def _vector_sequence_similarity_dot(x, y, locality=0.5):
nx = len(x)
ny = len(y)
z = np.dot(x, y.T)
m2 = 0.
for i in prange(nx):
for j in prange(ny):
seye = 1. / (np.abs(i - j) + 1)
zij = z[i, j] * (locality * (seye - 1) + 1)
z[i, j] = zij
m2 += z[i, :].max()
m1 = 0.
for j in prange(ny):
m1 += z[:, j].max()
return 0.5 * (m1 + m2) / (nx + ny)
def loops_NumbaJit_parallel(csm, e, h, r0, rm, kj):
""" Workaround for the prange error in jit. See documentation comment of
'loops_NumbaJit_parallel_FirstWritingOfSteer'.
For infos on the numba decorator see 'vectorizedOptimized_NumbaJit_Parallel'
"""
nFreqs = csm.shape[0]
nGridPoints = r0.shape[0]
nMics = csm.shape[1]
beamformOutput = np.zeros((nFreqs, nGridPoints), np.float64)
for cntFreqs in xrange(nFreqs):
kjj = kj[cntFreqs].imag
for cntGrid in prange(nGridPoints):
r01 = r0[cntGrid]
rs = r01 ** 2
temp1 = 0.0
for cntMics in xrange(nMics):
temp2 = 0.0
rm1 = rm[cntGrid, cntMics]
temp3 = np.float32(kjj * (rm1 - r01))
steerVec = (np.cos(temp3) - 1j * np.sin(temp3)) * rm1
for cntMics2 in xrange(cntMics):
rm1 = rm[cntGrid, cntMics2]
temp3 = np.float32(kjj * (rm1 - r01))
steerVec1 = (np.cos(temp3) - 1j * np.sin(temp3)) * rm1 # Steering vec is calculated redundantly--> very slow
temp2 += csm[cntFreqs, cntMics2, cntMics] * steerVec1
temp1 += 2 * (temp2 * steerVec.conjugate()).real
temp1 += (csm[cntFreqs, cntMics, cntMics] * steerVec.conjugate() * steerVec).real
beamformOutput[cntFreqs, cntGrid] = temp1 / rs
return beamformOutput
def lu_parallel_2(x):
upper = np.asfortranarray(np.zeros(x.shape, dtype=np.float64))
n = len(x)
lower = np.ascontiguousarray(np.eye(n))
for i in range(n):
for k in prange(i, n):
total = lower[i, :] @ upper[:, k]
upper[i][k] = x[i][k] - total
for k in prange(i + 1, n):
total = lower[k, :] @ upper[:, i]
lower[k][i] = (x[k][i] - total) / upper[i][i]
return lower, upper
def _eval_metric(clickthroughs, results_lookup, results_lookup_idx, metric):
"""Apply metric to clickthroughs
Parameters
----------
clickthroughs : np.ndarray
return value of self._simplify_df
results_lookup : np.ndarray
return value of self._build_results_lookup
results_lookup_idx : np.ndarray
return value of self._build_results_lookup_indexes
metric : callable
Must be decorated with numba.njit. Will be called with two
arguments. First a 2-d ndarray with the first dimension
representing the character length of the prefix search and the
second containing the page_id of the top-k results in rank order.
The second argument is the clicked page_id. The metric must
return a single float.
