Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Welcome To Ask or Share your Answers For Others

Categories

0 votes
400 views
in Technique[技术] by (71.8m points)

python - Problems With ANN BackProp/Gradient Checking.

Just wrote up my first Neural Network Class in python. Everything as far as I can tell should work, but there is some bug in it that I can't seem to find(Probably staring me right in the face). I first tried it on 10,000 examples of the MNIST data, then again when trying to replicate the sign function, and again when trying to replicate a XOR Gate. Every time, regardless of the # of epochs, it always produces output from all the output neurons(regardless of how many there may be) that are all roughly the same value, but the cost function seems to be going down. I am using batch gradient descent, all done using vectors(no loop for each training example).

#Neural Network Class

import numpy as np



class NeuralNetwork:

#methods
def __init__(self,layer_shape):
    #Useful Network Info
    self.__layer_shape = layer_shape
    self.__layers = len(layer_shape)

    #Initialize Random Weights
    self.__weights = [] 
    self.__weight_sizes = []
    for i in range(len(layer_shape)-1):
        current_weight_size = (layer_shape[i+1],layer_shape[i]+1)
        self.__weight_sizes.append(current_weight_size)
        self.__weights.append(np.random.normal(loc=0.1,scale=0.1,size=current_weight_size))

def sigmoid(self,z):
    return (1/(1+np.exp(-z)))

def sig_prime(self,z):
    return np.multiply(self.sigmoid(z),(1-self.sigmoid(z)))


def Feedforward(self,input,Train=False):
    self.__input_cases = np.shape(input)[0]

    #Empty list to hold the output of every layer.
    output_list = []
    #Appends the output of the the 1st input layer.
    output_list.append(input)

    for i in range(self.__layers-1):
        if i == 0:
            output = self.sigmoid(np.dot(np.concatenate((np.ones((self.__input_cases,1)),input),1),self.__weights[0].T))
            output_list.append(output)
        else:
            output = self.sigmoid(np.dot(np.concatenate((np.ones((self.__input_cases,1)),output),1),self.__weights[i].T))                 
            output_list.append(output)

    #Returns the final output if not training.         
    if Train == False:
        return output_list[-1]
    #Returns the entire output_list if need for training
    else:
        return output_list

def CostFunction(self,input,target,error_func=1):
    """Gives the cost of using a particular weight matrix 
    based off of the input and targeted output"""

    #Run the network to get output using current theta matrices.
    output = self.Feedforward(input)


    #####Allows user to choose Cost Functions.##### 

    #
    #Log Based Error Function
    #
    if error_func == 0:
        error = np.multiply(-target,np.log(output))-np.multiply((1-target),np.log(1-output))
        total_error = np.sum(np.sum(error))
    #    
    #Squared Error Cost Function
    #
    elif error_func == 1:
        error = (target - output)**2
        total_error = 0.5 * np.sum(np.sum(error))

    return total_error

def Weight_Grad(self,input,target,output_list):

            #Finds the Error Deltas for Each Layer
            # 
            deltas = []
            for i in range(self.__layers - 1):
                #Finds Error Delta for the last layer
                if i == 0:

                    error = (target-output_list[-1])

                    error_delta = -1*np.multiply(error,np.multiply(output_list[-1],(1-output_list[-1])))
                    deltas.append(error_delta)
                #Finds Error Delta for the hidden layers   
                else:
                    #Weight matrices have bias values removed
                    error_delta = np.multiply(np.dot(deltas[-1],self.__weights[-i][:,1:]),output_list[-i-1]*(1-output_list[-i-1]))
                    deltas.append(error_delta)

            #
            #Finds the Deltas for each Weight Matrix
            #
            Weight_Delta_List = []
            deltas.reverse()
            for i in range(len(self.__weights)):

                current_weight_delta = (1/self.__input_cases) * np.dot(deltas[i].T,np.concatenate((np.ones((self.__input_cases,1)),output_list[i]),1))
                Weight_Delta_List.append(current_weight_delta)
                #print("Weight",i,"Delta:","
",current_weight_delta)
                #print()

            #
            #Combines all Weight Deltas into a single row vector
            #
            Weight_Delta_Vector = np.array([[]])
            for i in Weight_Delta_List:

                Weight_Delta_Vector = np.concatenate((Weight_Delta_Vector,np.reshape(i,(1,-1))),1)
            return Weight_Delta_List        

def Train(self,input_data,target):
    #
    #Gradient Checking:
    #

    #First Get Gradients from first iteration of Back Propagation 
    output_list = self.Feedforward(input_data,Train=True)
    self.__input_cases = np.shape(input_data)[0]

