def __init__(self, my_lambda=0, file_name="dump.txt"):
super().__init__()
statistic = Statistic(output=file_name)
x, y, self.m = statistic.get_training_set()
self.X = x[:, :]
self.Y = y[:]
self.m = len(self.Y)
self.output_layer_size = 3
# reading the data
# data = loadmat('ex4data1.mat')
# self.X = data['X'][0:500, :]
# self.Y = data['y'][0:500, :]
# self.m = len(self.Y)
# self.output_layer_size = 10
self.my_lambda = my_lambda
self.input_layer_size = self.X.shape[1]
self.internal_layer_size = 25
self.theta1 = None
self.theta2 = None
self.w1 = None
self.w2 = None
self.weights = None
self.j = []
self._randomize_thetas()
class LogisticRegression(object):
def __init__(self, class_range=3, file_name="", my_lambda=1):
self.X = None
self.Y = None
self.weights = None
self.class_range = class_range
self.my_lambda = my_lambda
self.cost_data = [[], [], []]
self.stat = Statistic(output=file_name)
self.X, self.Y, self.m = self.stat.get_training_set()
self.weights = np.zeros([self.class_range, self.X.shape[1]])
def cost_function(self, initial_theta, X, y, i):
zero_theta = initial_theta.copy()
zero_theta[0] = 0
reg = (self.my_lambda * np.power(zero_theta, 2).sum())/2
sig = sigmoid(np.matmul(X, initial_theta))
cost = (-np.matmul(np.transpose(y), np.log(sig)) - np.matmul(np.transpose(1 - y), np.log(1 - sig)) + reg) / self.m
self.cost_data[i].append(cost)
return cost
def gradient(self, initial_theta, X, y, i):
zero_theta = initial_theta.copy()
zero_theta[0] = 0
sig = sigmoid(np.matmul(X, initial_theta))
reg = self.my_lambda * zero_theta
grad = (np.matmul(np.transpose(X), sig - y) + reg) / self.m
return grad
def optimize_one(self, index):
initial_theta = np.zeros(len(self.weights[index]))
y = self.Y == index
y = y.astype(int)
result = op.minimize(fun=self.cost_function,
x0=initial_theta,
args=(self.X, y, index),
method='TNC',
jac=self.gradient)
print(result)
return result.x
def optimize(self):
t = list()
for i in range(self.class_range):
t.append(self.optimize_one(i))
self.weights = np.asarray(t)
return self.weights
def predict(self, x):
w = np.matmul(self.weights, x).tolist()
print(w, w.index(max(w)))
return w.index(max(w))