def plan_path(self):
print("Searching for a path ...")
TARGET_ALTITUDE = 5
SAFETY_DISTANCE = 5
self.target_position[2] = TARGET_ALTITUDE
# TODO: read lat0, lon0 from colliders into floating point values
data = np.genfromtxt('colliders.csv', delimiter=',', dtype='str', max_rows=1)
_, lat0 = data[0].strip().split(' ')
_, lon0 = data[1].strip().split(' ')
# TODO: set home position to (lon0, lat0, 0)
self.set_home_position(float(lon0), float(lat0), 0)
# TODO: retrieve current global position
curr_glb_pos = [self._longitude, self._latitude, self._altitude]
# TODO: convert to current local position using global_to_local()
self._north, self._east, self._down = global_to_local(curr_glb_pos, self.global_home)
print('global home {0}, position {1}, local position {2}'.format(self.global_home, self.global_position,
self.local_position))
# Read in obstacle map
data = np.loadtxt('colliders.csv', delimiter=',', dtype='Float64', skiprows=2)
sampler = Sampler(data, SAFETY_DISTANCE)
self.polygons = sampler._polygons
nodes = sampler.sample(3)
print('nodes_len: ', len(nodes))
xx = [[p[0], p[1], TARGET_ALTITUDE, 0] for p in nodes]
data = msgpack.dumps(xx)
print(data)
self.connection._master.write(data)
def get_path(start=(0, 0, 6), goal=(331, 116, 19)):
data = np.loadtxt('colliders.csv',
delimiter=',',
dtype='Float64',
skiprows=2)
north_offset = -316
east_offset = -445
sampler = Sampler(data)
print("AAA")
polygons = sampler._polygons
print("BBB")
nodes = sampler.sample(3000)
# (339, 380) (441, 519)
# start = (0, 0, 6)
# goal = (331, 116, 19)
nodes += [start, goal]
print(nodes)
g = create_graph(nodes, k=10, polygons=polygons)
path, _ = a_star_graph(g, heuristic, start, goal)
waypoints = [[int(p[0]), int(p[1]), int(p[2]), 0] for p in path]
path_dict = {'path': waypoints}
print(waypoints)
with open('path.json', 'w') as json_file:
json.dump(path_dict, json_file)
with open('path.json', 'r') as json_file2:
dta = json.load(json_file2)
print(dta['path'])
return waypoints
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
expectations = [None for i in self.dbm.weights]
signal_up = [None for i in self.dbm.weights]
signal_down = [None for i in self.dbm.weights]
multiply_up = [None for i in self.dbm.weights]
multiply_down = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
signal_up[i] = self.dbm.signal(values[i], i+1)
signal_down[i] = self.dbm.signal(values[i+1], -(i+1))
multiply_up[i] = signal_up[i][:, tf.newaxis, :] - values[i][:, :, tf.newaxis] * self.dbm.weights[i]
multiply_down[i] = signal_down[i][:, :, tf.newaxis] - values[i+1][:, tf.newaxis, :] * self.dbm.weights[i]
# expectation[0]
x = multiply_up[0] + self.mariginalize2(multiply_up[1], self.dbm.weights[1], axis=2)[:, tf.newaxis, :]
y = multiply_down[0]
z = self.dbm.weights[0][tf.newaxis, :, :]
expectations[0] = self.mariginalize(x, y, z)
# expectation[1]
x = multiply_down[1] + self.mariginalize2( multiply_down[0], self.dbm.weights[0], axis=1)[:, :, tf.newaxis]
y = multiply_up[1]
z = self.dbm.weights[1][tf.newaxis, :, :]
expectations[1] = self.mariginalize(x, y, z)
return expectations
def create_graph_with_nodes(data, lat0, lon0, alt):
print('[start]create_grid_with_nodes')
sampler = Sampler(data)
global polygons
polygons = sampler._polygons
nodes = sampler.sample(300)
graph = create_graph(nodes, 10)
print("Number of edges", len(graph.edges))
grid = create_grid(data, sampler._zmax, 1)
nmin = np.min(data[:, 0])
emin = np.min(data[:, 1])
start = (lat0 + nmin, lon0 + emin, alt)
start_closest = closest_point(graph, start)
k = np.random.randint(len(graph.nodes))
print(k, len(graph.nodes))
goal = list(graph.nodes)[k]
print('start {0}'.format(start_closest))
print('goal {0}'.format(goal))
print('[end]create_grid_with_nodes')
return grid, graph, start_closest, goal
def create_graph_probabilistic(data,
drone_altitude,
safety_distance,
k=10,
sample_size=300):
sampler = Sampler(data, drone_altitude, safety_distance)
polygons = sampler._polygons
print("Polygons", len(polygons))
nodes = sampler.sample(sample_size)
print("Data", len(data))
print("Probabilistic node length", len(nodes))
g = nx.Graph()
tree = KDTree(nodes)
for n1 in nodes:
# for each node connect try to connect to k nearest nodes
idxs = tree.query([n1], k, return_distance=False)[0]
for idx in idxs:
n2 = nodes[idx]
if n2 == n1:
continue
if can_connect(n1, n2, polygons):
g.add_edge(n1, n2, weight=1)
return g
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
weight = self.dbm.weight_matrix(values)
return weight
def plan_path(self):
self.flight_state = States.PLANNING
print("Searching for a path ...")
