def test_forward_pass_multi_sample(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
out_buf = np.zeros(self.T.shape, dtype=self.T.dtype)
scc.forward_pass(self.theta, [self.X], out_buf)
assert_allclose(out_buf, self.T)
def test_dimensions(self):
lc = LinearCombination(5, 7)
sig = Sigmoid(7, 7)
scc = SequentialContainerConnection(5, 7, [lc, sig])
self.assertEqual(scc.input_dim, 5)
self.assertEqual(scc.output_dim, 7)
self.assertEqual(scc.get_param_dim(), 5*7)
def test_forward_pass_multi_sample(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
out_buf = np.zeros(self.T.shape, dtype=self.T.dtype)
scc.forward_pass(self.theta, [self.X], out_buf)
assert_allclose(out_buf, self.T)
def test_dimensions(self):
lc = LinearCombination(5, 7)
sig = Sigmoid(7, 7)
scc = SequentialContainerConnection(5, 7, [lc, sig])
self.assertEqual(scc.input_dim, 5)
self.assertEqual(scc.output_dim, 7)
self.assertEqual(scc.get_param_dim(), 5 * 7)
def test_backprop_multisample_zero_is_zero(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
in_error_buffers = [np.zeros_like(self.X)]
scc.backprop(self.theta, [self.X], self.T, np.zeros_like(self.T), in_error_buffers)
grad = np.zeros_like(self.theta)
scc.calculate_gradient(self.theta, grad, [self.X], self.T, in_error_buffers, np.zeros_like(self.T))
assert_allclose(in_error_buffers[0], np.zeros_like(self.X))
assert_allclose(grad, np.zeros_like(self.theta))
def test_forward_pass_single_samples(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
for x, t in zip(self.X, self.T):
t = np.atleast_2d(t)
x = np.atleast_2d(x)
out_buf = np.zeros_like(t)
scc.forward_pass(self.theta, [x], out_buf)
assert_allclose(out_buf, t)
def test_forward_pass_single_samples(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
for x, t in zip(self.X, self.T):
t = np.atleast_2d(t)
x = np.atleast_2d(x)
out_buf = np.zeros_like(t)
scc.forward_pass(self.theta, [x], out_buf)
assert_allclose(out_buf, t)
def test_backprop_multisample_zero_is_zero(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
in_error_buffers = [np.zeros_like(self.X)]
scc.backprop(self.theta, [self.X], self.T, np.zeros_like(self.T),
in_error_buffers)
grad = np.zeros_like(self.theta)
scc.calculate_gradient(self.theta, grad, [self.X], self.T,
in_error_buffers, np.zeros_like(self.T))
assert_allclose(in_error_buffers[0], np.zeros_like(self.X))
assert_allclose(grad, np.zeros_like(self.theta))
def test_backprop_multisample(self):
lc = LinearCombination(4, 1)
sig = Sigmoid(1, 1)
scc = SequentialContainerConnection(4, 1, [lc, sig])
assert_backprop_correct(scc, self.theta, [self.X],
np.ones_like(self.T))
