def test_numerical_gradient_checking(self):
label, image = next(mnist.read())
ninput = [pixel / 255 for row in image for pixel in row]
expected = [1 if i == label else 0 for i in range(10)]
nnet = NeuralNetwork([784, 16, 16, 10])
epsilon = 1e-5
numgrad = [np.empty(wmatrix.shape) for wmatrix in nnet.weight]
for k, wmatrix in enumerate(nnet.weight):
for i, w in np.ndenumerate(wmatrix):
wmatrix[i] = w - epsilon
nnet.feedforward(ninput)
a = nnet.get_error(expected)
wmatrix[i] = w + epsilon
nnet.feedforward(ninput)
b = nnet.get_error(expected)
numgrad[k][i] = (b - a) / 2 * epsilon
wmatrix[i] = w
error_gradient = nnet.get_error_gradient(expected)
unit = lambda v: v / norm(v) if (v != 0).any() else np.zeros(v.shape)
for k in range(len(nnet.weight)):
ag = error_gradient[k]
ng = numgrad[k]
print(f"custom = {norm(unit(ag) - unit(ng))}")
print(
f"derived from cs231 = {norm(unit(ag) * norm(ng) - ng) / max(norm(ag), norm(ng))}"
)
def backpropagation_main():
label, image = next(mnist.read())
ninput = [pixel / 255 for row in image for pixel in row]
expected = [1 if i == label else 0 for i in range(10)]
nnet = NeuralNetwork(DLAYERS, params=None)
# nnet = NeuralNetwork(DLAYERS, params=load_params())
for i in range(1000000000000):
guess = nnet.feedforward(ninput)
cost = nnet.get_error(expected)
print(f"[{i + 1}] cost = {cost}, guess = {guess}")
try:
nnet.backpropagate(expected)
except KeyboardInterrupt:
break
guess = nnet.feedforward(ninput)
cost = nnet.get_error(expected)
print(f"[{i + 1}] cost = {cost}")
save_params(nnet.params)