TARGET_ALTITUDE = 5
SAFETY_DISTANCE = 5
self.target_position[2] = TARGET_ALTITUDE
# TODO: read lat0, lon0 from colliders into floating point values
data = np.genfromtxt('colliders.csv',
delimiter=',',
dtype='str',
max_rows=1)
_, lat0 = data[0].strip().split(' ')
_, lon0 = data[1].strip().split(' ')
# TODO: set home position to (lon0, lat0, 0)
self.set_home_position(float(lon0), float(lat0), 0)
# TODO: retrieve current global position
curr_glb_pos = [self._longitude, self._latitude, self._altitude]
# TODO: convert to current local position using global_to_local()
self._north, self._east, self._down = global_to_local(
curr_glb_pos, self.global_home)
print('global home {0}, position {1}, local position {2}'.format(
self.global_home, self.global_position, self.local_position))
# Read in obstacle map
data = np.loadtxt('colliders.csv',
delimiter=',',
dtype='Float64',
skiprows=2)
sampler = Sampler(data, SAFETY_DISTANCE)
self.polygons = sampler._polygons
nodes = sampler.sample(300)
print('nodes_len: ', len(nodes))
g = self.create_graph(nodes, 10)
print('graph_edgs: ', len(g.edges))
start = self.local_position
goal = global_to_local([-122.396428, 37.795128, TARGET_ALTITUDE],
self.global_home)
start = self.find_closest_node(g.nodes, start)
goal = self.find_closest_node(g.nodes, goal)
path, cost = a_star_for_graph(g, heuristic, start, goal)
print('a_star_path: ', path)
path = self.prune_path(path)
print('prune_path: ', path)
if len(path) > 0:
# Convert path to waypoints
waypoints = [[p[0], p[1], TARGET_ALTITUDE, 0] for p in path]
# Set self.waypoints
self.waypoints = waypoints
# TODO: send waypoints to sim (this is just for visualization of waypoints)
self.send_waypoints()
def sample_nodes(data, grid_start, grid_goal, debug):
start_time = time.time()
sampler = Sampler(data)
# Extract all the polygons
polygons = sampler._polygons
nodes = sampler.sample(grid_start, grid_goal, NUM_SAMPLES)
stop_time = time.time()
if debug:
print("Time taken to build Sampler is: {0:5.2f}s".format(stop_time -
start_time))
return nodes, polygons
def nonbatch(task, method, N, M):
simulation_object = create_env(task)
d = simulation_object.num_of_features
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
psi_set = []
s_set = []
input_A = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(2 * simulation_object.feed_size))
input_B = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(2 * simulation_object.feed_size))
psi, s = get_feedback(simulation_object, input_A, input_B)
psi_set.append(psi)
s_set.append(s)
for i in range(1, N):
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples, axis=0)
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
input_A, input_B = run_algo(method, simulation_object, w_samples)
psi, s = get_feedback(simulation_object, input_A, input_B)
psi_set.append(psi)
s_set.append(s)
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
def nonbatch(task, method, N, M):
simulation_object = create_env(task)
d = simulation_object.num_of_features
w_true = 2*np.random.rand(d)-1
w_true = w_true / np.linalg.norm(w_true)
print('If in automated mode: true w = {}'.format(w_true/np.linalg.norm(w_true)))
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
psi_set = []
s_set = []
for i in range(N):
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1,1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples,axis=0)
print('Samples so far: ' + str(i))
print('w estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
print('Alignment = {}'.format(mean_w_samples.dot(w_true)/np.linalg.norm(mean_w_samples)))
input_A, input_B = run_algo(method, simulation_object, w_samples)
psi, s = get_feedback(simulation_object, input_A, input_B, w_true)
psi_set.append(psi)
s_set.append(s)
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1,1)
w_samples = w_sampler.sample(M)
print('Samples so far: ' + str(N))
print('w estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
print('Alignment = {}'.format(mean_w_samples.dot(w_true)/np.linalg.norm(mean_w_samples)))
class first_smci:
def __init__(self, dbm, sample_size=500, initial_update=1000, update_time=1):
self.sampler = None
self.dbm = dbm
self.initial_update = initial_update
self.update_time = update_time
self.sample_size = sample_size
self.mariginalize = self.dbm.propagation.first_smci_marginalize
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
signals = [None for i in self.dbm.layers]
signals[0] = self.dbm.signal(values[1], -1)
for i in range(1, len(self.dbm.layers)-1):
signals[i] = self.dbm.signal(values[i-1], i) + self.dbm.signal(values[i+1], -(i+1))
signals[-1] = self.dbm.signal(values[-2], len(self.dbm.weights))
multiply_up = [None for i in self.dbm.weights]
multiply_down = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
multiply_up[i] = values[i][:, :, tf.newaxis] * self.dbm.weights[i]
multiply_down[i] = values[i+1][:, tf.newaxis, :] * self.dbm.weights[i]
expectations = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
expectations[i] = self.mariginalize( signals[i][:, :, tf.newaxis]-multiply_down[i], signals[i+1][:, tf.newaxis, :]-multiply_up[i], self.dbm.weights[i])
return expectations
def batch(task, method, N, M, b):
if N % b != 0:
print('N must be divisible to b')
exit(0)
B = 20 * b
simulation_object = create_env(task)
d = simulation_object.num_of_features
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
psi_set = []
s_set = []
inputA_set = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(b, 2 * simulation_object.feed_size))
inputB_set = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(b, 2 * simulation_object.feed_size))
for j in range(b):
input_A = inputA_set[j]
input_B = inputB_set[j]
psi, s = get_feedback(simulation_object, input_A, input_B)
psi_set.append(psi)
s_set.append(s)
i = b
while i < N:
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples, axis=0)
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
print('Samples so far: ' + str(i))
inputA_set, inputB_set = run_algo(method, simulation_object, w_samples,
b, B)
for j in range(b):
input_A = inputA_set[j]
input_B = inputB_set[j]
psi, s = get_feedback(simulation_object, input_B, input_A)
psi_set.append(psi)
s_set.append(s)
i += b
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples, axis=0)
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
def nonbatch(task, criterion, query_type, epsilon, M):
simulation_object = create_env(task)
d = simulation_object.num_of_features
true_delta = 1 # make this None if you will also learn delta, and change the samplers below from sample_given_delta to sample (and of course remove the true_delta argument)
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
i = 0
score = np.inf
while score >= epsilon:
w_samples, delta_samples = w_sampler.sample_given_delta(M, query_type, true_delta)
mean_w_samples = np.mean(w_samples,axis=0)
print('w-estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
input_A, input_B, score = run_algo(criterion, simulation_object, w_samples, delta_samples)
if criterion == 'information':
print('Expected info gain = {}'.format(score))
elif criterion == 'volume':
print('Expected volume removal (meaningless scale) = {}'.format(score/M))
if score > epsilon:
phi_A, phi_B, s = get_feedback(simulation_object, input_A, input_B, query_type)
w_sampler.feed(phi_A, phi_B, [s])
i += 1
w_samples, delta_samples = w_sampler.sample_given_delta(M, query_type, true_delta)
mean_w_samples = np.mean(w_samples,axis=0)
print('w-estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
signals = [None for i in self.dbm.layers]
signals[0] = self.dbm.signal(values[1], -1)
for i in range(1, len(self.dbm.layers)-1):
signals[i] = self.dbm.signal(values[i-1], i) + self.dbm.signal(values[i+1], -(i+1))
signals[-1] = self.dbm.signal(values[-2], len(self.dbm.weights))
multiply_up = [None for i in self.dbm.weights]
multiply_down = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
multiply_up[i] = values[i][:, :, tf.newaxis] * self.dbm.weights[i]
multiply_down[i] = values[i+1][:, tf.newaxis, :] * self.dbm.weights[i]
expectations = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
expectations[i] = self.mariginalize( signals[i][:, :, tf.newaxis]-multiply_down[i], signals[i+1][:, tf.newaxis, :]-multiply_up[i], self.dbm.weights[i])
return expectations
class four_layer_second_smci:
def __init__(self, dbm, sample_size=500, initial_update=1000, update_time=1):
if len(dbm.layers) != 4:
raise ValueError("2-SMCI supports only 4-layered DBM.")
