def _idw_interpolate(sensor_data, target_grid_len):
'''
Args:
sensor_data -- np.2darray -- Rx data from a single Tx
target_grid_len -- int -- the target grid length for the Rx
Return:
np.1darray -- interpolated, shape = target_grid_len**2
'''
pre_gl = len(sensor_data)
factor = int(target_grid_len / pre_gl)
interpolate = np.zeros((target_grid_len, target_grid_len))
for new_x in range(target_grid_len):
for new_y in range(target_grid_len):
if new_x % factor == 0 and new_y % factor == 0 and sensor_data[
new_x // factor][
new_y // factor] != 0: # don't need to interpolate
interpolate[new_x][new_y] = sensor_data[new_x //
factor][new_y //
factor]
else:
v_x, v_y = new_x / factor, new_y / factor # virtual point in the coarse grid
points = [] # pick some points from the coarse grid
edge = int(8 / factor)
for pre_x in range(math.floor(v_x - edge),
math.ceil(v_x + 1) + edge):
if pre_x < 0 or pre_x >= pre_gl: # the x range
continue
for pre_y in range(math.floor(v_y - edge),
math.ceil(v_y + 1) + edge):
if pre_y >= 0 and pre_y < pre_gl and sensor_data[
pre_x][pre_y] != 0:
points.append((pre_x, pre_y,
distance((v_x, v_y),
(pre_x, pre_y))))
points = sorted(points,
key=lambda tup: tup[2]) # sort by distance
threshold = min(MultiIntepolate.NEIGHBOUR_NUM, len(points))
weights = np.zeros(threshold)
for i in range(threshold):
point = points[i]
dist = distance((v_x, v_y), point)
weights[i] = (
1. / dist
)**2 # inverse weighted distance or inverse weighted square
weights /= np.sum(weights) # normalize them
idw = 0
for i in range(threshold):
w = weights[i]
pre_rss = sensor_data[points[i][0]][points[i][1]]
idw += w * pre_rss
interpolate[new_x][new_y] = idw
return interpolate.reshape(target_grid_len * target_grid_len)
def _interpolate_idw(pre_inter, factor):
'''Fix one transmitters, interpolate the sensors
Args:
pre_inter -- np.1darray -- pre interpolated, shape = pre_gl*pre_gl
factor -- int
Return:
np.1darray -- interpolated, shape = gre_gl*gre_gl*factor*factor
'''
pre_gl = int(math.sqrt(
len(pre_inter))) # previous grid length (coarse grid)
pre_inter = pre_inter.reshape((pre_gl, pre_gl))
new_gl = pre_gl * factor # new grid length (find grid)
inter = np.zeros((new_gl, new_gl))
for new_x in range(new_gl):
for new_y in range(new_gl):
if new_x % factor == 0 and new_y % factor == 0: # don't need to interpolate
inter[new_x][new_y] = pre_inter[new_x // factor][new_y //
factor]
else:
v_x, v_y = float(new_x) / factor, float(
new_y
) / factor # virtual point in the coarse grid / real point in the fine grid
# pick some close points from the coarse grid
points = []
for pre_x in range(math.floor(v_x - 1),
math.ceil(v_x + 1) + 1):
for pre_y in range(math.floor(v_y - 1),
math.ceil(v_y + 1) + 1):
if pre_x >= 0 and pre_x < pre_gl and pre_y >= 0 and pre_y < pre_gl:
points.append((pre_x, pre_y,
distance((v_x, v_y),
(pre_x, pre_y))))
points = sorted(points,
key=lambda tup: tup[2]) # sort by distance
threshold = min(NEIGHBOR_NUM, len(points))
weights = np.zeros(threshold)
for i in range(threshold):
point = points[i]
dist = distance((v_x, v_y), point)
weights[i] = (
1. / dist
)**2 # inverse weighted distance or inverse weighted square
weights /= np.sum(weights) # normalize them
idw = 0
for i in range(threshold):
w = weights[i]
pre_rss = pre_inter[points[i][0]][points[i][1]]
idw += w * pre_rss
inter[new_x][new_y] = idw
return inter.reshape(new_gl * new_gl)
