# Checking convergence of 2-layer neural network in python

I am working with the following code:

import numpy as np

def sigmoid(x):
return 1.0/(1.0 + np.exp(-x))

def sigmoid_prime(x):
return sigmoid(x)*(1.0-sigmoid(x))

def tanh(x):
return np.tanh(x)

def tanh_prime(x):
return 1.0 - x**2

class NeuralNetwork:

def __init__(self, layers, activation='tanh'):
if activation == 'sigmoid':
self.activation = sigmoid
self.activation_prime = sigmoid_prime
elif activation == 'tanh':
self.activation = tanh
self.activation_prime = tanh_prime

# Set weights
self.weights = []
# layers = [2,2,1]
# range of weight values (-1,1)
# input and hidden layers - random((2+1, 2+1)) : 3 x 3
for i in range(1, len(layers) - 1):
r = 2*np.random.random((layers[i-1] + 1, layers[i] + 1)) -1
self.weights.append(r)
# output layer - random((2+1, 1)) : 3 x 1
r = 2*np.random.random( (layers[i] + 1, layers[i+1])) - 1
self.weights.append(r)

def fit(self, X, y, learning_rate=0.2, epochs=100000):
# Add column of ones to X
# This is to add the bias unit to the input layer
ones = np.atleast_2d(np.ones(X.shape[0]))
X = np.concatenate((ones.T, X), axis=1)

for k in range(epochs):
if k % 10000 == 0: print 'epochs:', k

i = np.random.randint(X.shape[0])
a = [X[i]]

for l in range(len(self.weights)):
dot_value = np.dot(a[l], self.weights[l])
activation = self.activation(dot_value)
a.append(activation)
# output layer
error = y[i] - a[-1]
deltas = [error * self.activation_prime(a[-1])]

# we need to begin at the second to last layer
# (a layer before the output layer)
for l in range(len(a) - 2, 0, -1):
deltas.append(deltas[-1].dot(self.weights[l].T)*self.activation_prime(a[l]))

# reverse
# [level3(output)->level2(hidden)]  => [level2(hidden)->level3(output)]
deltas.reverse()

# backpropagation
# 1. Multiply its output delta and input activation
#    to get the gradient of the weight.
# 2. Subtract a ratio (percentage) of the gradient from the weight.
for i in range(len(self.weights)):
layer = np.atleast_2d(a[i])
delta = np.atleast_2d(deltas[i])
self.weights[i] += learning_rate * layer.T.dot(delta)

def predict(self, x):
a = np.concatenate((np.ones(1).T, np.array(x)), axis=0)
for l in range(0, len(self.weights)):
a = self.activation(np.dot(a, self.weights[l]))
return a

if __name__ == '__main__':

nn = NeuralNetwork([2,2,1])

X = np.array([[0, 0],
[0, 1],
[1, 0],
[1, 1]])

y = np.array([0, 1, 1, 0])

nn.fit(X, y)

for e in X:
print(e,nn.predict(e))


While this converges well and fast when using the tanh, it does converge much slower when using the sigmoid ( in def __init__(self, layers, activation='tanh') change tanh to sigmoid ). I cannot find why that is. How do I improve the implementation for the sigmoid?

• Are you interested in reviews on aspects of the code unrelated to the sigmoid implementation? Commented May 4, 2016 at 19:14
• Of course I am!
– user
Commented May 4, 2016 at 19:21

## The reasons for the speed discrepancy

1. The reason for the differences in timing are because evaluating sigmoid_prime() takes far longer than tanh_prime(). You can see this if you use a line profiler such as the line_profiler module.

2. Is tanh_prime() supposed to be the derivative of tanh()? If so, you might want to double-check your formula. The derivative of tanh(x) is 1. - tanh(x)**2, not 1. - x**2.

3. In fact, if you use the the actual definition of the derivative of tanh(), the timings become much more similar.

def tanh_prime_alt(x):
return 1 - tanh(x)**2

foo = np.random.rand(10000)
%timeit -n 100 tanh_prime(foo)
%timeit -n 100 tanh_prime_alt(foo)
%timeit -n 100 sigmoid_prime(foo)

100 loops, best of 3: 10.2 µs per loop
100 loops, best of 3: 116 µs per loop
100 loops, best of 3: 279 µs per loop


So with this alternate tanh_prime(), the sigmoid method is now only 2× slower, not 20× slower. I should emphasize that (a) I don't know enough about neural networks to know if 1. - x**2 is an appropriate expression or approximation to the actual derivative of tanh(), but if it is in fact OK, then (b) the reason that activation = 'tanh' is so much faster is because of this approximation/error.

4. The remaining 2× difference is because in your factored expression of sigmoid_prime(), you are needlessly evaluating sigmoid() twice. I'd instead do this:

def sigmoid_prime_alt(x):
sig_x = sigmoid(x)
return sig_x - sig_x**2


As expected, this speeds things up two-fold relative to your original definition.

foo = np.random.rand(10000)
%timeit -n 100 sigmoid_prime(foo)
%timeit -n 100 sigmoid_prime_alt(foo)

100 loops, best of 3: 248 µs per loop
100 loops, best of 3: 132 µs per loop

5. Since the sigmoid() function and the tanh() function are related by tanh(x) = 2 * sigmoid(2*x) - 1, i.e. sigmoid(x) = (1 + tanh(x/2.))/2, then iff you are OK with the weird 1 - x**2 approximation for tanh_prime(), you should be able to work out a similar approximation for sigmoid_prime().

6. You might be interested in the autograd module, which provides a generalized capability to compute symbolic derivatives of most NumPy code.

1. Why are your weights Python lists instead of NumPy arrays? If you're already using NumPy, you might as well use it wherever you can.
2. You probably don't need the for l in range(len(self.weights)): loop, do you? Can't you use NumPy array slicing and the matrix capabilities of np.dot() to replace this loop?
3. If you are going to loop, you don't need to do for l in range(len(self.weights)) and then reference self.weights[l]. You can do for weight in self.weights: and then reference weight in your loop code, for example.