def compute_weights(self):
'''
Computes all weights of G and of G_tilde and returns them as a tuple (w, w_tilde).
'''
self.w = np.zeros([len(self.images), self.num_columns,
self.K]) # node weights
self.w_tilde = np.zeros_like(self.w)
# fill in node weights
for t in range(len(self.images)):
for i in range(self.num_columns):
from_x = int(self.centers[t][0] +
self.col_vectors[i, 0] * self.min_radius[0])
from_y = int(self.centers[t][1] +
self.col_vectors[i, 1] * self.min_radius[1])
to_x = int(self.centers[t][0] +
self.col_vectors[i, 0] * self.max_radius[0])
to_y = int(self.centers[t][1] +
self.col_vectors[i, 1] * self.max_radius[1])
coords = bham.bresenhamline(np.array([[from_x, from_y]]),
np.array([[to_x, to_y]]))
num_pixels = len(coords)
for k in range(self.K):
start = int(k * float(num_pixels) / self.K)
end = max(start + 1, start + num_pixels / self.K)
self.w[t, i, k] = -1 * self.compute_weight_at(
t, coords[start:end])
for i in range(self.num_columns):
self.w_tilde[t, i, 0] = self.w[t, i, 0]
for k in range(1, self.K):
self.w_tilde[t, i,
k] = self.w[t, i, k] - self.w[t, i, k - 1]
def do_wave(BC, wave_origin, t0=0.0):
assert (type(wave_origin) is tuple) #will save heartach later
##set speed of sound
speed_sound = N * 340
#get dimensionality of space
dims = len(wave_origin)
#generate the time grid and set it to the initial time
time_grid = np.ones_like(BC, dtype=float) * t0
#generate basis with zero around the origin
basis = [
np.arange(-wave_origin[i], BC.shape[i] - wave_origin[i])
for i in range(dims)
]
#create a mesh
m_grid = np.meshgrid(*basis[::-1])
#stack it to another dimension
m_grid_ar = np.stack(m_grid)
#get disntance
dists = np.linalg.norm(m_grid_ar, axis=0)
#and use speed of sound to add on
time_grid = time_grid + (dists / speed_sound)
##now we worry about what get hit
visited = np.zeros_like(BC, dtype=np.bool_)
#generate the rays that land in boundary
#get coordinates of all the boundary
edge_idx = []
edge_idx.append([(0, i) for i in range(BC.shape[1])]) #top
edge_idx.append([(i, BC.shape[1]) for i in range(0, BC.shape[1])]) #right
edge_idx.append([(BC.shape[0], i) for i in range(BC.shape[1])]) #bottom
edge_idx.append([(i, -1) for i in range(BC.shape[1])]) #left
edge_idx = [x for y in edge_idx for x in y]
#loop through all the points
for idx in edge_idx:
#simplify the rise over run
sp = np.array(wave_origin)[np.newaxis, :]
idx = np.array(idx)[np.newaxis, :]
#init the move list
#it contains the id of the basis vector to take
move_list = bresenhamline(sp, idx)
move_list = [tuple(move_list[s, :]) for s in range(move_list.shape[0])]
for cur_pos in move_list:
try:
visited[cur_pos] = True
except:
break
if BC[cur_pos]:
#hit a boundary!!!
break
visited[wave_origin] = True
return visited, time_grid
def compute_weights(image,
center,
K,
max_radii,
col_vectors,
inverse_order=False,
min_radii=[0, 0, 0]):
'''
Computes all weights of G and of G_tilde and returns them as a tuple (w, w_tilde).