def get_log10_weight(to_tx_dists, tx, rx0):
'''get the weights from project_dists
Args:
project_dists -- np.1darray
tx -- (float, float)
rx0 -- (float, float) -- location to be interpolated
Return:
np.1darray
'''
for i, dist in enumerate(to_tx_dists):
if dist == 0:
to_tx_dists[
i] = IpsnInterpolate.ILDW_DIST # this value can tweak
log10_dist_to_tx = np.log10(to_tx_dists)
reference_dist = np.log10(distance(tx, rx0))
log10_dist_to_rx0 = log10_dist_to_tx - reference_dist
log10_dist_to_rx0 = np.absolute(log10_dist_to_rx0)
weight = np.zeros(len(log10_dist_to_rx0))
for i, dist in enumerate(log10_dist_to_rx0):
if dist > 0:
weight[i] = (1. / dist)
maxx = max(weight)
for i, w in enumerate(weight):
if w == 0:
weight[i] = 2 * maxx if maxx > 0 else 1
return weight
def find_d0_pd0(self, num=4):
'''find some largest RSS, average them as the P(d0) and d0
Args:
num (int):
'''
means = []
dic = defaultdict(list)
for i in range(len(self.locations)):
for j in range(len(self.locations)):
if i == j:
continue
dist = distance(self.locations[i], self.locations[j])
dic[dist].append((i, j, self.means[i, j]))
means.append(self.means[i][j])
# X, Y = [], []
threshold = sorted(np.array(means))[-num]
d0, pd0 = [], []
for key, vals in sorted(
dic.items()): # distance --> [ (loc1, loc2) ...]
for val in vals:
i, j, mean = val
if mean >= threshold:
d0.append(key)
pd0.append(mean)
if len(d0) == num:
return np.array(d0).mean(), np.array(pd0).mean()
def _ildw(self, tx_data, tx_1dindex):
'''Interpolate one Tx
Args:
tx_data -- np.1darray -- contains zeros values
Return:
tx_data -- np.1darray -- the zero values are filled up
'''
tx = (tx_1dindex // self.full_grid_len, tx_1dindex % self.full_grid_len
) # ildw requires the location of the transmitter
grid_pathloss = tx_data.reshape(self.full_grid_len, self.full_grid_len)
grid_interpolate = np.copy(grid_pathloss)
# find the zero elements and impute them
for x in range(grid_pathloss.shape[0]):
for y in range(grid_pathloss.shape[1]):
if grid_pathloss[x][
y] != 0.0: # skip the ones that don't need to interpolate
continue
points = []
d = self.range
for i in range(x - d, x + d + 1):
for j in range(y - d, y + d + 1):
if i < 0 or i >= grid_pathloss.shape[
0] or j < 0 or j >= grid_pathloss.shape[1]:
continue
if grid_pathloss[i][
j] == 0.0: # only use the known point to interpolate
continue
dist = distance((x, y), (i, j))
points.append((i, j, dist))
points = sorted(points, key=lambda tup: tup[2])
threshold = min(IpsnInterpolate.NEIGHBOR_NUM, len(points))
weights = np.zeros(threshold)
dist_to_tx = np.zeros(threshold)
for i in range(threshold):
nei = (points[i][0], points[i][1])
dist_to_tx[i] = distance(tx, nei)
rx0 = (x, y)
weights = IpsnInterpolate.get_log10_weight(dist_to_tx, tx, rx0)
weights /= np.sum(weights)
idw_pathloss = 0
for i in range(threshold):
w = weights[i]
rss = grid_pathloss[points[i][0]][points[i][1]]
idw_pathloss += w * rss
grid_interpolate[x][y] = idw_pathloss
return grid_interpolate.reshape(self.full_grid_len *
self.full_grid_len)
def combine_sensor_data(self):
'''Combine sensor data from multiply granularity. each granularity has full data in 1 dimension (from the SPLAT!)
'''
grid_lens = sorted(TxMultiGran.RX_CELL_LEN.keys()) # [5, 10, 20]
if len(grid_lens) != len(Global.GRAN_LEVEL):
print((self.x, self.y),
'length of grid_lens and granularity level doesn\'t match !')
