def new_target(self, uav, noisy_states):
ogrid = [np.linspace(*dim, num=res) for dim, res in zip(
self.get_effective_area(), self.grid_resolution)]
x, y, z = meshgrid_nd(*ogrid)
acq = self.acquisition_fn(
np.column_stack((x.flat, y.flat, z.flat)), uav, noisy_states)
max_idx = np.unravel_index(np.argmax(acq), x.shape)
x0 = np.array([x[max_idx], y[max_idx], z[max_idx]])
x, val, unused = fmin_l_bfgs_b(
NegateFn(self.acquisition_fn).eval_with_derivative, x0,
args=(uav, noisy_states,), bounds=self.get_effective_area(),
pgtol=1e-10, factr=1e2)
idx = np.argmax(self.acquisition_fn.predictor.y_train.data)
x0 = np.asarray(noisy_states[uav].position)
for dx in 5 * rnd.randn(5):
x2, val2, unused = fmin_l_bfgs_b(
NegateFn(self.acquisition_fn).eval_with_derivative, x0 + dx,
args=(uav, noisy_states,), bounds=self.get_effective_area(),
pgtol=1e-10, factr=1e2)
if val2 < val:
x = x2
idx = np.argmax(self.acquisition_fn.predictor.y_train.data)
x0 = self.acquisition_fn.predictor.x_train.data[idx]
for dx in 5 * rnd.randn(5):
x2, val2, unused = fmin_l_bfgs_b(
NegateFn(self.acquisition_fn).eval_with_derivative, x0 + dx,
args=(uav, noisy_states,), bounds=self.get_effective_area(),
pgtol=1e-10, factr=1e2)
if val2 < val:
x = x2
return [x]
def init(self, conf):
area = conf['area']
self.area = np.asarray(area)
self._positions = self.fileh.createEArray(
self.fileh.root, 'positions', tables.FloatAtom(),
(self.client.numUAVs, 0, 3), expectedrows=self.expected_steps,
title='Noiseless positions (numUAVs x timesteps x 3) of the UAVs '
'over time.')
self.fileh.createArray(
self.fileh.root, 'sample_locations', self.client.get_locations(),
title='Locations where prediction was requested '
'(num locations x 3)')
self.fileh.createArray(
self.fileh.root, 'reference_samples',
np.asarray(self.client.get_reference_samples()),
title='Reference samples (num locations)')
ogrid = [np.linspace(*dim, num=res) for dim, res in zip(
area, [29, 29, 9])]
x, y, z = meshgrid_nd(*ogrid)
locations = np.column_stack((x.flat, y.flat, z.flat))
self.fileh.createArray(
self.fileh.root, 'gt_locations', locations,
title='Locations where ground truth was evaluated '
'(num locations x 3)')
self.fileh.createArray(
self.fileh.root, 'gt_samples',
np.asarray(self.client.get_samples(locations)),
title='Ground truth samples (num locations)')
def predict_on_volume(predictor, area, grid_resolution):
ogrid = [np.linspace(*dim, num=res) for dim, res in zip(
area, grid_resolution)]
x, y, z = meshgrid_nd(*ogrid)
pred, mse = predictor.predict(
np.column_stack((x.flat, y.flat, z.flat)), eval_MSE=True)
np.maximum(0, pred, out=pred)
assert x.shape == y.shape and y.shape == z.shape
pred = pred.reshape(x.shape)
mse = mse.reshape(x.shape)
return pred, mse, (x, y, z)
def _plot_plume(self):
area = self.conf['area']
ogrid = [np.linspace(*dim, num=res) for dim, res in zip(
area, (29, 29, 9))]
x, y, z = meshgrid_nd(*ogrid)
values = griddata(
self.data.root.gt_locations.read(),
self.data.root.gt_samples.read(),
np.column_stack((x.flat, y.flat, z.flat))).reshape(x.shape)
self._truth_volume = self.plot_volume2(
(x, y, z), values, 0.1, figure=self.truth.mayavi_scene)
if not self.plain:
mlab.points3d(
*self.data.root.sample_locations.read().T, scale_factor=5,
color=(0.7, 0.0, 0.0), figure=self.truth.mayavi_scene)