Parameters:
- image the image data (3d numpy array)
- center tupel with center coordinates (x, y, z)
- K number of nodes along each col_vector
- max_radii tupel containing number of pixels in (x, y, z) directions
- col_vectors vectors pointing on the unit sphere in all directions to be sampled
- inverse_order if true, the computed weights will be sorted such that they go from
the outside to the inside of the given image (along same vectors)
- min_radii defines the min_radii to start sampling the col_vectors at
'''
num_columns = len(col_vectors)
w = np.zeros([num_columns, K]) # node weights
w_tilde = np.zeros([num_columns, K])
# fill in node weights
for i in range(num_columns):
from_x = int(center[0] + col_vectors[i, 0] * min_radii[0])
from_y = int(center[1] + col_vectors[i, 1] * min_radii[1])
from_z = int(center[2] + col_vectors[i, 2] * min_radii[2])
to_x = int(center[0] + col_vectors[i, 0] * max_radii[0])
to_y = int(center[1] + col_vectors[i, 1] * max_radii[1])
to_z = int(center[2] + col_vectors[i, 2] * max_radii[2])
coords = bham.bresenhamline(np.array([[from_x, from_y, from_z]]),
np.array([[to_x, to_y, to_z]]))
num_pixels = len(coords)
for k in range(K):
start = int(k * float(num_pixels) / K)
end = max(start + 1, start + num_pixels / K)
w[i, k] = -1 * compute_weight(image, coords[start:end])
if inverse_order:
w = w[:, ::-1]
for i in range(num_columns):
w_tilde[i, 0] = w[i, 0]
for k in range(1, K):
w_tilde[i, k] = w[i, k] - w[i, k - 1]
return w, w_tilde
def get_column_flowvectors(self, oid, frame, column_id):
'''
Returns all flowvector for one column of the net surface graph.
The length of the returned list of 2-tupel (flow-vectors x,y) corresponds to the
number of pixels along the bresenham line from this objects seedpoint and the point
the net surface flow found the outline.
'''
flow = self.flows[frame]
point_c = self.object_seedpoints[oid][frame]
point_b = self.netsurfs[oid][frame].get_surface_point(column_id)
coords = bham.bresenhamline(np.array([point_c]), np.array([point_b]))
vectors = []
for c in coords:
fx, fy = flow[c[1], c[0]].T
vectors.append((fx, fy))
return vectors
def _process_ray(i, j, xray_img, ct_img3d, board_m, light_pos, light_pos_m):
# Calculate the voxels along the line segment connecting the light source with the sensor board pixel
px3d_m = board_m.pixel_to_3d_hg([i, j])
# TODO efficiency: intersect with the CT volume before running Bresenham to avoid making redundant calculations
voxel_list = bresenhamline(light_pos_m[np.newaxis, :3], px3d_m[:3], max_iter=-1)
# (assuming the entire CT volume is between the light source and the board)
valid_voxels = np.logical_and(voxel_list >= 0, voxel_list < ct_img3d.shape).all(axis=1)
voxel_list = voxel_list[valid_voxels].astype(np.int)
# Calculate the distance the ray travels per voxel
px3d = board_m.pixel_to_3d_hg([i, j])
ray_direction_m = px3d_m[:3] - light_pos_m[:3]
ray_direction = px3d[:3] - light_pos
principal_dimension = np.argmax(np.abs(ray_direction_m))
# The distance is 1/cos(angle between the line segment and the principal direction)
distance_per_voxel = np.abs(norm(ray_direction) /
ray_direction[principal_dimension])
# Sum the voxels along the line segment
voxel_values = ct_img3d[tuple(voxel_list.T)] # index into the CT volume
xray_img[i, j] = np.sum(voxel_values) * distance_per_voxel
def compute_weights(self, inverse_order=False):
'''
Computes all weights of G and of G_tilde and returns them as a tuple (w, w_tilde).