Returns
-------
np.ndarray
result of applying metric to each row of clickthroughs
"""
out = np.empty(clickthroughs.shape[0], dtype=np.float32)
for i in numba.prange(clickthroughs.shape[0]):
cat_id, clickpage = clickthroughs[i]
# list of cat_ids for prefix searches on cat_id
# from shortest to longest
results_list_idx = results_lookup_idx[cat_id]
# the search results for all prefix searches of cat_id
# up to self.max_prefix_len
searchterm_results = results_lookup[results_list_idx]
out[i] = metric(searchterm_results, clickpage)
return out
def gini_coefficient(y):
r"""
Implements the Gini inequality index
Parameters
-----------
y : array_like(float)
Array of income/wealth for each individual. Ordered or unordered is fine
Returns
-------
Gini index: float
The gini index describing the inequality of the array of income/wealth
References
----------
https://en.wikipedia.org/wiki/Gini_coefficient
"""
n = len(y)
i_sum = np.zeros(n)
for i in prange(n):
for j in range(n):
i_sum[i] += abs(y[i] - y[j])
return np.sum(i_sum) / (2 * n * np.sum(y))
def loops_NumbaJit_parallelSlow(csm, r0, rm, kj):
nFreqs = csm.shape[0]
nGridPoints = len(r0)
nMics = csm.shape[1]
beamformOutput = np.zeros((nFreqs, nGridPoints), np.float64)
for cntFreqs in xrange(nFreqs):
kjj = kj[cntFreqs].imag
for cntGrid in prange(nGridPoints):
r01 = r0[cntGrid]
rs = r01 ** 2
temp1 = 0.0
for cntMics in xrange(nMics):
temp2 = 0.0
rm1 = rm[cntGrid, cntMics]
temp3 = np.float32(kjj * (rm1 - r01))
steerVec = (np.cos(temp3) - 1j * np.sin(temp3)) * rm1
for cntMics2 in xrange(cntMics):
rm1 = rm[cntGrid, cntMics2]
temp3 = np.float32(kjj * (rm1 - r01))
steerVec1 = (np.cos(temp3) - 1j * np.sin(temp3)) * rm1
temp2 += csm[cntFreqs, cntMics2, cntMics] * steerVec1
temp1 += 2 * (temp2 * steerVec.conjugate()).real
temp1 += (csm[cntFreqs, cntMics, cntMics] * steerVec.conjugate() * steerVec).real
beamformOutput[cntFreqs, cntGrid] = (temp1 / rs).real
return beamformOutput
def neighbor_average_waveform(waveforms, neighbors, lwt):
"""
Obtain the average waveform built from the neighbors of each pixel
Parameters
----------
waveforms : ndarray
Waveforms stored in a numpy array.
Shape: (n_chan, n_pix, n_samples)
neighbors : ndarray
2D array where each row is [pixel index, one neighbor of that pixel].
Changes per telescope.
Can be obtained from
`ctapipe.instrument.CameraGeometry.neighbor_matrix_where`.
lwt: int
Weight of the local pixel (0: peak from neighbors only,
1: local pixel counts as much as any neighbor)
Returns
-------
average_wf : ndarray
Average of neighbor waveforms for each pixel.
Shape: (n_chan, n_pix, n_samples)
"""
n_neighbors = neighbors.shape[0]
sum_ = waveforms * lwt
n = np.zeros(waveforms.shape)
for i in prange(n_neighbors):
pixel = neighbors[i, 0]
neighbor = neighbors[i, 1]
for channel in range(waveforms.shape[0]):
sum_[channel, pixel] += waveforms[channel, neighbor]
n[channel, pixel] += 1
return sum_ / n
def calc_fitness(population, exposure_indices, target, v_alpha, h_alpha, v_beta, h_beta):
fitness = np.zeros(population.shape[1],dtype=np.float32)
exposure = np.zeros(target.shape,dtype=np.float32)
field = np.zeros(target.shape, dtype=np.float32)
for p in range(population.shape[1]):
field *= 0
set_doses_field(field, exposure_indices, population[:, p])
exposure = calc_exposure(field, v_alpha, h_alpha, v_beta, h_beta)