    Weight_Delta_List = self.Weight_Grad(input_data,target,output_list)  

    #Creates List of Gradient Approx arrays set to zero.
    grad_approx_list = []
    for i in self.__weight_sizes:
        current_grad_approx = np.zeros(i)
        grad_approx_list.append(current_grad_approx)


    #Compute Approx. Gradient for every Weight Change
    for W in range(len(self.__weights)):
        for index,value in np.ndenumerate(self.__weights[W]):
            orig_value = self.__weights[W][index]      #Saves the Original Value
            print("Orig Value:", orig_value)

            #Sets weight to  weight +/- epsilon
            self.__weights[W][index] = orig_value+.00001
            cost_plusE = self.CostFunction(input_data, target)

            self.__weights[W][index] = orig_value-.00001
            cost_minusE = self.CostFunction(input_data, target)

            #Solves for grad approx:
            grad_approx = (cost_plusE-cost_minusE)/(2*.00001)
            grad_approx_list[W][index] = grad_approx

            #Sets Weight Value back to its original value
            self.__weights[W][index] = orig_value


    #
    #Print Gradients from Back Prop. and Grad Approx. side-by-side:
    #

    print("Back Prop. Grad","","Grad. Approx")
    print("-"*15,"","-"*15)
    for W in range(len(self.__weights)):
        for index, value in np.ndenumerate(self.__weights[W]):
            print(self.__weights[W][index],""*3,grad_approx_list[W][index])

    print("
"*3)
    input_ = input("Press Enter to continue:")


    #
    #Perform Weight Updates for X number of Iterations
    #
    for i in range(10000):
    #Run the network
        output_list = self.Feedforward(input_data,Train=True)
        self.__input_cases = np.shape(input_data)[0]

        Weight_Delta_List = self.Weight_Grad(input_data,target,output_list)


        for w in range(len(self.__weights)):
            #print(self.__weights[w])
            #print(Weight_Delta_List[w])
            self.__weights[w] = self.__weights[w] - (.01*Weight_Delta_List[w]) 


    print("Done")`

I even implememented Gradient Checking and the values are different, and I thought I would try replacing the Back Propagation updates with the Approx. Gradient Checking values, but that gave the same results, causing me to doubt even my Gradient Checking code.

Here are some of the values being produced when training for the XOR Gate:

Back Prop. Grad: 0.0756102610697 0.261814503398 0.0292734023876 Grad Approx: 0.05302210631166 0.0416095559674 0.0246847342122 Cost: Before Training: 0.508019225507 After Training 0.50007095103 (After 10000 Epochs) Output for 4 different examples(after training): [ 0.49317733] [ 0.49294556] [ 0.50489004] [ 0.50465824]

So my question is, is there any obvious problem with my Back Propagation, or my gradient checking? Are there any usual problems when a ANN shows these symptoms(Outputs are all roughly the same/Cost is going down)?

See Question&Answers more detail:os

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome To Ask or Share your Answers For Others

1 Reply

0 votes
by (71.8m points)

I'm not very proficient at reading python code, but your gradient list for XOR contains 3 elements, corresponding for 3 weights. I assume, that these are two inputs and one bias for a single neuron. If true, such network can not learn XOR (minimun NN that can learn XOR need two hidden neurons and one output unit). Now, looking at Feedforward function, if np.dot computes what it name says (i.e dot product of two vectors), and sigmoid is scalar, then this will always correspond to output of one neuron and I don't see the way how you can add more neurons to the layers with this code.

Following advice could be useful to debug any newly implemented NN:

1) Don't start with MNIST or even XOR. Perfectly good implementation may fail to learn XOR because it can easily fell into local minima and you could spent a lot of time hunting for non-existent error. A good starting point will be AND function, that can be learned with single neuron

2) Check forward computation pass by manually computing results on few examples. thats easy to do with small number of weights. Then try to train it with numerical gradient. If it fails, then either your numerical gradient is wrong (check that by hand) or training procedure is wrong. (it can fail to work if you set too large learning rate, but otherwise training must converge since error surface is convex).

3) once you can train it with numerical grad, debug your analytical gradients (check gradient per neuron, and then gradient for individual weights). That again can be computed manually and compared to what you see.

4) Upon completion of step 3, if everything works OK, add one hidden layer and repeat steps 2 and 3 with AND function.

5) after everything works with AND, you can move to XOR function and other more complicated tasks.

This procedure may seems time consuming, but it almost aways results in working NN in the end


与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
OGeek|极客中国-欢迎来到极客的世界,一个免费开放的程序员编程交流平台!开放,进步,分享!让技术改变生活,让极客改变未来! Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Click Here to Ask a Question

...