self.sampler = None
self.dbm = dbm
self.initial_update = initial_update
self.update_time = update_time
self.sample_size = sample_size
self.mariginalize = self.dbm.propagation.first_smci_marginalize
self.mariginalize2 = self.dbm.propagation.second_smci_marginalize
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
expectations = [None for i in self.dbm.weights]
signal_up = [None for i in self.dbm.weights]
signal_down = [None for i in self.dbm.weights]
multiply_up = [None for i in self.dbm.weights]
multiply_down = [None for i in self.dbm.weights]
for i,_ in enumerate(self.dbm.weights):
signal_up[i] = self.dbm.signal(values[i], i+1)
signal_down[i] = self.dbm.signal(values[i+1], -(i+1))
multiply_up[i] = signal_up[i][:, tf.newaxis, :] - values[i][:, :, tf.newaxis] * self.dbm.weights[i]
multiply_down[i] = signal_down[i][:, :, tf.newaxis] - values[i+1][:, tf.newaxis, :] * self.dbm.weights[i]
# expectation[0]
x = multiply_up[0] + self.mariginalize2( signal_down[2][:, tf.newaxis, :] + multiply_up[1], self.dbm.weights[1], axis=2)[:, tf.newaxis, :]
y = multiply_down[0]
z = self.dbm.weights[0][tf.newaxis, :, :]
expectations[0] = self.mariginalize(x, y, z)
# expectation[1]
x = multiply_down[1] + self.mariginalize2( multiply_down[0], self.dbm.weights[0], axis=1)[:, :, tf.newaxis]
y = multiply_up[1] + self.mariginalize2( multiply_up[2], self.dbm.weights[2], axis=2)[:, tf.newaxis, :]
z = self.dbm.weights[1][tf.newaxis, :, :]
expectations[1] = self.mariginalize(x, y, z)
# expectation[2]
x = multiply_up[2]
y = multiply_down[2] + self.mariginalize2( signal_up[0][:, :, tf.newaxis] + multiply_down[1], self.dbm.weights[1], axis=1)[:, :, tf.newaxis]
z = self.dbm.weights[2][tf.newaxis, :, :]
expectations[2] = self.mariginalize(x, y, z)
return expectations
def read_ply(self, file_name):
num_samples = self.num_samples // len(self.files_list)
if self.file_index == len(self.files_list) - 1:
num_samples = num_samples + (self.num_samples - (num_samples * len(self.files_list)))
root, ext = os.path.splitext(file_name)
if not os.path.isfile(root + ".npy"):
ply = PlyData.read(file_name)
vertex = ply['vertex']
(x, y, z) = (vertex[t] for t in ('x', 'y', 'z'))
points = zip(x.ravel(), y.ravel(), z.ravel())
np.save(root + ".npy", points)
else:
points = np.load(root + ".npy")
#load normals
if os.path.isfile(root + "_normals" + ".ply"):
if not os.path.isfile(root + "_normals" + ".npy"):
ply1 = PlyData.read(root + "_normals" + ".ply")
vertex = ply1['vertex']
(nx, ny, nz) = (vertex[t] for t in ('nx', 'ny', 'nz'))
self.normals = np.asarray(zip(nx.ravel(), ny.ravel(), nz.ravel()))
np.save(root + "_normals" + ".npy", self.normals)
else:
self.normals = np.load(root + "_normals" + ".npy")
if self.add_noise:
self.data = utils.add_noise_normal(points, std=self.nois_std)
else:
self.data = np.asarray(points)
self.pc_diameter = utils.get_pc_diameter(self.data)
self.l = self.relL*self.pc_diameter
rot = utils.angle_axis_to_rotation(self.rotation_angle, self.rotation_axis)
self.data = utils.transform_pc(self.data, rot)
#plotutils.show_pc(self.data)
#mlab.show()
#TODO: better sampling
print "sampling file: ", file_name
self.samples, self.sample_indices = Sampler.sample(self.data, -1, min_num_point=-1, file_name=file_name, sampling_algorithm=self.sampling_algorithm)
#self.samples, self.sample_indices = Sampler.sample(self.data, -1, num_samples, file_name=file_name, sampling_algorithm=self.sampling_algorithm)
#self.samples = self.samples[0:num_samples]
#self.sample_indices = self.sample_indices[0:num_samples]
self.tree = spatial.KDTree(self.data)
return self.data
def nonbatch(task, method, N, M, checkpoints=None):
if checkpoints is None:
checkpoints = []
checkpointed_weights = []
simulation_object = create_env(task)
d = simulation_object.num_of_features
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
psi_set = []
s_set = []
input_A = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(2 * simulation_object.feed_size))
input_B = np.random.uniform(low=2 * lower_input_bound,
high=2 * upper_input_bound,
size=(2 * simulation_object.feed_size))
psi, s = get_feedback_auto(
simulation_object, input_A,
input_B) # psi is the difference, s is the 1 or -1 signal
psi_set.append(psi)
s_set.append(s)
for i in range(1, N):
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples, axis=0)
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
if i in checkpoints:
checkpointed_weights.append(mean_w_samples /
np.linalg.norm(mean_w_samples))
print("Weights saved at iteration {}".format(i))
input_A, input_B = run_algo(method, simulation_object, w_samples)
psi, s = get_feedback_auto(simulation_object, input_A, input_B)
psi_set.append(psi)
s_set.append(s)
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1, 1)
w_samples = w_sampler.sample(M)
checkpointed_weights.append(mean_w_samples /
np.linalg.norm(mean_w_samples))
print('w-estimate = {}'.format(mean_w_samples /
np.linalg.norm(mean_w_samples)))
return checkpointed_weights
class montecarlo:
def __init__(self, dbm, sample_size=500, initial_update=1000, update_time=1):
self.sampler = None
self.dbm = dbm
self.initial_update = initial_update
self.update_time = update_time
self.sample_size = sample_size
def expectation(self):
if self.sampler is None:
self.sampler = Sampler(self.dbm, self.sample_size, self.initial_update, self.update_time)
values = self.sampler.sampling()
weight = self.dbm.weight_matrix(values)
return weight
def read_ply(self, file_name, num_samples=1000, sample_class_start=0, add_noise =False,
noise_prob=0.3, noise_factor=0.02, noise_std=0.1, sampling_algorithm=SampleAlgorithm.Uniform,
rotation_axis=[0, 0, 1], rotation_angle=0):
root, ext = os.path.splitext(file_name)
if not os.path.isfile(root + ".npy"):
ply = PlyData.read(file_name)
vertex = ply['vertex']
(x, y, z) = (vertex[t] for t in ('x', 'y', 'z'))
points = zip(x.ravel(), y.ravel(), z.ravel())
np.save(root + ".npy", points)
else:
points = np.load(root + ".npy")
#load normals
if os.path.isfile(root + "_normals" + ".ply"):
if not os.path.isfile(root + "_normals" + ".npy"):
ply1 = PlyData.read(root + "_normals" + ".ply")
vertex = ply1['vertex']
(nx, ny, nz) = (vertex[t] for t in ('nx', 'ny', 'nz'))
self.normals = np.asarray(zip(nx.ravel(), ny.ravel(), nz.ravel()))
np.save(root + "_normals" + ".npy", self.normals)
else:
self.normals = np.load(root + "_normals" + ".npy")
if add_noise:
print "adding noise to model.."
mr = utils.model_resolution(np.array(points))
#mr = 0.404
print "model resolution: ", mr
self.data = utils.add_noise_normal(np.array(points), mr, noise_std)
else:
self.data = np.asarray(points)
rot = utils.angle_axis_to_rotation(rotation_angle, rotation_axis)
self.data = utils.transform_pc(self.data, rot)
#plotutils.show_pc(self.data)
#mlab.show()
#TODO: better sampling
self.samples, self.sample_indices = Sampler.sample(self.data, -1, num_samples-1, file_name=file_name, pose=rot, sampling_algorithm=sampling_algorithm)
self.tree = spatial.KDTree(self.data)
self.sample_class_start = sample_class_start
self.sample_class_current = sample_class_start
self.num_samples = self.samples.shape[0]
print "num samples: ", self.num_samples
logging.basicConfig(filename='example.log',level=logging.DEBUG)
return self.data
def batch(task, method, N, M, b):
if N % b != 0:
print('N must be divisible to b')
exit(0)
B = 20*b
simulation_object = create_env(task)
d = simulation_object.num_of_features
w_true = 2*np.random.rand(d)-1
w_true = w_true / np.linalg.norm(w_true)
print('If in automated mode: true w = {}'.format(w_true/np.linalg.norm(w_true)))
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
psi_set = []
s_set = []
i = 0
while i < N:
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1,1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples,axis=0)
print('Samples so far: ' + str(i))
print('w estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
print('Alignment = {}'.format(mean_w_samples.dot(w_true)/np.linalg.norm(mean_w_samples)))
inputA_set, inputB_set = run_algo(method, simulation_object, w_samples, b, B)
for j in range(b):
input_A = inputA_set[j]
input_B = inputB_set[j]
psi, s = get_feedback(simulation_object, input_B, input_A, w_true)
psi_set.append(psi)
s_set.append(s)
i += b
w_sampler.A = psi_set
w_sampler.y = np.array(s_set).reshape(-1,1)
w_samples = w_sampler.sample(M)
mean_w_samples = np.mean(w_samples, axis=0)
print('Samples so far: ' + str(N))
print('w estimate = {}'.format(mean_w_samples/np.linalg.norm(mean_w_samples)))
print('Alignment = {}'.format(mean_w_samples.dot(w_true)/np.linalg.norm(mean_w_samples)))
def find_threshold(num_weighted_samples,
num_random_samples,
reward_values,
num_membership_queries=0,
task='driver',
method="nonbatch"):
# first, sample the trajectories from the distribution\
simulation_object = create_env(task)
d = simulation_object.num_of_features
lower_input_bound = [x[0] for x in simulation_object.feed_bounds]
upper_input_bound = [x[1] for x in simulation_object.feed_bounds]
w_sampler = Sampler(d)
# set the reward weights of the sampler
# set the number of membership queries as a log function of the total # of samples
num_membership_queries = max(
num_membership_queries,
int(math.ceil(math.log(num_weighted_samples + num_random_samples))))
reward_traj_set = collect_trajectories(simulation_object, method,
num_weighted_samples, reward_values)
random_traj_set = collect_trajectories(simulation_object, "random",
num_random_samples, reward_values)
w_true = np.array([0.56687795, -0.51010378, 0.5178173, 0.38769675])
svm_reward_set = reward_traj_set[:
num_membership_queries] + collect_trajectories(
simulation_object, method,
num_weighted_samples, reward_values)
#adding n more samples to the svm dataset --> n + log(n) samples
svm_random_set = random_traj_set[:
num_membership_queries] + collect_trajectories(
simulation_object, "random",
num_weighted_samples, reward_values)
#adding n more samples to the svm dataset --> n + log(n) samples
full_traj_set = reward_traj_set + random_traj_set
# sort the trajectories by reward
sorted_lattice = lattice.sort_on_rewards(full_traj_set)
#test set trajectories sampled from w_true
f_reward = open('reward_test_set.obj', 'rb')
reward_traj_set_test = pickle.load(f_reward)
f_reward.close()
#test set trajectories sampled randomly
f_random = open('random_test_set.obj', 'rb')
random_traj_set_test = pickle.load(f_random)
f_random.close()
#get data and labels for the test set
x = []
y = []
r = []
reward_traj_set_test = reward_traj_set_test + random_traj_set_test