#fitness[p] = np.sum(np.abs(np.subtract(target,exposure)))#/exposure_indices.shape[0]
buf = 0
for i in prange(target.shape[0]):
for j in range(target.shape[1]):
if target[i,j] > 0:
buf += np.abs(exposure[i,j]-target[i,j])**2
fitness[p] = fitness[p] + buf
# #fitness[p] += np.sum(population[:,p]>np.var(population[:,p]))#/exposure_indices.shape[0]
# buf = 0
# for i in range(field.shape[0]):
# for j in range(field.shape[1]):
# if field[i,j] > 0:
# neighbors = np.array([field[i+1,j],field[i-1,j],field[i,j+1],field[i,j-1],field[i+1,j+1],field[i+1,j-1],field[i-1,j+1],field[i-1,j-1]],dtype=np.float32)
# notzero = np.sum(neighbors>0)+1
# buf += np.abs(field[i,j]- (field[i,j] +field[i+1,j]+field[i,j+1]+field[i-1,j]+field[i,j-1]+field[i+1,j+1]+field[i+1,j-1]+field[i-1,j+1]+field[i-1,j-1])/notzero)**2
# fitness[p] = fitness[p] + buf
fitness /= exposure_indices.shape[0]
#fitness /= fitness.mean()
return fitness
def get_tem_potential_numba(time, R0, vnorm, vdir, center, rcut, inte1,
rr, dr, fr_val, conj_c2r, l, m, jmx, ind_lm, ind_lmm, V_time):
"""
Numba version of the computation of the external potential in time
for tem calculations
"""
for it in nb.prange(time.shape[0]):
R_sub = R0 + vnorm*vdir*(time[it] - time[0]) - center
norm = np.sqrt(np.dot(R_sub, R_sub))
if norm > rcut:
I1 = inte1/(norm**(l+1))
I2 = 0.0
else:
rsub_max = (np.abs(rr - norm)).argmin() # find_nearrest_index(rr, norm)
I1 = np.sum(fr_val[0:rsub_max+1]*
rr[0:rsub_max+1]**(l+2)*rr[0:rsub_max+1])
I2 = np.sum(fr_val[rsub_max+1:]*
rr[rsub_max+1:]/(rr[rsub_max+1:]**(l-1)))
I1 = I1*dr/(norm**(l+1))
I2 = I2*(norm**l)*dr
clm_tem = csphar_numba(R_sub, l)
clm = (4*np.pi/(2*l+1))*clm_tem[ind_lm]*(I1 + I2)
clmm = (4*np.pi/(2*l+1))*clm_tem[ind_lmm]*(I1 + I2)
rlm = c2r_lm(conj_c2r, jmx, clm, clmm, m)
V_time[it] = rlm + 0.0j
def kernel_run(array, alpha, result):
for i in prange(array.shape[0]):
sqr_norm = 0.
for d in range(array.shape[1]):
sqr_norm += array[i, d]**2 * alpha[d]
result[i] = math.sqrt(sqr_norm)
def __vector_sequence_similarity_euclid(x, y, z, nx, ny, locality=0.5):
nx = len(x)
ny = len(y)
m1 = 0.
m2 = 0.
for i in prange(nx):
xi = x[i]
for j in prange(ny):
yj = y[j]
zij = (1. - (((xi - yj) ** 2).sum() ** 0.5))
seye = 1. / (np.abs(i - j) + 1)
zij *= locality * (seye - 1) + 1
z[i, j] = zij
m2 += z[i, :].max()
for j in prange(ny):
m1 += z[:, j].max()
return (m1 + m2) / (nx + ny)
def prange_reduction2():
"""
>>> prange_reduction2()
49999995000000.0
"""
sum = 0.0
for i in numba.prange(10000000):
sum += i
return sum
def simple_prange_lastprivate():
"""
>>> simple_prange_lastprivate()
10
"""
var = 20
for i in numba.prange(1):
var = 10
return var
def calc_pi(NUM):
counter = 0
for i in prange(NUM):
x = random.random()
y = random.random()
if x*x+y*y < 1.0:
counter += 1
pi = 4.0*counter/NUM
return pi
def convolve_with_vector(field,exposure,v,h):
buf = np.zeros(field.shape,dtype=np.float32)
for j in prange(field.shape[1]):
for i in range(field.shape[0]):
if field[i,j] > 0:
for k in range(v.shape[0]):
fi = i+k-int((v.shape[0]-1)/2)
if fi >= 0 and fi < field.shape[0]:
buf[fi,j] += field[i,j]*v[k]
for i in prange(field.shape[0]):
for j in range(field.shape[1]):
if buf[i,j] < 0:
for k in range(h.shape[0]):
fj = j+k-int((h.shape[0])/2)
if fj >= 0 and fj < field.shape[1]:
exposure[i,fj] += buf[i,j]*h[k]
