + str(l + 1)] = beta2 * s["dW" + str(l + 1)] + (1 - beta2) * np.square(grads["dW" + str(l + 1)])
        s["db" + str(l + 1)] = beta2 * s["db" + str(l + 1)] + (1 - beta2) * np.square(grads["db" + str(l + 1)])
         
        #計算第二階段的偏差修正后的估計值,輸入:"s , beta2 , t",輸出:"s_corrected"
        s_corrected["dW" + str(l + 1)] = s["dW" + str(l + 1)] / (1 - np.power(beta2,t))
        s_corrected["db" + str(l + 1)] = s["db" + str(l + 1)] / (1 - np.power(beta2,t))
        
        #更新參數,輸入: "parameters, learning_rate, v_corrected, s_corrected, epsilon". 輸出: "parameters".
        parameters["W" + str(l + 1)] = parameters["W" + str(l + 1)] - learning_rate * (v_corrected["dW" + str(l + 1)] / np.sqrt(s_corrected["dW" + str(l + 1)] + epsilon))
        parameters["b" + str(l + 1)] = parameters["b" + str(l + 1)] - learning_rate * (v_corrected["db" + str(l + 1)] / np.sqrt(s_corrected["db" + str(l + 1)] + epsilon))
    
    return (parameters,v,s)

 

 

現在用類似之前的模型來寫個程序

def model(X,Y,layers_dims,optimizer,learning_rate=0.0007,
          mini_batch_size=64,beta=0.9,beta1=0.9,beta2=0.999,
          epsilon=1e-8,num_epochs=10000,print_cost=True,is_plot=True):
    
    """
    可以運行在不同優化器模式下的3層神經網絡模型。
    
    參數:
        X - 輸入數據,維度為(2,輸入的數據集里面樣本數量)
        Y - 與X對應的標簽
        layers_dims - 包含層數和節點數量的列表
        optimizer - 字符串類型的參數,用於選擇優化類型,【 "gd" | "momentum" | "adam" 】
        learning_rate - 學習率
        mini_batch_size - 每個小批量數據集的大小
        beta - 用於動量優化的一個超參數
        beta1 - 用於計算梯度后的指數衰減的估計的超參數
        beta1 - 用於計算平方梯度后的指數衰減的估計的超參數
        epsilon - 用於在Adam中避免除零操作的超參數,一般不更改
        num_epochs - 整個訓練集的遍歷次數,(視頻2.9學習率衰減,1分55秒處,視頻中稱作“代”),相當於之前的num_iteration
        print_cost - 是否打印誤差值,每遍歷1000次數據集打印一次,但是每100次記錄一個誤差值,又稱每1000代打印一次
        is_plot - 是否繪制出曲線圖
        
    返回:
        parameters - 包含了學習后的參數
        
    """
    L = len(layers_dims)
    costs = []
    t = 0 #每學習完一個minibatch就增加1
    seed = 10 #隨機種子
    
    #初始化參數
    parameters = opt_utils.initialize_parameters(layers_dims)
    
