Using relu activation function destroys the model

Question

I have implemented a neural network with 1 hidden layer for classification . It uses sigmoid activation function and cross entropy loss . But while watching a cs231n lecture , I came across relu activation function which converges much faster. So, I used relu activation for hidden layer but the accuracy has drastically decreased to 30-40% from more than 90% . Previously I struggled with relu because cost function was always tending to infinity because output of relu can be 0. I fixed that by always adding a small value in the log.

Following are the most important code snippets that I have modified from previous version where I had used sigmoid activation. I could not highlight the part that I had changed so I have added a #changed comment. I will the entire code if someone wants to take a closer look.

snippets:

activation functions:

def relu(arg):  #I have tried both relu and leaky relu
    return 1*(arg<0)*0.0001*arg + (arg>=0)*arg

def reluGrad(arg):
    for i in range(arg.shape[0]):
        for j in range(arg.shape[1]):
            if arg[i][j]>0:
                arg[i][j]=1
            else:
                arg[i][j]=0
    return arg
def softmax(x):
    x = x.transpose()
    e_x = np.exp(x - np.max(x))
    return (e_x / e_x.sum(axis=0)).transpose()

forward prop:

a1 = np.insert(data,0,np.ones(len(data)),1).astype(np.float64)
    z2 = a1.dot(theta1)
    a2 = relu(z2) #changed
    a2 = np.insert(a2,0,np.ones(len(a2)),1)
    z3 = a2.dot(theta2)
    a3 = softmax(z3) #changed

compute the cost:

cost = -(output*(np.log(a3))+(1-output)*(np.log(1-a3))).sum()
cost = (1/len(data))*cost + (lamb/(2*len(data)))*((np.delete(theta1,0,0)**2).sum() + (np.delete(theta2,0,0)**2).sum())

backProp:

sigma3 = a3-output
sigma2 = (sigma3.dot(np.transpose(theta2)))* reluGrad(np.insert(z2,0,np.ones(len(z2)),1)) #changed
sigma2 = np.delete(sigma2,0,1)
delta2 = (np.transpose(a2)).dot(sigma3)
delta1 = (np.transpose(a1)).dot(sigma2)

grad1 = delta1/len(data) + (lamb/len(data))*np.insert(np.delete(theta1,0,0),0,np.zeros(len(theta1[0])),0)
grad2 = delta2/len(data) + (lamb/len(data))*np.insert(np.delete(theta2,0,0),0,np.zeros(len(theta2[0])),0)

#update theta

theta1 = theta1 - alpha*grad1
theta2 = theta2 - alpha*grad2

Why does the accuracy decrease ? What's wrong in this implementation with relu function ?

Using relu activation function destroys the model

0 Answers0