Stochastic gradient descent example

astrofrog.py

import numpy as np

N = 5000
I_true = np.random.randn(N)
D = I_true[:, None] - I_true[None, :]

I0 = np.zeros_like(I_true)

eta = 1.0 / N  # Learning rate.
tol = 1.25e-11  # Error tolerance.

inds = np.arange(N)
for n in range(500):
    np.random.shuffle(inds)
    for i in inds:
        I0[i] += eta * np.sum(D[i, :] - I0[i, None] + I0[None, :])
    error = np.sum((D - I0[:, None] + I0[None, :]) ** 2) / N
    print error
    if error < tol:
        break

print np.sum((I0 - I_true) ** 2) / N