    #選擇優化器  注意注意,重點操作!優化用的是啥就指望這個呢!這里是初始化V與S
if optimizer == "gd": pass #不使用任何優化器,直接使用梯度下降法 elif optimizer == "momentum": v = initialize_velocity(parameters) #使用動量 elif optimizer == "adam": v, s = initialize_adam(parameters)#使用Adam優化 else: print("optimizer參數錯誤,程序退出。") exit(1) #開始學習 for i in range(num_epochs): #定義隨機 minibatches,我們在每次遍歷數據集之后增加種子以重新排列數據集,使每次數據的順序都不同 seed = seed + 1 minibatches = random_mini_batches(X,Y,mini_batch_size,seed) for minibatch in minibatches: #選擇一個minibatch (minibatch_X,minibatch_Y) = minibatch #前向傳播 A3 , cache = opt_utils.forward_propagation(minibatch_X,parameters) #計算誤差 cost = opt_utils.compute_cost(A3 , minibatch_Y) #反向傳播 grads = opt_utils.backward_propagation(minibatch_X,minibatch_Y,cache) #更新參數 注意注意 重點操作第二步,這里頭是更新不同的參數 if optimizer == "gd": parameters = update_parameters_with_gd(parameters,grads,learning_rate) elif optimizer == "momentum": parameters, v = update_parameters_with_momentun(parameters,grads,v,beta,learning_rate) elif optimizer == "adam": t = t + 1 #這個t是用來修正v和s的corrected的 parameters , v , s = update_parameters_with_adam(parameters,grads,v,s,t,learning_rate,beta1,beta2,epsilon) #記錄誤差值 if i % 100 == 0: costs.append(cost) #是否打印誤差值 if print_cost and i % 1000 == 0: print("" + str(i) + "次遍歷整個數據集,當前誤差值:" + str(cost)) #是否繪制曲線圖 if is_plot: plt.plot(costs) plt.ylabel('cost') plt.xlabel('epochs (per 100)') plt.title("Learning rate = " + str(learning_rate)) plt.show() return parameters

普通

#使用普通的梯度下降
layers_dims = [train_X.shape[0],5,2,1]
parameters = model(train_X, train_Y, layers_dims, optimizer="gd",is_plot=True)

adam優化后

layers_dims = [train_X.shape[0], 5, 2, 1]
#使用Adam優化的梯度下降
parameters = model(train_X, train_Y, layers_dims, optimizer="adam",is_plot=True)

預測(這里一直不是很懂,回頭再看看opt_utils的代碼研究研究)

#預測
preditions = opt_utils.predict(train_X,train_Y,parameters)

#繪制分類圖
plt.title("Model with Adam optimization")
axes = plt.gca()
axes.set_xlim([-1.5, 2.5])
axes.set_ylim([-1, 1.5])
opt_utils.plot_decision_boundary(lambda x: opt_utils.predict_dec(parameters, x.T), train_X, train_Y)

 

 

 

 