for node in reward_traj_set_test:
#print(node.reward_value)
x.append(node.features)
reward = np.sum(np.dot(w_true, node.features))
r.append(reward)
if reward < 0.74:
y.append(0)
else:
y.append(1)
print(y)
#now, begin getting membership query feedback on things
bsearch_reward_bound, labeled_data = membership_threshold(
sorted_lattice, simulation_object, get_labels=True)
svm_bsearch_coeff, svm_bsearch_inter, clssfr_bsearch = svm_threshold(
svm_reward_set, simulation_object, labeled_samples=labeled_data)
svm_reward_coeff, svm_reward_inter, clssfr_reward = svm_threshold(
svm_reward_set, simulation_object)
svm_random_coeff, svm_random_inter, clssfr_random = svm_threshold(
svm_random_set, simulation_object)
# finished process
print("Reward boundary retrieved from binary search method is {}".format(
bsearch_reward_bound))
print(
"SVM coefficient and intercept for same queries as binary search are: {} and {}"
.format(svm_bsearch_coeff, svm_bsearch_inter))
print(
"SVM coefficient and intercept for reward-sampled queries are: {} and {}"
.format(svm_reward_coeff, svm_reward_inter))
print(
"SVM coefficient and intercept for random-sampled queries are: {} and {}"
.format(svm_random_coeff, svm_random_inter))
print("Reward weights for task are {}".format(reward_values))
acc_bsearch = get_accuracy(r,
y,
reward_bound=bsearch_reward_bound,
clssfr=None)
acc_svm_learnt = get_accuracy(x,
y,
reward_bound=None,
clssfr=clssfr_reward)
acc_svm_random = get_accuracy(x,
y,
reward_bound=None,
clssfr=clssfr_random)
print("Accuracy for binary search is ", acc_bsearch)
print("Accuracy for svm with reward-sampled queries is ", acc_svm_learnt)
print("Accuracy for svm with randomly-sampled queries is ", acc_svm_random)
if args.cell_type == 1:
rh = np.zeros([1, args.internal_size * (args.layers * 2)])
else:
rh = np.zeros([1, args.internal_size * (args.layers)])
out_str = ""
for k in range(1000):
ryo, rh = sess.run([Yo, H],
feed_dict={
X: ry,
dropout_prob: 1.0,
initial_state: rh,
batchsize: 1
})
char_int_list = []
for i in range(1):
char_int = Sampler.sample_by_prop(ryo[i], VOCAB_SIZE, 0)
out_str += str(translation["INT_TO_CHAR"][char_int])
char_int_list.append(char_int)
ry = np.array([char_int_list])
with open(
args.out_folder + "samples/sample_" + timestamp +
"_epoch_" + str(epoch + 1) + ".txt", "wb") as fd:
fd.write(
b"======================================== SAMPLE ========================================\n"
)
fd.write(out_str.encode())
fd.write(
b"\n====================================== END SAMPLE ======================================\n"
)
if (epoch + 1) % 10 == 0 and LAST_DIVISION != (epoch + 1):
learn_rate /= 2 # from the seq2seq paper but a wee bit later epoch wise
enc_dec.build_trainer(src, src_mask, trg, trg_mask)
enc_dec.build_sampler()
if configuration['reload']:
enc_dec.load()
sample_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=1,
maxlen=configuration['seq_len_src'], stochastic=True)
valid_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=configuration['beam_size'],
maxlen=3*configuration['seq_len_src'], stochastic=False)
sampler = Sampler(sample_search, **configuration)
bleuvalidator = BleuValidator(valid_search, **configuration)
# train function
train_fn = enc_dec.train_fn
if configuration.get('with_layernorm', False):
update_fn = enc_dec.update_fn
# train data
ds = DStream(**configuration)
# valid data
vs = get_devtest_stream(data_type='valid', input_file=None, **configuration)
# main_loop
# modified by Zhaopeng Tu, 2016-07-14
def main(config, tr_stream, dev_stream):
# Create Theano variables
logger.info('Creating theano variables')
source_char_seq = tensor.lmatrix('source_char_seq')
source_sample_matrix = tensor.btensor3('source_sample_matrix')
source_char_aux = tensor.bmatrix('source_char_aux')
source_word_mask = tensor.bmatrix('source_word_mask')
target_char_seq = tensor.lmatrix('target_char_seq')
target_char_aux = tensor.bmatrix('target_char_aux')
target_char_mask = tensor.bmatrix('target_char_mask')
target_sample_matrix = tensor.btensor3('target_sample_matrix')
target_word_mask = tensor.bmatrix('target_word_mask')
target_resample_matrix = tensor.btensor3('target_resample_matrix')
target_prev_char_seq = tensor.lmatrix('target_prev_char_seq')
target_prev_char_aux = tensor.bmatrix('target_prev_char_aux')
target_bos_idx = tr_stream.trg_bos
target_space_idx = tr_stream.space_idx['target']
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'],
config['src_dgru_nhids'],
config['enc_nhids'],
config['src_dgru_depth'],
config['bidir_encoder_depth'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['trg_dgru_nhids'], config['trg_igru_nhids'],
config['dec_nhids'], config['enc_nhids'] * 2,
config['transition_depth'], config['trg_igru_depth'],
config['trg_dgru_depth'], target_space_idx,
target_bos_idx)
representation = encoder.apply(source_char_seq, source_sample_matrix,
source_char_aux, source_word_mask)
cost = decoder.cost(representation, source_word_mask, target_char_seq,
target_sample_matrix, target_resample_matrix,
target_char_aux, target_char_mask, target_word_mask,
target_prev_char_seq, target_prev_char_aux)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
for layer_n in range(config['src_dgru_depth']):
