def train(self, debug=False, interval=1000):
f = lambda x: self.feed_forward()
start_time = datetime.now()
if debug:
print("Test System :")
print("learning rate:", self.learning_rate, "iteration count:",
self.iteration_count)
print(self.name, "initial loss value :", self.loss_val(), "\n")
for step in range(self.iteration_count + 1):
self.W2 -= self.learning_rate * numerical_derivative(f, self.W2)
self.b2 -= self.learning_rate * numerical_derivative(f, self.b2)
self.W3 -= self.learning_rate * numerical_derivative(f, self.W3)
self.b3 -= self.learning_rate * numerical_derivative(f, self.b3)
if debug and step % interval == 0:
print(self.name, "step :", step, ", loss_value :",
self.loss_val())
if debug:
print(self.name, 'train result:')
print("final loss value :", self.loss_val())
print("final W2 =", self.W2)
print("final b2 =", self.b2)
print("final W3 =", self.W3)
print("final b3 =", self.b3)
print(self.name, "Elapsed time :",
datetime.now() - start_time, "\n")
def train(self, input, target):
self.input_data = input
self.target_data = target
f = lambda x: self.feed_forward()
self.W2 -= self.learning_rate * numerical_derivative(f, self.W2)
self.b2 -= self.learning_rate * numerical_derivative(f, self.b2)
self.W3 -= self.learning_rate * numerical_derivative(f, self.W3)
self.b3 -= self.learning_rate * numerical_derivative(f, self.b3)
def train(self, debug=False, interval=5):
for step in range(self.epochs):
for i in range(len(self.xdata)):
self.input_data = self.xdata[i]
self.target_data = self.tdata[i]
f = lambda x: self.feed_forward()
for i in range(len(self.W)):
self.W[i] -= self.learning_rate * numerical_derivative(
f, self.W[i])
self.b[i] -= self.learning_rate * numerical_derivative(
f, self.b[i])
if debug and step % interval == 0:
print("epoch:", step, "loss_val:", self.loss_val())
def train(self, debug=False, interval=1000):
if debug:
print("initial error value :", self.error_val(), "initial W =",
self.W, "\nb =", self.b)
start_time = datetime.now()
f = lambda x: self.loss_func()
for step in range(self.iteration_count + 1):
self.W -= self.learning_rate * numerical_derivative(f, self.W)
self.b -= self.learning_rate * numerical_derivative(f, self.b)
if debug and step % interval == 0:
print("step :", step, "error value :", self.error_val(), "W =",
self.W, "b =", self.b)
if debug:
print("final error value :", self.error_val(), "final W =", self.W,
"\nb =", self.b)
print("Elapsed time :", datetime.now() - start_time)
xdata = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
and_tdata = np.array([0, 0, 0, 1]).reshape(4, 1)
or_tdata = np.array([0, 1, 1, 1]).reshape(4, 1)
nand_tdata = np.array([1, 1, 1, 0]).reshape(4, 1)
xor_tdata = np.array([0, 1, 1, 0]).reshape(4, 1)
test_data = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
input_node_num = 2
hidden_node_num = 4
output_node_num = 1
W2 = np.random.rand(input_node_num, hidden_node_num)
b2 = np.random.rand(hidden_node_num)
W3 = np.random.rand(hidden_node_num, output_node_num)
b3 = np.random.rand(output_node_num)
# xor problem solved!
learning_rate = 1e-2
f = lambda x: feed_forward(xdata, xor_tdata)
for step in range(30000):
W2 -= learning_rate * numerical_derivative(f, W2)
b2 -= learning_rate * numerical_derivative(f, b2)
W3 -= learning_rate * numerical_derivative(f, W3)
b3 -= learning_rate * numerical_derivative(f, b3)
if step % 1000 == 0:
print("step :", step, "loss_value :", loss_val(xdata, xor_tdata))
for data in test_data:
print('predict', data, ':', predict(data))
# error_val does exactly same as what loss_func function does.
def error_val(x, t):
y = np.dot(x, W) + b
return np.sum((t - y)**2) / len(x)
def predict(x):
y = np.dot(x, W) + b
return y
if __name__ == "__main__":
# 1e-2 ~ 1e-8
learning_rate = 1e-5
f = lambda x: loss_func(x_data, t_data)
print("initial error value :", error_val(x_data, t_data), "W =", W,
"\nb =", b)
# 8000 ~ 500000
for step in range(20001):
W -= learning_rate * numerical_derivative(f, W)
b -= learning_rate * numerical_derivative(f, b)
if step % 10000 == 0:
print("step :", step, "error value :", error_val(x_data, t_data),
" W =", W, "\nb =", b)
print("final error value :", error_val(x_data, t_data), " W =", W, "\nb =",
b)
test_data = [[4, 4, 4, 4], [-3, 0, 9, -1], [7, -9, 2, -8]]
for data in test_data:
print('predict', data, ':', predict(np.array(data)))