opt_utils的代碼

  1 # -*- coding: utf-8 -*-
  2 
  3 #opt_utils.py
  4 
  5 import numpy as np
  6 import matplotlib.pyplot as plt
  7 import sklearn
  8 import sklearn.datasets
  9 
 10 def sigmoid(x):
 11     """
 12     Compute the sigmoid of x
 13  
 14     Arguments:
 15     x -- A scalar or numpy array of any size.
 16  
 17     Return:
 18     s -- sigmoid(x)
 19     """
 20     s = 1/(1+np.exp(-x))
 21     return s
 22  
 23 def relu(x):
 24     """
 25     Compute the relu of x
 26  
 27     Arguments:
 28     x -- A scalar or numpy array of any size.
 29  
 30     Return:
 31     s -- relu(x)
 32     """
 33     s = np.maximum(0,x)
 34     
 35     return s
 36 
 37 
 38 def load_params_and_grads(seed=1):
 39     np.random.seed(seed)
 40     W1 = np.random.randn(2,3)
 41     b1 = np.random.randn(2,1)
 42     W2 = np.random.randn(3,3)
 43     b2 = np.random.randn(3,1)
 44  
 45     dW1 = np.random.randn(2,3)
 46     db1 = np.random.randn(2,1)
 47     dW2 = np.random.randn(3,3)
 48     db2 = np.random.randn(3,1)
 49     
 50     return W1, b1, W2, b2, dW1, db1, dW2, db2
 51     
 52 def initialize_parameters(layer_dims):
 53     """
 54     Arguments:
 55     layer_dims -- python array (list) containing the dimensions of each layer in our network
 56     
 57     Returns:
 58     parameters -- python dictionary containing your parameters "W1", "b1", ..., "WL", "bL":
 59                     W1 -- weight matrix of shape (layer_dims[l], layer_dims[l-1])
 60                     b1 -- bias vector of shape (layer_dims[l], 1)
 61                     Wl -- weight matrix of shape (layer_dims[l-1], layer_dims[l])
 62                     bl -- bias vector of shape (1, layer_dims[l])
 63                     
 64     Tips:
 65     - For example: the layer_dims for the "Planar Data classification model" would have been [2,2,1]. 
 66     This means W1's shape was (2,2), b1 was (1,2), W2 was (2,1) and b2 was (1,1). Now you have to generalize it!
 67     - In the for loop, use parameters['W' + str(l)] to access Wl, where l is the iterative integer.
 68     """
 69     
 70     np.random.seed(3)
 71     parameters = {}
 72     L = len(layer_dims) # number of layers in the network
 73  
 74     for l in range(1, L):
 75         parameters['W' + str(l)] = np.random.randn(layer_dims[l], layer_dims[l-1])*  np.sqrt(2 / layer_dims[l-1])
 76         parameters['b' + str(l)] = np.zeros((layer_dims[l], 1))
 77         
 78         assert(parameters['W' + str(l)].shape == layer_dims[l], layer_dims[l-1])
 79         assert(parameters['W' + str(l)].shape == layer_dims[l], 1)
 80         
 81     return parameters
 82     
 83 def forward_propagation(X, parameters):
 84     """
 85     Implements the forward propagation (and computes the loss) presented in Figure 2.
 86     
 87     Arguments:
 88     X -- input dataset, of shape (input size, number of examples)
 89     parameters -- python dictionary containing your parameters "W1", "b1", "W2", "b2", "W3", "b3":
 90                     W1 -- weight matrix of shape ()
 91                     b1 -- bias vector of shape ()
 92                     W2 -- weight matrix of shape ()
 93                     b2 -- bias vector of shape ()
 94                     W3 -- weight matrix of shape ()
 95                     b3 -- bias vector of shape ()
 96     
 97     Returns:
 98     loss -- the loss function (vanilla logistic loss)
 99     """
100     
101     # retrieve parameters
102     W1 = parameters["W1"]
103     b1 = parameters["b1"]
104     W2 = parameters["W2"]
105     b2 = parameters["b2"]
106     W3 = parameters["W3"]
107     b3 = parameters["b3"]
108     
109     # LINEAR -> RELU -> LINEAR -> RELU -> LINEAR -> SIGMOID
110     z1 = np.dot(W1, X) + b1
111     a1 = relu(z1)
112     z2 = np.dot(W2, a1) + b2
113     a2 = relu(z2)
114     z3 = np.dot(W3, a2) + b3
115     a3 = sigmoid(z3)
116     
117     cache = (z1, a1, W1, b1, z2, a2, W2, b2, z3, a3, W3, b3)
118     
119     return a3, cache
120  
121 def backward_propagation(X, Y, cache):
122     """
123     Implement the backward propagation presented in figure 2.
124     
125     Arguments:
126     X -- input dataset, of shape (input size, number of examples)
127     Y -- true "label" vector (containing 0 if cat, 1 if non-cat)