'''
assert not self.image is None
self.w = np.zeros([self.num_columns, self.K]) # node weights
self.w_tilde = np.zeros([self.num_columns, self.K])
# fill in node weights
for i in range(self.num_columns):
from_x = int(self.center[0] +
self.col_vectors[i, 0] * self.min_radii[0])
from_y = int(self.center[1] +
self.col_vectors[i, 1] * self.min_radii[1])
from_z = int(self.center[2] +
self.col_vectors[i, 2] * self.min_radii[2])
to_x = int(self.center[0] +
self.col_vectors[i, 0] * self.max_radii[0])
to_y = int(self.center[1] +
self.col_vectors[i, 1] * self.max_radii[1])
to_z = int(self.center[2] +
self.col_vectors[i, 2] * self.max_radii[2])
coords = bham.bresenhamline(np.array([[from_x, from_y, from_z]]),
np.array([[to_x, to_y, to_z]]))
num_pixels = len(coords)
for k in range(self.K):
start = int(k * float(num_pixels) / self.K)
end = max(start + 1, start + num_pixels / self.K)
self.w[i, k] = -1 * self.compute_weight_at(coords[start:end])
if inverse_order:
self.w = self.w[:, ::-1]
for i in range(self.num_columns):
self.w_tilde[i, 0] = self.w[i, 0]
for k in range(1, self.K):
self.w_tilde[i, k] = self.w[i, k] - self.w[i, k - 1]
centers = [ndimage.center_of_mass(labels==i) for i in range(1, num_maxima+1)]
radii = [data[center] for center in centers]
return np.array(centers), np.array(radii)
def rasterize():
pass
solid_nodes, solid_edges = map(np.array, load_pdb('CHA'))
solid_nodes -= solid_nodes.min(axis=0)
solid_nodes *= 4
coord_pair = solid_nodes[solid_edges]
discretized = []
for a, b in coord_pair:
point_list = bresenham.bresenhamline(np.atleast_2d(a), b, -1).astype(int).tolist()
discretized.extend([tuple(point) for point in point_list])
array = np.array(discretized)
size = array.max(axis=0) - array.min(axis=0) + 1
canvas = np.ones(size, dtype=bool)
offset = array.min(axis=0)
for idx, _ in np.ndenumerate(canvas):
if idx in discretized:
canvas[idx] = 0
# mpl_tools.visualize(canvas)
centers, radii = extract_spheres(canvas)
def calcWork(v,
w,
currentsGrid_u,
currentsGrid_v,
targetSpeed_mps,
geotransform=None,
timeIn=0,
interval=3600,
pixelsize_m=1,
distMeas="haversine"):
'''
This function calculates the work applied by a vehicle
between two points, given rasters with u, v force components
v: start point (row, col) in force rasters
w: end point (row, col) in force rasters
currentsGrid_u: 3D Raster of forces u components.
[time index, row, column] = u
currentsGrid_v: 3D Raster of forces v components.
[time index, row, column] = v
'''
elapsed = float(timeIn)
(index, rem) = divmod(elapsed, interval)
index = min(floor(index), currentsGrid_u.shape[0] - 2)
if v != w:
# Heading
vecs = (w[1] - v[1], w[0] - v[0])
dotprod = vecs[0] * 1 + vecs[1] * 0
maga = pow(vecs[0] * vecs[0] + vecs[1] * vecs[1], 0.5)
heading_rad = 0.0
if (maga != 0):
costheta = dotprod / maga
heading_rad = acos(costheta)
# Distance
rows = currentsGrid_u.shape[0]
if distMeas == "haversine":
v_latlon = grid2world(v[0], v[1], geotransform, rows)
w_latlon = grid2world(w[0], w[1], geotransform, rows)
hdist = haversine.haversine(v_latlon, w_latlon) * 1000
elif distMeas == "euclidean-scaled":
hdist = pow(
(pow(v[0] - w[0], 2) + pow(v[1] - w[1], 2)), 0.5) * pixelsize_m
elif distMeas == "euclidean":
hdist = pow((pow(v[0] - w[0], 2) + pow(v[1] - w[1], 2)), 0.5)
# Work
work = 0
print(v, w)
b = list(bresenham.bresenhamline(np.array([v]), np.array([w])))
hdist_ = hdist / len(b)
for p in b[:]:
xA = targetSpeed_mps * cos(heading_rad)
yA = targetSpeed_mps * sin(heading_rad)
# Interpolate between nearest time's currents
if currentsGrid_u.shape[0] > 1:
ua_ = currentsGrid_u[index, p[0], p[1]] * cos(