return
# map granularity level to grid length # {1200 --> 5, 400 --> 10, 0 --> 20}
gran_level2grid_len = {}
for gran_level, grid_len in zip(reversed(Global.GRAN_LEVEL),
grid_lens):
gran_level2grid_len[gran_level] = grid_len
# map grid length to ratio. the ratio of finest_rx_grid_len to different grid_len
finest_rx_grid_len = grid_lens[-1]
grid_len2ratio = {}
for grid_len in grid_lens:
grid_len2ratio[grid_len] = int(
finest_rx_grid_len / grid_len) # {5 --> 4, 10 --> 2, 20 --> 1}
self.sensor_data = np.zeros(
(finest_rx_grid_len, finest_rx_grid_len)) # fill up this array
t_x = self.x * grid_len2ratio[TxMultiGran.TX_GRID_LEN]
t_y = self.y * grid_len2ratio[TxMultiGran.TX_GRID_LEN]
for x in range(finest_rx_grid_len):
for y in range(finest_rx_grid_len):
dist = distance((x, y), (t_x, t_y)) * TxMultiGran.RX_CELL_LEN[
finest_rx_grid_len] # distance in meters
gran_level = self.get_gran_level(Global.GRAN_LEVEL, dist)
grid_len = gran_level2grid_len[gran_level]
ratio = grid_len2ratio[grid_len]
if x % ratio == 0 and y % ratio == 0: # granularity control by this ratio
if TxMultiGran.TX_GRID_LEN == grid_lens[0]:
# when the Tx granularity equals to the lowest Rx granularity
gran_data = self.granularity_data[grid_len]
gran_grid_len = int(math.sqrt(len(gran_data)))
index = x // ratio * gran_grid_len + y // ratio
self.sensor_data[x][y] = gran_data[index]
else:
# when the Tx granularity not equal to the lowest Rx granularity, only use the finest granularity
gran_data = self.granularity_data[finest_rx_grid_len]
gran_grid_len = int(math.sqrt(len(gran_data)))
index = x * gran_grid_len + y
self.sensor_data[x][y] = gran_data[index]
if self.debug:
num_wireless_link = np.count_nonzero(self.sensor_data)
TxMultiGran.NUM_WIRELESS_LINK += num_wireless_link
print('Tx = ({}, {})'.format(self.x, self.y),
'; number of wireless links to sensors', num_wireless_link,
'; Total = {}'.format(TxMultiGran.NUM_WIRELESS_LINK))
def predict(self, tx, rx):
'''Predice the receiver's RSS at rx, when the transmitter is at tx
Args:
tx (tuple(float, float))
rx (tuple(float, float))
'''
dist = distance(tx, rx)
tx = Point(tx[0], tx[1])
rx = Point(rx[0], rx[1])
nW = self.wall.count_intersect(tx, rx)
nW = nW if nW <= self.C else self.C
# offset = 0
# if dist < 2 and nW == 0:
# offset = 6.37
return self.p_d0 - 10 * self.n * math.log10(
dist / self.d0) - nW * self.waf + self.intercept
def compute_errors(self, inter_data, dist_close, dist_far):
'''Compute the interpolation errors
Args:
inter_data -- np.2darray
dist_close -- int
dist_far -- int -- distance threshold between tx and rx
coarse_gran -- int
'''
grid_is_inter = np.zeros(
(self.full_grid_len, self.full_grid_len)
) # the places for sensors in coarse grid, no interpolation happening
for x in self.index:
for y in self.index:
grid_is_inter[x, y] = 1.