128     cache -- cache output from forward_propagation()
129     
130     Returns:
131     gradients -- A dictionary with the gradients with respect to each parameter, activation and pre-activation variables
132     """
133     m = X.shape[1]
134     (z1, a1, W1, b1, z2, a2, W2, b2, z3, a3, W3, b3) = cache
135     
136     dz3 = 1./m * (a3 - Y)
137     dW3 = np.dot(dz3, a2.T)
138     db3 = np.sum(dz3, axis=1, keepdims = True)
139     
140     da2 = np.dot(W3.T, dz3)
141     dz2 = np.multiply(da2, np.int64(a2 > 0))
142     dW2 = np.dot(dz2, a1.T)
143     db2 = np.sum(dz2, axis=1, keepdims = True)
144     
145     da1 = np.dot(W2.T, dz2)
146     dz1 = np.multiply(da1, np.int64(a1 > 0))
147     dW1 = np.dot(dz1, X.T)
148     db1 = np.sum(dz1, axis=1, keepdims = True)
149     
150     gradients = {"dz3": dz3, "dW3": dW3, "db3": db3,
151                  "da2": da2, "dz2": dz2, "dW2": dW2, "db2": db2,
152                  "da1": da1, "dz1": dz1, "dW1": dW1, "db1": db1}
153     
154     return gradients
155  
156 def compute_cost(a3, Y):
157     
158     """
159     Implement the cost function
160     
161     Arguments:
162     a3 -- post-activation, output of forward propagation
163     Y -- "true" labels vector, same shape as a3
164     
165     Returns:
166     cost - value of the cost function
167     """
168     m = Y.shape[1]
169     
170     logprobs = np.multiply(-np.log(a3),Y) + np.multiply(-np.log(1 - a3), 1 - Y)
171     cost = 1./m * np.sum(logprobs)
172     
173     return cost
174  
175 def predict(X, y, parameters):
176     """
177     This function is used to predict the results of a  n-layer neural network.
178     
179     Arguments:
180     X -- data set of examples you would like to label
181     parameters -- parameters of the trained model
182     
183     Returns:
184     p -- predictions for the given dataset X
185     """
186     
187     m = X.shape[1]
188     p = np.zeros((1,m), dtype = np.int)
189     
190     # Forward propagation
191     a3, caches = forward_propagation(X, parameters)
192     
193     # convert probas to 0/1 predictions
194     for i in range(0, a3.shape[1]):
195         if a3[0,i] > 0.5:
196             p[0,i] = 1
197         else:
198             p[0,i] = 0
199  
200     # print results
201  
202     #print ("predictions: " + str(p[0,:]))
203     #print ("true labels: " + str(y[0,:]))
204     print("Accuracy: "  + str(np.mean((p[0,:] == y[0,:]))))
205     
206     return p
207  
208 def predict_dec(parameters, X):
209     """
210     Used for plotting decision boundary.
211     
212     Arguments:
213     parameters -- python dictionary containing your parameters 
214     X -- input data of size (m, K)
215     
216     Returns
217     predictions -- vector of predictions of our model (red: 0 / blue: 1)
218     """
219     
220     # Predict using forward propagation and a classification threshold of 0.5
221     a3, cache = forward_propagation(X, parameters)
222     predictions = (a3 > 0.5)
223     return predictions
224  
225 def plot_decision_boundary(model, X, y):
226     # Set min and max values and give it some padding
227     x_min, x_max = X[0, :].min() - 1, X[0, :].max() + 1
228     y_min, y_max = X[1, :].min() - 1, X[1, :].max() + 1
229     h = 0.01
230     # Generate a grid of points with distance h between them
231     xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
232     # Predict the function value for the whole grid
233     Z = model(np.c_[xx.ravel(), yy.ravel()])
234     Z = Z.reshape(xx.shape)
235     # Plot the contour and training examples
236     plt.contourf(xx, yy, Z, cmap=plt.cm.Spectral)
237     plt.ylabel('x2')
238     plt.xlabel('x1')
239     plt.scatter(X[0, :], X[1, :], c=y, cmap=plt.cm.Spectral)
240     plt.show()
241  
242 def load_dataset(is_plot = True):
243     np.random.seed(3)
244     train_X, train_Y = sklearn.datasets.make_moons(n_samples=300, noise=.2) #300 #0.2 
245     # Visualize the data
246     if is_plot:
247         plt.scatter(train_X[:, 0], train_X[:, 1], c=train_Y, s=40, cmap=plt.cm.Spectral);
248     train_X = train_X.T
249     train_Y = train_Y.reshape((1, train_Y.shape[0]))
250     
251     return train_X, train_Y
opt_utils詳細代碼

 


免責聲明!

本站轉載的文章為個人學習借鑒使用,本站對版權不負任何法律責任。如果侵犯了您的隱私權益,請聯系本站郵箱yoyou2525@163.com刪除。



 
粵ICP備18138465號   © 2018-2025 CODEPRJ.COM