encoder.decimator.dgru.transitions[layer_n].weights_init = Orthogonal()
for layer_n in range(config['bidir_encoder_depth']):
encoder.children[
1 + layer_n].prototype.recurrent.weights_init = Orthogonal()
if config['trg_igru_depth'] == 1:
decoder.interpolator.igru.weights_init = Orthogonal()
else:
for layer_n in range(config['trg_igru_depth']):
decoder.interpolator.igru.transitions[
layer_n].weights_init = Orthogonal()
for layer_n in range(config['trg_dgru_depth']):
decoder.interpolator.feedback_brick.dgru.transitions[
layer_n].weights_init = Orthogonal()
for layer_n in range(config['transition_depth']):
decoder.transition.transitions[layer_n].weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(str(shape), count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(str(value.get_value().shape), name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Set extensions
logger.info("Initializing extensions")
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
train_monitor = CostCurve([cost, gradient_norm, step_norm],
config=config,
after_batch=True,
before_first_epoch=True,
prefix='tra')
extensions = [
train_monitor,
Timing(),
Printing(every_n_batches=config['print_freq']),
FinishAfter(after_n_batches=config['finish_after']),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
generated = decoder.generate(representation, source_word_mask)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[config['transition_depth']])
) # generated[transition_depth] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
transition_depth=config['transition_depth'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(source_char_seq,
source_sample_matrix,
source_char_aux,
source_word_mask,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main_sample():
pc = utils.read_ply('E:/Fac/BMC Master/Thesis/Models/bunny/reconstruction/plytest/bun_zipper.ply')
sample_points = Sampler.sample(pc, -1, 1000, sampling_algorithm=SampleAlgorithm.Uniform)
mlab.points3d(sample_points[:, 0], sample_points[:, 1], sample_points[:, 2], color=(1, 1, 1), mode='point')
mlab.show()
print 'done'
print(data) # ## Step 2 - Sample Points # # # You may want to limit the z-axis values. # In[70]: from sampling import Sampler # TODO: sample points randomly # then use KDTree to find nearest neighbor polygon # and test for collision num_samp = 500 sampler = Sampler(data) polygons = sampler._polygons nodes = sampler.sample(num_samp) print(len(nodes)) # ## Step 3 - Connect Nodes # # Now we have to connect the nodes. There are many ways they might be done, it's completely up to you. The only restriction being no edge connecting two nodes may pass through an obstacle. # # NOTE: You can use `LineString()` from the `shapely` library to create a line. Additionally, `shapely` geometry objects have a method `.crosses` which return `True` if the geometries cross paths, for instance your `LineString()` with an obstacle `Polygon()`! # In[71]: # TODO: connect nodes # Suggested method # 1) cast nodes into a graph called "g" using networkx
def create_prm(data):
sampler = Sampler(data)
polygons = sampler._polygons
nodes = sampler.sample(300)
g = create_graph(nodes, 10,polygons)
return nodes,g
def plan_path(self):
self.flight_state = States.PLANNING
print("Searching for a path ...")
TARGET_ALTITUDE = 5
SAFETY_DISTANCE = 5
self.target_position[2] = TARGET_ALTITUDE
# TODO: read lat0, lon0 from colliders into floating point values
filename = 'colliders.csv'
data = np.loadtxt(filename, delimiter=';,', dtype='str')[0].split(", ")
lat0, lon0 = [float(d.split(" ")[1]) for d in data]
# TODO: set home position to (lon0, lat0, 0)
self.set_home_position(lon0, lat0, 0)
print("self.local_position ", self.local_position)
# TODO: retrieve current global position
current_glbl_pos = [self._longitude, self._latitude,self._altitude]
print(current_glbl_pos)
print(self.global_position)
# TODO: convert to current local position using global_to_local()
current_lcl_pos = global_to_local(current_glbl_pos, self.global_home)
print("current_lcl_pos ", len(current_lcl_pos))
print('global home {0}, position {1}, local position {2}'.format(self.global_home, self.global_position,
self.local_position))
# Read in obstacle map
data = np.loadtxt('colliders.csv', delimiter=',', dtype='Float64', skiprows=2)
# Define a grid for a particular altitude and safety margin around obstacles
_, north_offset, east_offset = create_grid(data, TARGET_ALTITUDE, SAFETY_DISTANCE)
print("North offset = {0}, east offset = {1}".format(north_offset, east_offset))
# Define starting point on the grid (this is just grid center)
# grid_start = (-north_offset, -east_offset)
# TODO: convert start position to current position rather than map center
grid_start = (int(current_lcl_pos[0]), int(current_lcl_pos[1]))
# Set goal as some arbitrary position on the grid
# grid_goal = (-north_offset + 10, -east_offset + 10)
# Example: sampling 300 points and removing
# ones conflicting with obstacles.
# print("Start of sampling...")
sampler = Sampler(data)
polygons = sampler.polygons
nodes = sampler.sample(300)
nodes.insert(0, (grid_start[0], grid_start[1], 5.0))
# Build a graph using the sampled nodes
# and connect each nodes to its 10th closest nodes
print("Start graph building...")
