def test_get_differentials(self):
derivatives = self.network.get_derivatives()
self.network.update_network(self.X)
# Calculate derivative numerically using taylor series
dw = 1E-6
for test_i, layer in enumerate(self.network.data):
for test_j in range(len(layer)):
analytical = derivatives[test_i][test_j](self.Y)
for k in range(self.network.data[test_i][test_j].w.shape[0]):
if type(self.network.data[test_i][test_j]) == Softmax:
for l in range(self.network.data[0][0].number_classes):
self.network.data[0][0].W[k][l] -= dw
output1 = self.network.update_network(self.X)[0]
self.network.data[0][0].W[k][l] += 2 * dw
output2 = self.network.update_network(self.X)[0]
numerical = (log_likelihood(self.Y, output2) -
log_likelihood(self.Y, output1)) / (
2 * dw)
self.assertAlmostEqual(
cp.asnumpy(numerical),
cp.asnumpy(analytical[k][l]), 5)
else:
self.network.data[test_i][test_j].w[k] -= dw
output1 = self.network.update_network(self.X)[0]
self.network.data[test_i][test_j].w[k] += 2 * dw
output2 = self.network.update_network(self.X)[0]
numerical = (log_likelihood(self.Y, output2) -
log_likelihood(self.Y, output1)) / (2 *
dw)
self.assertAlmostEqual(cp.asnumpy(numerical),
cp.asnumpy(analytical[k]), 5)
],
[
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components))),
SingleNeuron(list(range(n_components)))
]])
predictions, x = network.update_network(X2_train)
ll_train = log_likelihood(Y_train, predictions)
predictions, x = network.update_network(X2_test)
ll_test = log_likelihood(Y_test, predictions)
print(f'Initial ll_train = {ll_train/n_train}, ll_test = {ll_test/(n_test)}')
# To load previously trained weights
# w_data = np.load('cifar-10/w_data.npy', allow_pickle=True)
# network.set_weights(w_data)
network.train(X2_train, Y_train, 50)
predictions, x = network.update_network(X2_train)
ll_train = log_likelihood(Y_train, predictions)
predictions, x = network.update_network(X2_test)
ll_test = log_likelihood(Y_test, predictions)
print(f'Final ll_train = {ll_train/n_train}, ll_test = {ll_test/(n_test)}\n')
def test_log_likelihood(self):
output = self.network.update_network(self.X)[0]
self.assertEqual(log_likelihood(self.Y, output).shape, ())