def test_input_gradient(self):
"""
"""
us = np.ones((1, 2))
cost_grad =\
TwoWheeledConfigModule.gradient_cost_fn_with_input(None, us)
# numeric grad
eps = 1e-4
expected_grad = np.zeros((1, 2))
for i in range(2):
tmp_u = us.copy()
tmp_u[0, i] = us[0, i] + eps
forward = \
TwoWheeledConfigModule.input_cost_fn(tmp_u)
forward = np.sum(forward)
tmp_u = us.copy()
tmp_u[0, i] = us[0, i] - eps
backward = \
TwoWheeledConfigModule.input_cost_fn(tmp_u)
backward = np.sum(backward)
expected_grad[0, i] = (forward - backward) / (2. * eps)
assert cost_grad == pytest.approx(expected_grad)
def test_calc_costs(self):
# make config
config = TwoWheeledConfigModule()
# set
pred_len = 5
state_size = 3
input_size = 2
pop_size = 2
pred_xs = np.ones((pop_size, pred_len, state_size))
g_xs = np.ones((pop_size, pred_len, state_size)) * 0.5
input_samples = np.ones((pop_size, pred_len, input_size)) * 0.5
costs = config.input_cost_fn(input_samples)
expected_costs = np.ones((pop_size, pred_len, input_size))*0.5
assert costs == pytest.approx(expected_costs**2 * np.diag(config.R))
costs = config.state_cost_fn(pred_xs, g_xs)
expected_costs = np.ones((pop_size, pred_len, state_size))*0.5
assert costs == pytest.approx(expected_costs**2 * np.diag(config.Q))
costs = config.terminal_state_cost_fn(pred_xs[:, -1, :],\
g_xs[:, -1, :])
expected_costs = np.ones((pop_size, state_size))*0.5
assert costs == pytest.approx(expected_costs**2 * np.diag(config.Sf))