currentsGrid_v[index, p[0], p[1]])
va_ = currentsGrid_u[index, p[0], p[1]] * sin(
currentsGrid_v[index, p[0], p[1]])
try:
ub_ = currentsGrid_u[index + 1, p[0], p[1]] * cos(
currentsGrid_v[index + 1, p[0], p[1]])
vb_ = currentsGrid_u[index + 1, p[0], p[1]] * sin(
currentsGrid_v[index + 1, p[0], p[1]])
except:
ub_ = currentsGrid_u[index, p[0], p[1]] * cos(
currentsGrid_v[index + 1, p[0], p[1]])
vb_ = currentsGrid_u[index, p[0], p[1]] * sin(
currentsGrid_v[index + 1, p[0], p[1]])
u_ = ua_ * (1 - (rem / interval)) + ub_ * ((rem / interval))
v_ = va_ * (1 - (rem / interval)) + vb_ * ((rem / interval))
uv_ = np.array([u_, v_])
cmag = np.sqrt(uv_.dot(uv_))
cdir = atan2(v_, u_)
# Convert to knots for sanity check
#knots = cmag * 1.94384
#print(knots)
else:
# Static currents -> can't interpolate in time
cmag = currentsGrid_u[0, p[0], p[1]]
cdir = currentsGrid_v[0, p[0], p[1]]
xB = cmag * cos(cdir)
yB = cmag * sin(cdir)
# Calculate applied force,
# given desired and environment forces
dV = (xB - xA, yB - yA)
magaDV = pow(dV[0] * dV[0] + dV[1] * dV[1], 0.5)
dirDV = 0.0
dotprod = dV[0] * 1 + dV[1] * 0
if magaDV != 0:
costheta = dotprod / magaDV
dirDV = acos(costheta)
work += magaDV * hdist_
# Update time
elapsed += hdist_ / targetSpeed_mps
(index, rem) = divmod(elapsed, interval)
index = min(floor(index), currentsGrid_u.shape[0] - 2)
else:
work = 0
return work, elapsed
return np.array(centers), np.array(radii)
def rasterize():
pass
solid_nodes, solid_edges = map(np.array, load_pdb('CHA'))
solid_nodes -= solid_nodes.min(axis=0)
solid_nodes *= 4
coord_pair = solid_nodes[solid_edges]
discretized = []
for a, b in coord_pair:
point_list = bresenham.bresenhamline(np.atleast_2d(a), b,
-1).astype(int).tolist()
discretized.extend([tuple(point) for point in point_list])
array = np.array(discretized)
size = array.max(axis=0) - array.min(axis=0) + 1
canvas = np.ones(size, dtype=bool)
offset = array.min(axis=0)
for idx, _ in np.ndenumerate(canvas):
if idx in discretized:
canvas[idx] = 0
# mpl_tools.visualize(canvas)
centers, radii = extract_spheres(canvas)
def calcWork(path,
n,
regionGrid,
targetSpeed_mps=1,
currentsGrid_u=None,
currentsGrid_v=None,
currentsTransform=None,
regionTransform=None,
rewardGrid=None,
timeIn=0,
interval=3600,
pixelsize_m=1):
'''
This function calculates cost to move vehicle along path
'''
# Initialize costs: distance, obstacles, work, reward
costs = np.zeros(4)
# Get all cells along path
cells_ = np.array([
bresenham.bresenhamline(np.array([path[i]]), np.array([path[i + 1]]))
for i, p in enumerate(path[:-1])
])
# Number of cells in each path segment
lens = np.array([len(c) for c in cells_])
cells = np.vstack(cells_)
# Obstacle penalty
costs[1] = np.sum((grid[cells[:, 0], cells[:, 1]]))
# Reward
if rewardGrid is not None:
costs[3] = np.sum((rewardGrid[cells[:, 0], cells[:, 1]]))
# Points in (lon, lat)
lonlat = np.vstack(
(regionTransform[1] * path[:, 1] + regionTransform[2] * path[:, 0] +
regionTransform[0], regionTransform[4] * path[:, 1] +
regionTransform[5] * path[:, 0] + regionTransform[3])).T
# Haversine distances, converted from km to meters
hdists = haversine_np(lonlat[:-1][:, 0], lonlat[:-1][:, 1],
lonlat[1:][:, 0], lonlat[1:][:, 1]) * 1000
# Distance penalty
costs[0] = sum(hdists)
# If no water currents, return now
if currentsGrid_u is None:
costs[2] = costs[0]
return costs[2], costs[0], costs[1], costs[3]
# Calculate headings
vecs = np.flip(path[1:] - path[:-1], axis=1)
dotprod = vecs[:, 0]
maga = np.sqrt(vecs[:, 0] * vecs[:, 0] + vecs[:, 1] * vecs[:, 1])
costheta = vecs[:, 0] / maga # Should check if div by 0?
