def test_learn_xnor_operations(self):
sgd = Sgd(sizes=[2, 3, 1])
print('Training...')
for _ in range(500):
sgd.update_network(input=numpy.array([[0], [0]]), expected_output=numpy.array([[1]]), mini_batch=2, eta=2.0)
sgd.update_network(input=numpy.array([[0], [1]]), expected_output=numpy.array([[0]]), mini_batch=2, eta=2.0)
sgd.update_network(input=numpy.array([[1], [0]]), expected_output=numpy.array([[0]]), mini_batch=2, eta=2.0)
sgd.update_network(input=numpy.array([[1], [1]]), expected_output=numpy.array([[1]]), mini_batch=2, eta=2.0)
print('Testing...')
output1: List[float] = sgd.feed_forward(input=numpy.array([[0], [0]]))
self.assertEqual(round(output1[0][0]), 1.0)
output4: List[float] = sgd.feed_forward(input=numpy.array([[0], [1]]))
self.assertEqual(round(output4[0][0]), 0.0)
output3: List[float] = sgd.feed_forward(input=numpy.array([[1], [0]]))
self.assertEqual(round(output3[0][0]), 0.0)
output2: List[float] = sgd.feed_forward(input=numpy.array([[1], [1]]))
self.assertEqual(round(output2[0][0]), 1.0)
def test_learn_handwriting(self):
sgd = Sgd(sizes=[Mnist.IMAGE_SIZE * Mnist.IMAGE_SIZE, 30, 10])
mnist_datas = Mnist.retrieve_mnist_datas(
filename='../../classes/mnist/mnist.raw')
print('Training...')
for mnist_data in mnist_datas[:50000]:
image_bytes = mnist_data['image_bytes']
input = numpy.frombuffer(buffer=image_bytes,
dtype='uint8').reshape(
(len(image_bytes), 1))
normalized_input = input / 255
number = mnist_data['number']
expected_output = numpy.zeros(shape=(10, 1))
expected_output[number] = 1.0
sgd.update_network(input=normalized_input,
expected_output=expected_output,
mini_batch=10,
eta=4.0)
print('Testing...')
correct = 0
total = 0
for mnist_data in mnist_datas[50000:]:
image_bytes = mnist_data['image_bytes']
input = numpy.frombuffer(buffer=image_bytes,
dtype='uint8').reshape(
(len(image_bytes), 1))
normalized_input = input / 255
number = mnist_data['number']
output: List[float] = sgd.feed_forward(input=normalized_input)
output_prob: List[float] = sgd.softmax(z=output)
output_index = numpy.argmax(output_prob)
if output_index == number:
correct += 1
total += 1
percent_correct = 100 * correct / total
print(
f'Correct {correct} out of {total} ({percent_correct:.1f}% correct)'
)
self.assertGreater(percent_correct, 80)
def test_softmax_02(self):
z = numpy.array([2.0, 1.0, 0.0, -1.0, -2.0])
output = Sgd.softmax(z).tolist()
self.assertEqual(output, [0.6364086465588308, 0.23412165725273662, 0.0861285444362687, 0.03168492079612427,
0.011656230956039609])
self.assertAlmostEqual(sum(output), 1.0)
def test_softmax_01(self):
z = numpy.array([2.0, 1.0, 0.0])
output = Sgd.softmax(z).tolist()
self.assertEqual(output, [0.6652409557748219, 0.24472847105479764, 0.09003057317038046])
self.assertAlmostEqual(sum(output), 1.0)
def test_sigmoid_prime(self):
z = numpy.array([-1.0, 0.0, 1.0])
output = Sgd.sigmoid_prime(z).tolist()
self.assertEqual(output, [0.19661193324148185, 0.25, 0.19661193324148185])
def test_sigmoid_prime_f(self):
self.assertAlmostEqual(Sgd.sigmoid_prime_f(-1.0), 0.19661193324148185)
self.assertAlmostEqual(Sgd.sigmoid_prime_f(0.0), 0.25)
self.assertAlmostEqual(Sgd.sigmoid_prime_f(1.0), 0.19661193324148185)
def test_sigmoid(self):
z = numpy.array([-1.0, 0.0, 1.0])
output = Sgd.sigmoid(z).tolist()
self.assertEqual(output, [0.2689414213699951, 0.5, 0.7310585786300049])
def test_sigmoid_f(self):
self.assertAlmostEqual(Sgd.sigmoid_f(-1.0), 0.2689414213699951)
self.assertAlmostEqual(Sgd.sigmoid_f(0.0), 0.5)
self.assertAlmostEqual(Sgd.sigmoid_f(1.0), 0.7310585786300049)