t0 = time.time()
graph = create_graph(polygons, nodes, 11)
print('graph took {0} seconds to build'.format(time.time() - t0))
start = list(graph.nodes)[0]
len_path = 0
while len_path == 0:
k = np.random.randint(len(graph.nodes))
goal = list(graph.nodes)[k]
path, _ = a_star(graph, heuristic, start, goal)
len_path = len(path)
pruned_path = prune_path(path, polygons)
waypoints = [[int(p[0]), int(p[1]), int(p[2]), 0] for p in pruned_path]
self.waypoints = waypoints
# TODO: send waypoints to sim (this is just for visualization of waypoints)
self.send_waypoints()
import networkx as nx
import numpy as np
import matplotlib.pyplot as plt
from shapely.geometry import Polygon, Point, LineString
from queue import PriorityQueue
from sampling import Sampler
import numpy.linalg as LA
from sklearn.neighbors import KDTree
print(nx.__version__)
filename = 'colliders.csv'
data = np.loadtxt(filename, delimiter=',', dtype='Float64', skiprows=2)
sampler = Sampler(data)
def can_connect(n1, n2):
polygons = sampler._polygons
l = LineString([n1, n2])
for p in polygons:
minofn1n2 = min(n1[2], n2[2])
if p.crosses(1) and p.height >= minofn1n2:
return False
return True
def create_graph():
# print(polygons)
import sys
sys.path.append('src/utils')
from config_utils import settings
if __name__ == '__main__':
sequence_length = settings.model.sequence_length
vocabulary_size = settings.model.vocabulary_size
hidden_size = settings.model.hidden_size
print ('Creating model with configuration: {0}'.format(settings.model))
model = seq2seq_attention(sequence_length, vocabulary_size, hidden_size)
print ('Loading model weights from {0}'.format(settings.model.weights_path))
model.load_weights('models/seq2seq_weights.h5')
vocabulary_file = settings.data.vocabulary_path
with open(vocabulary_file, 'r') as handle:
vocabulary = json.load(handle)
sampler = Sampler(model, vocabulary, sequence_length)
while True:
question = raw_input('>>')
response = sampler.respond(question, greedy=True)
print (response)
for t in (.7, .8, .9):
response = sampler.respond(question, temperature=t)
print (response)
def main(config, tr_stream, dev_stream):
# Create Theano variables
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
# Test values
'''
theano.config.compute_test_value = 'warn'
source_sentence.tag.test_value = numpy.random.randint(10, size=(10, 10))
target_sentence.tag.test_value = numpy.random.randint(10, size=(10, 10))
source_sentence_mask.tag.test_value = \
numpy.random.rand(10, 10).astype('float32')
target_sentence_mask.tag.test_value = \
numpy.random.rand(10, 10).astype('float32')
sampling_input.tag.test_value = numpy.random.randint(10, size=(10, 10))
'''
# Construct model
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
# Initialize model
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
cg = ComputationGraph(cost)
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
print('Parameter shapes')
for shape, count in Counter(shapes).most_common():
print(' {:15}: {}'.format(shape, count))
# Set up training algorithm
algorithm = GradientDescent(cost=cost,
params=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Set up beam search and sampling computation graphs
sampling_representation = encoder.apply(sampling_input,
tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
samples, = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(ComputationGraph(
generated[1])) # generated[1] is the next_outputs
# Set up training model
training_model = Model(cost)
enc_param_dict = Selector(encoder).get_params()
dec_param_dict = Selector(decoder).get_params()
gh_model_name = '/data/lisatmp3/firatorh/nmt/wmt15/trainedModels/blocks/sanity/refGHOG_adadelta_40k_best_bleu_model.npz'
tmp_file = numpy.load(gh_model_name)
gh_model = dict(tmp_file)
tmp_file.close()
for key in enc_param_dict:
print '{:15}: {}'.format(enc_param_dict[key].get_value().shape, key)
for key in dec_param_dict:
print '{:15}: {}'.format(dec_param_dict[key].get_value().shape, key)
enc_param_dict['/bidirectionalencoder/embeddings.W'].set_value(
gh_model['W_0_enc_approx_embdr'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_state'].set_value(
gh_model['W_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_update'].set_value(
gh_model['G_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_reset'].set_value(
gh_model['R_enc_transition_0'])
enc_param_dict['/bidirectionalencoder/fwd_fork/fork_inputs.W'].set_value(
gh_model['W_0_enc_input_embdr_0'])
enc_param_dict['/bidirectionalencoder/fwd_fork/fork_inputs.b'].set_value(
gh_model['b_0_enc_input_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/fwd_fork/fork_update_inputs.W'].set_value(
gh_model['W_0_enc_update_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/fwd_fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_enc_reset_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_state'].set_value(
gh_model['W_back_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_update'].set_value(
gh_model['G_back_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_reset'].set_value(
gh_model['R_back_enc_transition_0'])
enc_param_dict['/bidirectionalencoder/back_fork/fork_inputs.W'].set_value(
gh_model['W_0_back_enc_input_embdr_0'])
enc_param_dict['/bidirectionalencoder/back_fork/fork_inputs.b'].set_value(
gh_model['b_0_back_enc_input_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/back_fork/fork_update_inputs.W'].set_value(
gh_model['W_0_back_enc_update_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/back_fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_back_enc_reset_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/lookupfeedbackwmt15/lookuptable.W'].set_value(
gh_model['W_0_dec_approx_embdr'])
#dec_param_dict['/decoder/sequencegenerator/readout/lookupfeedback/lookuptable.W'].set_value(gh_model['W_0_dec_approx_embdr'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/maxout_bias.b'].set_value(
gh_model['b_0_dec_hid_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax0.W'].set_value(
gh_model['W1_dec_deep_softmax']) # Missing W1
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax1.W'].set_value(
gh_model['W2_dec_deep_softmax'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax1.b'].set_value(
gh_model['b_dec_deep_softmax'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_states.W'].set_value(
gh_model['W_0_dec_hid_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_feedback.W'].set_value(
gh_model['W_0_dec_prev_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_weighted_averages.W'].set_value(
gh_model['W_0_dec_repr_readout'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_weighted_averages.b'].set_value(
gh_model['b_0_dec_repr_readout'])
dec_param_dict['/decoder/sequencegenerator/fork/fork_inputs.b'].set_value(
gh_model['b_0_dec_input_embdr_0'])
dec_param_dict['/decoder/sequencegenerator/fork/fork_inputs.W'].set_value(
gh_model['W_0_dec_input_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/fork/fork_update_inputs.W'].set_value(
gh_model['W_0_dec_update_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_dec_reset_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_inputs.W'].set_value(
gh_model['W_0_dec_dec_inputter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_inputs.b'].set_value(
gh_model['b_0_dec_dec_inputter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_update_inputs.W'].set_value(
gh_model['W_0_dec_dec_updater_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_update_inputs.b'].set_value(
gh_model['b_0_dec_dec_updater_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_reset_inputs.W'].set_value(
gh_model['W_0_dec_dec_reseter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_reset_inputs.b'].set_value(
gh_model['b_0_dec_dec_reseter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_state'].set_value(
gh_model['W_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_update'].set_value(
gh_model['G_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_reset'].set_value(
gh_model['R_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/state_trans/transform_states.W'].set_value(
gh_model['B_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/preprocess.W'].set_value(
gh_model['A_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/energy_comp/linear.W'].set_value(
gh_model['D_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder/state_initializer/linear_0.W'].set_value(
gh_model['W_0_dec_initializer_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder/state_initializer/linear_0.b'].set_value(
gh_model['b_0_dec_initializer_0'])
config['val_burn_in'] = -1
# Initialize main loop
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=[
FinishAfter(after_n_batches=1),
Sampler(model=search_model,
config=config,
data_stream=tr_stream,
every_n_batches=config['sampling_freq']),
BleuValidator(
sampling_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_eos_idx=config['src_eos_idx'],
trg_eos_idx=config['trg_eos_idx'],
before_training=True,
before_batch=True), #every_n_batches=config['bleu_val_freq']),
TrainingDataMonitoring([cost], after_batch=True),
#Plot('En-Fr', channels=[['decoder_cost_cost']],
# after_batch=True),
Printing(after_batch=True)
])