def simple_prange_reduction():
"""
>>> simple_prange_reduction()
15
"""
var = 10
for i in numba.prange(1):
var += 5
return var
def loops_Njit_Parallel_Prange(csm, SpecAllChn):
nFreqs = csm.shape[0]
nMics = csm.shape[1]
for cntFreq in range(nFreqs):
for cntRow in prange(nMics):
temp = np.conj(SpecAllChn[cntFreq, cntRow])
for cntColumn in range(nMics):
csm[cntFreq, cntRow, cntColumn] += temp * SpecAllChn[cntFreq, cntColumn]
return csm
def _downsample_2d_std_var(src, mask, use_mask, method, fill_value, mode_rank, out):
src_w, src_h, out_w, out_h = _get_dimensions(src, out)
if src_w == out_w and src_h == out_h:
return src
if out_w > src_w or out_h > src_h:
raise ValueError("invalid target size")
scale_x = src_w / out_w
scale_y = src_h / out_h
for out_y in prange(out_h):
src_yf0 = scale_y * out_y
src_yf1 = src_yf0 + scale_y
src_y0 = int(src_yf0)
src_y1 = int(src_yf1)
wy0 = 1.0 - (src_yf0 - src_y0)
wy1 = src_yf1 - src_y1
if wy1 < _EPS:
wy1 = 1.0
if src_y1 > src_y0:
src_y1 -= 1
for out_x in range(out_w):
src_xf0 = scale_x * out_x
src_xf1 = src_xf0 + scale_x
src_x0 = int(src_xf0)
src_x1 = int(src_xf1)
wx0 = 1.0 - (src_xf0 - src_x0)
wx1 = src_xf1 - src_x1
if wx1 < _EPS:
wx1 = 1.0
if src_x1 > src_x0:
src_x1 -= 1
v_sum = 0.0
w_sum = 0.0
wv_sum = 0.0
wvv_sum = 0.0
for src_y in range(src_y0, src_y1 + 1):
wy = wy0 if (src_y == src_y0) else wy1 if (src_y == src_y1) else 1.0
for src_x in range(src_x0, src_x1 + 1):
wx = wx0 if (src_x == src_x0) else wx1 if (src_x == src_x1) else 1.0
v = src[src_y, src_x]
if np.isfinite(v) and not (use_mask and mask[src_y, src_x]):
w = wx * wy
v_sum += v
w_sum += w
wv_sum += w * v
wvv_sum += w * v * v
if w_sum < _EPS:
out[out_y, out_x] = fill_value
else:
out[out_y, out_x] = (wvv_sum * w_sum - wv_sum * wv_sum) / w_sum / w_sum
if method == DS_STD:
out = np.sqrt(out)
return out
def lnlike_normal_v(o, m, e, lcids):
m = atleast_2d(m)
npv = m.shape[0]
npt = o.size
lnl = zeros(npv)
for i in prange(npv):
for j in range(npt):
k = lcids[j]
lnl[i] += -log(e[i,k]) - 0.5*log(2*pi) - 0.5*((o[j]-m[i,j])/e[i,k])**2
return lnl
def prange_reduction_and_privates():
"""
>>> prange_reduction_and_privates()
100.0
"""
sum = 10.0
for i in numba.prange(10):
j = i * 2
sum += j
return sum
def calculate_beta_sobolev(tau_sobolevs, beta_sobolevs):
for i in prange(len(tau_sobolevs)):
if tau_sobolevs[i] > 1e3:
beta_sobolevs[i] = tau_sobolevs[i]**-1
elif tau_sobolevs[i] < 1e-4:
beta_sobolevs[i] = 1 - 0.5 * tau_sobolevs[i]
else:
beta_sobolevs[i] = (1 - np.exp(-tau_sobolevs[i])) / (
tau_sobolevs[i])
return beta_sobolevs
def ea_iter_v(t, t0, p, e, w):
Ma = mean_anomaly(t, t0, p, e, w)
ec = e*sin(Ma)/(1.0 - e*cos(Ma))
for j in prange(len(t)):
for k in range(15):
ect = ec[j]
ec[j] = e*sin(Ma[j]+ec[j])
if (abs(ect-ec[j]) < 1e-4):
break
Ea = Ma + ec
return Ea
def check_limits(population):
"""
Check if all individuals of the are bigger than 0, meaning that no negativ exposure is allowed
:param population: Population matrix
:return: corrected population
"""
for i in prange(population.shape[1]):
for j in range(population.shape[0]):
if population[j, i] < 0:
population[j, i] = 0
return population
def _fj(j, l, m, a, b):
"""From Handbook of Computational Quantum Chemistry by David B. Cook
in chapter 7.7.1 -- Essentially a FOILing of the pre-exponential
cartesian power dependence in one dimension."""
