def logistic_gradient(X, Y, gd_lambda, descent=True, epsilon_accepted=1e-6, max_iterations=10000000):
accepted = False
iterations = 0
epsilon = 1
X['b'] = np.ones(len(X))
m = X.shape[1] # number of cols
print 'sh0: {} len(X): {}'.format(m, len(X))
w_old = np.zeros(m)
while not accepted:
w_new = np.zeros(w_old.shape)
for j in range(len(w_old)): # by col
delta = 0.0
for i in range(len(X)): # by row
delta += (Y.values[i] - sigmoid(np.dot(w_old, X.values[i]))) * X.values[i, j]
w_new[j] = w_old[j] + gd_lambda * delta
if np.any(np.isnan(w_new)):
raise ValueError('NAN is found on iteration {}'.format(iterations))
epsilon = sum(np.abs(w_new - w_old))/len(w_new)
print 'epsilon: {}'.format(epsilon)
print 'w:'
print '{} iterations, w: {}'.format(iterations, w_new[:])
w_old = w_new
if epsilon < epsilon_accepted:
accepted = True
if iterations >= max_iterations:
accepted = True
iterations += 1
return w_new
def predict(df, model, binary=False, logistic=False):
if 'b' not in df.columns:
df['b'] = 1
if binary:
cutoff = .5
if binary and logistic:
predictions = [sigmoid(x) for x in np.dot(df, model)]
else:
predictions = np.dot(df, model)
if binary:
for p in range(len(predictions)):
if predictions[p] < cutoff:
predictions[p] = 0
else:
predictions[p] = 1
return predictions
def testLogGradient2():
X = np.random.random(size=[10, 2])
y = utils.sigmoid(X[:, 0]* .5 + 2 * X[:, 1] + 3)
df = pd.DataFrame(data=X)
w = gd.logistic_gradient(df, y, .05)
print w
