CODEWITHSUNDEEP
python
Dzung Nguyen
Dzung Nguyen

Reputation: 3944

How to use class methods as callable function

I'm running a gradient checking in Python as following:

class NeuralNetwork:
    def gradient_checking(self):
        m = 5
        theta1 = debuginitializeweights(self.size[1], self.size[0])
        theta2 = debuginitializeweights(self.size[2], self.size[1])
        thetas = vectorize([theta1, theta2], self.size)

        X = debuginitializeweights(m, self.size[0] - 1)
        y = 1 + np.mod(np.array(range(m)), 3)

        return scipy.optimize.check_grad(self.cost, self.grad, thetas, [X, y])

where the signature of class methods are:

def cost(self, thetas, X, label):   
def grad(self, thetas, X, label):

However when running the gradient checking it is saying that 

File "/home/andrey/Data/Hacking/ML/Coursera_Andrew/neuralnetwork.py", line 142, in gradient_checking
    return check_grad(self.cost, self.grad, thetas, [X, y])
  File "/usr/local/lib/python2.7/dist-packages/scipy/optimize/optimize.py", line 656, in check_grad
    return sqrt(sum((grad(x0, *args) -
TypeError: grad() takes exactly 4 arguments (3 given)

How can I fix this error?

Upvotes: 0

Views: 85

Answers (2)

Dzung Nguyen
Dzung Nguyen

Reputation: 3944

Turns out this is simply optional argument problem. Here is a quick example:

 from scipy.optimize import check_grad
class NeuralNetwork:
    def cost(self, x, A, b):
        return x[0] ** 2 - 0.5 * x[1] ** 3 + A * b

    def grad(self, x, A, b):
        return [2 * x[0] + A, -1.5 * x[1]**2 +b]

a = NeuralNetwork()
print a.cost([1.5, -1.5], 10, 1)
print check_grad(a.cost, a.grad, [1.5, -1.5], 10, 1)

Previously I did:

 check_grad(a.cost, a.grad, [1.5, -1.5], (10, 1))

That's why it keeps missing arguments.

Upvotes: 0

skrrgwasme
skrrgwasme

Reputation: 9633

The stack trace tells you exactly what you need to know:

TypeError: grad() takes exactly 4 arguments (3 given)

Your grad signature reflects the 4 argument requirement:

def grad(self, thetas, X, label):

I see you're attempting to use tuple unpacking in your call to grad():

return sqrt(sum((grad(x0, *args))))

The implicit self being passed when grad() is called will take the self position in the argument list, x0 will take the thetas position, leaving X and label to be filled by *args. Try printing args or examining it with PDB to confirm that it contains two items. Since you're not getting a ValueError from trying to unpack a non-iterable, it probably is of the right type. It sounds like it probably doesn't have both items in it that you expected.

Upvotes: 1

Related Questions