tot, i, f = 0., max(0, j - m), min(j, l) + 1
for k in prange(i, f):
tot += (choose(l, k) *
choose(m, int(j - k)) *
a ** (l - k) *
b ** (m + k - j))
return tot
def _nin(l, m, pa, pb, p, N):
"""From Handbook of Computational Quantum Chemistry by David B. Cook
in chapter 7.7.1 -- Sums the result of _fj over the total angular momentum
in one dimension."""
ltot = l + m
if not ltot: return N
tot = 0.
for j in prange(int(ltot // 2 + 1)):
tot += (_fj(2 * j, l, m, pa, pb) *
dfac21(j) / (2 * p) ** j)
return tot * N
def simple_prange_shared():
"""
>>> simple_prange_shared()
20L
"""
result = np.empty(1, dtype=np.int64)
shared = 20
for i in numba.prange(1):
result[0] = shared
return result[0]
def simple_prange_private():
"""
>>> simple_prange_private()
10L
"""
result = np.empty(1, dtype=np.int64)
var = 20
for i in numba.prange(1):
var = 10
result[0] = var
return result[0]
def kde(X):
b = 0.5
points = np.array([-1.0, 2.0, 5.0])
N = points.shape[0]
n = X.shape[0]
exps = 0
for i in prange(n):
p = X[i]
d = (-(p-points)**2)/(2*b**2)
m = np.min(d)
exps += m-np.log(b*N)+np.log(np.sum(np.exp(d-m)))
return exps
def mutate(arr,sigma,mutation_rate):
for i in prange(arr.shape[0]):
if np.random.random() < mutation_rate:
#mutation = np.random.normal()*sigma
#if mutation > sigma*1.0:
# mutation = sigma
#if mutation < -sigma*1.0:
# mutation = -sigma
mutation = (np.random.random()-0.5) * sigma
# mutation = arr[i]*2*(np.random.random()-0.5) * 0.05
arr[i] = arr[i] + mutation
return arr
def prange_shared_privates_reductions(shared):
"""
>>> prange_shared_privates_reductions(2.0)
100.0
"""
sum = 10.0
for i in numba.prange(10):
j = i * shared
sum += j
shared = 3.0
return sum
def make1Drec(wt, lenwt1d, band, scale2d, nx, ny, nz, mode='haar'):
# 1d wavelet decomposition
if mode == 'haar':
print "Using Haar recomp"
F = np.array([0.5, 0.5])
p = 1
Lo_R = np.sqrt(2) * F / np.sum(F)
Hi_R = Lo_R[::-1].copy()
first = 2 - p % 2
Hi_R[first::2] = -Hi_R[first::2]
if mode == '9/7':
print "Using 9/7 recomp"
Df = np.array([0.0267487574110000,-0.0168641184430000,-0.0782232665290000,0.266864118443000,\
0.602949018236000,0.266864118443000,-0.0782232665290000,-0.0168641184430000,\