# Train!
main_loop.run()
# Simple synthetic data
points = np.random.random(
(10000, 2)
) # Generated data, the input data should be a numpy array with the shape (n, 2)
categories = np.random.randint(
0, 10, 10000
) # Generated label, multi-class sampling method would consider the label information as an reason to select or not select an item. It would be a np.zeros(n) as default.
# Datasets used in our study
all_data = np.load(os.path.join('data', 'abalone.npz'))
points, categories = all_data['positions'], all_data['labels']
print(points.shape, categories.shape)
sampler = Sampler()
sampler.set_data(
points, categories
) # For single-class sampling methods like random sampling, categories is not needed to be provided
sampling_method = RandomSampling # You can choose your desired sampling method.
rs_args = {
'sampling_rate':
0.3 # You can set the sampling ratio and other specific params for different sampling methods here.
}
sampler.set_sampling_method(
sampling_method,
**rs_args) # Set Random Sampling for the sampler with necessary params
sampled_point, sampled_category = sampler.get_samples(
) # Get the sampling result
import numpy as np
import matplotlib.pyplot as plt
from shapely.geometry import Polygon, Point, LineString
from queue import PriorityQueue
#Step 1 - Load Data
# This is the same obstacle data from the previous lesson.
filename = '../data/colliders.csv'
data = np.loadtxt(filename, delimiter=',', dtype='Float64', skiprows=2)
print(data)
#Step 2 - Sample Points in free space
from sampling import Sampler
sampler = Sampler(data)
polygons = sampler._polygons
# Example: sampling 100 points and removing
# ones conflicting with obstacles.
nodes = sampler.sample(300)
print(len(nodes))
#Step 3 - Connect Nodes
import numpy.linalg as LA
from sklearn.neighbors import KDTree
def can_connect(n1, n2):
l = LineString([n1, n2])
def read_ply(self, file_name):
num_samples = self.num_samples // len(self.files_list)
if self.file_index == len(self.files_list) - 1:
num_samples = num_samples + (self.num_samples - (num_samples * len(self.files_list)))
root, ext = os.path.splitext(file_name)
if not os.path.isfile(root + ".npy"):
ply = PlyData.read(file_name)
vertex = ply['vertex']
(x, y, z) = (vertex[t] for t in ('x', 'y', 'z'))
points = zip(x.ravel(), y.ravel(), z.ravel())
np.save(root + ".npy", points)
else:
points = np.load(root + ".npy")
if self.add_noise:
self.data = utils.add_noise(points, prob=self.noise_prob, factor=self.noise_factor)
else:
self.data = np.asarray(points)
#if self.data.shape[0] > 2e5:
# self.data, _ = Sampler.sample(self.data, -1, 2e5, sampling_algorithm=self.sampling_algorithm)
pc_diameter = utils.get_pc_diameter(self.data)
self.l = self.relL*pc_diameter
rot = utils.angle_axis_to_rotation(self.rotation_angle, self.rotation_axis)
self.data = utils.transform_pc(self.data, rot)
#plotutils.show_pc(self.data)
#mlab.show()
#TODO: better sampling
print "sampling file: ", file_name
self.samples, self.sample_indices = Sampler.sample(self.data, -1, num_samples, file_name=file_name, sampling_algorithm=self.sampling_algorithm)
self.samples = self.samples[0:num_samples]
self.sample_indices = self.sample_indices[0:num_samples]
self.tree = spatial.KDTree(self.data)
#TODO:Intergrate with num_samples for consistency
if self.filter_bad_samples:
temp_file_samples = 'temp/' + os.path.basename(file_name) + '_' + str(num_samples) + '_filter' + str(self.filter_threshold) + '.npy'
print 'samples file: ', temp_file_samples
if os.path.isfile(temp_file_samples):
self.sample_indices = np.load(temp_file_samples)
self.samples = self.data[self.sample_indices]
else:
self.samples, self.sample_indices = Sampler.sample(self.data, -1, num_samples*2, sampling_algorithm=self.sampling_algorithm)
self.samples = self.samples[0:num_samples*2]
self.sample_indices = self.sample_indices[0:num_samples*2]
sample_indices_temp = []
for idx in self.sample_indices:
if self.is_good_sample(self.data[idx], self.filter_threshold):
sample_indices_temp.append(idx)
if len(sample_indices_temp) >= num_samples:
break
assert (len(sample_indices_temp) >= num_samples)
self.sample_indices = np.asarray(sample_indices_temp[0:num_samples])
self.samples = self.data[self.sample_indices]
np.save(temp_file_samples, self.sample_indices)
#plotutils.show_pc(self.samples)
#mlab.show()
logging.basicConfig(filename='example.log',level=logging.DEBUG)
return self.data