size = len(inter_data)
errors_all = [] # errors for all tx-rx
errors_close = [] # errors for tx-rx with small distance
errors_far = [] # errors for tx-rx with large distance
for i in range(size):
tx = (i // self.full_grid_len, i % self.full_grid_len)
for j in range(size):
rx = (j // self.full_grid_len, j % self.full_grid_len)
if grid_is_inter[rx[0]][rx[1]] == 1. or i == j:
continue
error = inter_data[i][j] - self.full_data[i][j]
dist = distance(tx, rx)
if dist < dist_close:
errors_close.append(error)
elif dist > dist_far:
errors_far.append(error)
errors_all.append(error)
mean_error = np.mean(errors_all)
mean_absolute_error = np.mean(np.absolute(errors_all))
mean_error_close = np.mean(errors_close)
mean_absolute_error_close = np.mean(np.absolute(errors_close))
mean_error_far = np.mean(errors_far)
mean_absolute_error_far = np.mean(np.absolute(errors_far))
std = np.std(np.absolute(errors_all))
std_close = np.std(np.absolute(errors_close))
std_far = np.std(np.absolute(errors_far))
return Output(None, mean_error, mean_absolute_error, mean_error_close, mean_absolute_error_close, \
mean_error_far, mean_absolute_error_far, std, std_close, std_far)
def _idw(self, tx_data):
'''Interpolate one Tx
Args:
tx_data -- np.1darray -- contains zeros values
Return:
tx_data -- np.1darray -- the zero values are filled up
'''
grid_pathloss = tx_data.reshape(self.full_grid_len, self.full_grid_len)
grid_interpolate = np.copy(grid_pathloss)
# find the zero elements and impute them
for x in range(grid_pathloss.shape[0]):
for y in range(grid_pathloss.shape[1]):
if grid_pathloss[x][
y] != 0.0: # skip the ones that don't need to interpolate
continue
points = []
d = self.range
for i in range(x - d, x + d + 1):
for j in range(y - d, y + d + 1):
if i < 0 or i >= grid_pathloss.shape[
0] or j < 0 or j >= grid_pathloss.shape[1]:
continue
if grid_pathloss[i][
j] == 0.0: # only use the known point to interpolate
continue
dist = distance((x, y), (i, j))
points.append((i, j, dist))
points = sorted(points, key=lambda tup: tup[2])
threshold = min(IpsnInterpolate.NEIGHBOR_NUM, len(points))
weights = np.zeros(threshold)
for i in range(threshold):
dist = points[i][2]
weights[i] = (1. / dist)**IpsnInterpolate.IDW_EXPONENT
weights /= np.sum(weights)
idw_pathloss = 0
for i in range(threshold):
w = weights[i]
rss = grid_pathloss[points[i][0]][points[i][1]]
idw_pathloss += w * rss
grid_interpolate[x][y] = idw_pathloss
return grid_interpolate.reshape(self.full_grid_len *
self.full_grid_len)
def main0():
'''Read the Utah data, location transform
'''
means, stds, locations, wall = read_utah_data(path='dataUtah')
lt = LocationTransform(locations, cell_len=1)
print(means)
print(stds)
print(locations)
print(wall)
print(lt)
errors = []
for real_loc in lt.real_location:
gridcell = lt.real_2_gridcell(real_loc)
real_loc2 = lt.gridcell_2_real(gridcell)
error = distance(real_loc, real_loc2)
errors.append(error)
print('({:5.2f}, {:5.2f}) -> ({:2d}, {:2d}) -> ({:5.2f}, {:5.2f}); error = {:3.2f}'.format(real_loc[0], real_loc[1], gridcell[0], gridcell[1], real_loc2[0], real_loc2[1], error))
plot_cdf(errors)
def regresssion_n_waf(self):
'''Do a regression to get the n and waf
'''
X, y = [], []
for i in range(len(self.locations)):
for j in range(len(self.locations)):
if i == j:
continue
dist = distance(self.locations[i], self.locations[j])
tx = Point(self.locations[i][0], self.locations[i][1])
rx = Point(self.locations[j][0], self.locations[j][1])
nW = self.wall.count_intersect(tx, rx)
nW = nW if nW <= self.C else self.C
y.append(self.means[i][j] - self.p_d0)
# y.append(self.means[i][j])
X.append([math.log10(dist / self.d0), nW])
# X.append([math.log10(dist), nW])
reg = LinearRegression(fit_intercept=True).fit(X, y)
print('Regression score:', reg.score(X, y))
return reg.coef_[0], reg.coef_[1], reg.intercept_
def main1():
mean, stds, locations, wall = read_utah_data()
# print(mean)
# print(stds)
# print(locations)
lt = LocationTransform(locations, cell_len=1)
# print(lt)
num = len(mean)