0.0267487574110000])
Rf = np.array([-0.0456358815570000,-0.0287717631140000,0.295635881557000,0.557543526229000,\
0.295635881557000,-0.0287717631140000,-0.0456358815570000])
lr = len(Rf)
ld = len(Df)
lmax = max(lr, ld)
if lmax % 2:
lmax += 1
Rf = np.hstack(
[np.zeros((lmax - lr) / 2), Rf,
np.zeros((lmax - lr + 1) / 2)])
Df = np.hstack(
[np.zeros((lmax - ld) / 2), Df,
np.zeros((lmax - ld + 1) / 2)])
p = 1
first = 2 - p % 2
Lo_R1 = np.sqrt(2) * Df / np.sum(Df)
Hi_R1 = Lo_R1[::-1].copy()
Hi_R1[first::2] = -Hi_R1[first::2]
#Hi_D1=Hi_R1[::-1].copy()
#Lo_D1=Lo_R1[::-1].copy()
Lo_R2 = np.sqrt(2) * Rf / np.sum(Rf)
#Hi_R2 = Lo_R2[::-1].copy()
#Hi_R2[first::2] = -Hi_R2[first::2]
#Hi_D2=Hi_R2[::-1].copy()
#Lo_D2=Lo_R2[::-1].copy()
Lo_R = Lo_R2
Hi_R = Hi_R1
(h1, g1) = (Lo_R, Hi_R) #(h0,g0)=(Lo_D1,Hi_D2)
tmp = np.zeros((nz * scale2d, nx, ny))
for bd in prange(scale2d):
for corx in prange(nx):
for cory in prange(ny):
tmpwt = wt[lenwt1d * bd:lenwt1d * (bd + 1), corx, cory].copy()
sig = tmpwt[:np.int(band[0])].copy() #JGMOD
start = np.int(band[0]) #JGMOD
for sc in np.arange(np.size(band) - 2):
last = start + np.int(band[sc + 1]) #JGMOD
detail = tmpwt[start:last].copy()
lsig = 2 * sig.size
s = band[sc + 2]
appInt = np.zeros(lsig - 1)
appInt[::2] = sig.copy()
appInt = fast_convolution(appInt, h1, 1)
first = np.int(np.floor(float(np.size(appInt) - s) /
2.)) #JGMOD
last = np.int(
np.size(appInt) -
np.ceil(float(np.size(appInt) - s) / 2.)) #JGMOD
appInt = appInt[first:last]
detailInt = np.zeros(lsig - 1)
detailInt[::2] = detail.copy()
detailInt = fast_convolution(detailInt, g1, 1)
detailInt = detailInt[first:last]
sig = appInt + detailInt
start = last
tmp[bd * nz:(bd + 1) * nz, corx, cory] = sig.copy()
return tmp
def _gadf(X_cos, X_sin, n_samples, image_size):
X_gadf = np.empty((n_samples, image_size, image_size))
for i in prange(n_samples):
X_gadf[i] = np.outer(X_sin[i], X_cos[i]) - np.outer(X_cos[i], X_sin[i])
return X_gadf
def _process_proximity_line(source_line, x_coords, y_coords, pan_near_x,
pan_near_y, is_forward, line_id, width,
max_distance, line_proximity, values,
distance_metric):