heading_rad = np.arccos(costheta)
# Approx distance of each cell
hdists_ = hdists / lens
# Expand path-length variables to cell-length
hdists_e = np.hstack(
[np.zeros(c.shape[0]) + hdists_[i] for i, c in enumerate(cells_)])
# Estimate elapsed time to reach each cell
ela = [(hdists_e[i] / targetSpeed_mps) * i for i in range(cells.shape[0])]
# Vehicle target velocities
xA = targetSpeed_mps * np.cos(heading_rad)
yA = targetSpeed_mps * np.sin(heading_rad)
xA = np.hstack(
[np.zeros(c.shape[0]) + xA[i] for i, c in enumerate(cells_)])
yA = np.hstack(
[np.zeros(c.shape[0]) + yA[i] for i, c in enumerate(cells_)])
# Calculate indices for accessing water current grids
rem_idx = np.divmod(ela, interval)
p = np.array(cells[:]).astype("int")
idx_1 = np.minimum(rem_idx[0], currentsGrid_u.shape[0] - 2).astype("int")
idx_2 = idx_1 + 1
# Interpolate all needed water currents
uv_ = np.array(
(
(currentsGrid_u[idx_1, p[:, 0], p[:, 1]] * np.cos(currentsGrid_v[idx_1, p[:, 0], p[:, 1]])) \
* (1 - (rem_idx[1] / interval)) + \
(currentsGrid_u[idx_2, p[:, 0], p[:, 1]]) * np.cos((currentsGrid_v[idx_2, p[:, 0], p[:, 1]])) \
* (rem_idx[1] / interval),
(currentsGrid_u[idx_1, p[:, 0], p[:, 1]]) * np.sin((currentsGrid_v[idx_1, p[:, 0], p[:, 1]])) \
* (1 - (rem_idx[1] / interval)) + \
(currentsGrid_u[idx_2, p[:, 0], p[:, 1]]) * np.sin((currentsGrid_v[idx_2, p[:, 0], p[:, 1]])) \
* (rem_idx[1] / interval)
)
).T
# Calculate applied force required by vehicle to maintain target velocity
cmag = np.sqrt(np.sum(uv_ * uv_, axis=1))
cdir = np.arctan2(uv_[:, 1], uv_[:, 0])
xB = cmag * np.cos(cdir)
yB = cmag * np.sin(cdir)
dV = np.array((xB - xA, yB - yA)).T
magaDV = np.power(dV[:, 0] * dV[:, 0] + dV[:, 1] * dV[:, 1], 0.5)
dotprod = dV[:, 0]
costheta = dotprod / magaDV # Prob need to check for div by 0 ??
dwork = magaDV * hdists_e
# Work penalty
costs[2] = np.sum(dwork)
return costs[2], costs[0], costs[1], costs[3]