with open('dataUtah/means.txt', 'w') as f:
for i in range(num):
for j in range(num):
f.write('{:6.2f} '.format(mean[i][j]))
f.write('\n')
with open('dataUtah/stds.txt', 'w') as f:
for std in stds:
f.write('{:.3f}\n'.format(std))
with open('dataUtah/locations.txt', 'w') as f:
for i in range(num):
cell = lt.grid_location[i]
real_loc = lt.real_location[i]
f.write('{:3d} - '.format(cell[0] * lt.grid_len + cell[1]))
f.write('({:2d}, {:2d}) - '.format(cell[0], cell[1]))
f.write('({:7.4f}, {:7.4f})\n'.format(real_loc[0], real_loc[1]))
errors = []
for real_loc in lt.real_location:
gridcell = lt.real_2_gridcell(real_loc)
real_loc2 = lt.gridcell_2_real(gridcell)
error = distance(real_loc, real_loc2)
errors.append(error)
print(
'({:5.2f}, {:5.2f}) -> ({:2d}, {:2d}) -> ({:5.2f}, {:5.2f}); error = {:3.2f}'
.format(real_loc[0], real_loc[1], gridcell[0], gridcell[1],
real_loc2[0], real_loc2[1], error))
plot_cdf(errors)
def correct(self):
X, Y = [], []
for i in range(len(self.locations)):
for j in range(len(self.locations)):
if i == j:
continue
dist = distance(self.locations[i], self.locations[j])
tx = Point(self.locations[i][0], self.locations[i][1])
rx = Point(self.locations[j][0], self.locations[j][1])
nW = self.wall.count_intersect(tx, rx)
nW = nW if nW <= self.C else self.C
X.append(dist)
Y.append(self.means[i][j] + nW * self.waf)
# Y.append(self.means[i][j])
plt.rcParams['font.size'] = 20
plt.figure(figsize=(10, 10))
plt.scatter(np.log10(X), Y)
# plt.scatter(X, Y)
plt.title('Wall Correction')
plt.xlabel('Log Distance (m)')
plt.ylabel('RSS (dBm)')
plt.ylim([-85, -30])
plt.savefig('visualize/RSS-logdist-wallcorrect.png')
def _idw_interpolate_2(pre_inter, target_grid_len, tx, tx_pl):
'''Fix one transmitters, interpolate the sensors
Args:
pre_inter -- np.1darray -- pre interpolated, shape = pre_gl*pre_gl
target_grid_len -- int -- the Tx that needs to interpolate Rx
factor -- int
Return:
np.1darray -- interpolated, shape = gre_gl*gre_gl*factor*factor
'''
pre_gl = int(math.sqrt(
len(pre_inter))) # previous grid length (coarse grid)
pre_inter = pre_inter.reshape((pre_gl, pre_gl))
factor = int(target_grid_len / pre_gl)
tx_x, tx_y = tx // target_grid_len, tx % target_grid_len
inter = np.zeros((target_grid_len, target_grid_len))
for new_x in range(target_grid_len):
for new_y in range(target_grid_len):
if new_x % factor == 0 and new_y % factor == 0: # don't need to interpolate
inter[new_x][new_y] = pre_inter[new_x // factor][new_y //
factor]
else:
v_x, v_y = float(new_x) / factor, float(
new_y
) / factor # virtual point in the coarse grid / real point in the fine grid
# pick some close points from the coarse grid
points = []
for pre_x in range(math.floor(v_x - 1),
math.ceil(v_x + 1) + 1):
for pre_y in range(math.floor(v_y - 1),
math.ceil(v_y + 1) + 1):
if pre_x >= 0 and pre_x < pre_gl and pre_y >= 0 and pre_y < pre_gl:
points.append(
(pre_x, pre_y,
distance((v_x, v_y), (pre_x, pre_y)))
) # the distance between the virtual Rx in the coarse grid and coarse Rx
dist_to_tx_fine = distance(
(new_x, new_y),
(tx_x, tx_y)) # distance in the fine grid
if dist_to_tx_fine < 4:
dist_to_tx_coarse = distance(
(v_x, v_y), (tx_x / factor, tx_y / factor))
points.append(
(tx_x / factor, tx_y / factor, dist_to_tx_coarse)
) # the additional Rx at the same location as the Tx
points = sorted(points,
key=lambda tup: tup[2]) # sort by distance
threshold = min(MultiIntepolate.NEIGHBOUR_NUM, len(points))
weights = np.zeros(threshold)
for i in range(threshold):
point = points[i]
dist = distance((v_x, v_y), point)
dist = 0.01 if dist == 0.0 else dist
weights[i] = (
1. / dist
)**2 # inverse weighted distance or inverse weighted square
weights /= np.sum(weights) # normalize them
idw = 0
for i in range(threshold):
w = weights[i]
try:
pre_rss = pre_inter[points[i][0]][points[i][1]]
except:
pre_rss = tx_pl # the additional Rx at the same location as the Tx
idw += w * pre_rss
inter[new_x][new_y] = idw
return inter.reshape(target_grid_len * target_grid_len)