# Process proximity for a line of pixels in an image
#
# source_line: 1d ndarray, input data
# pan_near_x: 1d ndarray
# pan_near_y: 1d ndarray
# is_forward: boolean, will we loop forward through pixel?
# line_id: np.int64, index of the source_line in the image
# width: np.int64, image width. It is the number of pixels in the
# source_line
# max_distance: np.float64, maximum distance considered.
# line_proximity: 1d numpy array of type np.float64,
# calculated proximity from source_line
# values: 1d numpy array of type np.uint8,
# A list of target pixel values to measure the distance from.
# If this option is not provided proximity will be computed
# from non-zero pixel values.
# Currently pixel values are internally processed as integers
# Return: 1d numpy array of type np.float64.
# Corresponding proximity of source_line.
start = width - 1
end = -1
step = -1
if is_forward:
start = 0
end = width
step = 1
n_values = len(values)
for pixel in prange(start, end, step):
is_target = False
# Is the current pixel a target pixel?
if n_values == 0:
is_target = source_line[pixel] != 0
else:
for i in prange(n_values):
if source_line[pixel] == values[i]:
is_target = True
if is_target:
line_proximity[pixel] = 0.0
pan_near_x[pixel] = pixel
pan_near_y[pixel] = line_id
continue
# Are we near(er) to the closest target to the above (below) pixel?
near_distance_square = max_distance**2 * 2.0
if pan_near_x[pixel] != -1:
# distance_square
dist_sqr = _distance(x_coords[pan_near_x[pixel]], x_coords[pixel],
y_coords[pan_near_y[pixel]],
y_coords[line_id], distance_metric)
if dist_sqr < near_distance_square:
near_distance_square = dist_sqr
else:
pan_near_x[pixel] = -1
pan_near_y[pixel] = -1
# Are we near(er) to the closest target to the left (right) pixel?
last = pixel - step
if pixel != start and pan_near_x[last] != -1:
dist_sqr = _distance(x_coords[pan_near_x[last]], x_coords[pixel],
y_coords[pan_near_y[last]], y_coords[line_id],
distance_metric)
if dist_sqr < near_distance_square:
near_distance_square = dist_sqr
pan_near_x[pixel] = pan_near_x[last]
pan_near_y[pixel] = pan_near_y[last]
# Are we near(er) to the closest target to the
# topright (bottom left) pixel?
tr = pixel + step
if tr != end and pan_near_x[tr] != -1:
dist_sqr = _distance(x_coords[pan_near_x[tr]], x_coords[pixel],
y_coords[pan_near_y[tr]], y_coords[line_id],
distance_metric)
if dist_sqr < near_distance_square:
near_distance_square = dist_sqr
pan_near_x[pixel] = pan_near_x[tr]
pan_near_y[pixel] = pan_near_y[tr]
# Update our proximity value.
if pan_near_x[pixel] != -1 and not np.isnan(source_line[pixel])\
and max_distance * max_distance >= near_distance_square\
and (line_proximity[pixel] < 0 or
near_distance_square <
line_proximity[pixel] * line_proximity[pixel]):
line_proximity[pixel] = sqrt(near_distance_square)
return
def rbf(X, w):
result = np.zeros((X.shape[0], ), dtype=np.float32)
for idx in prange(X.shape[0]):
result[idx] = -np.linalg.norm(X[idx, :] - w)**2 / (2 * gamma**2)
return np.exp(result)
def construct_lower(L, U, nrow, ncol):
for i in prange(ncol):
for j in range(i, nrow):
L[j, i] /= U[i, i]
L[i, i] = 1
def tensor_conv1d(
out,
out_shape,
out_strides,
out_size,
input,
input_shape,
input_strides,
weight,
weight_shape,
weight_strides,
reverse,
):
"""
1D Convolution implementation.
Given input tensor of
`batch, in_channels, width`
and weight tensor
`out_channels, in_channels, k_width`
Computes padded output of
`batch, out_channels, width`
`reverse` decides if weight is anchored left (False) or right.
(See diagrams)
Args:
out (array): storage for `out` tensor.
out_shape (array): shape for `out` tensor.
out_strides (array): strides for `out` tensor.
out_size (int): size of the `out` tensor.
input (array): storage for `input` tensor.
input_shape (array): shape for `input` tensor.
input_strides (array): strides for `input` tensor.
weight (array): storage for `input` tensor.
weight_shape (array): shape for `input` tensor.
weight_strides (array): strides for `input` tensor.
reverse (bool): anchor weight at left or right
"""
batch_, out_channels, out_width = out_shape
batch, in_channels, width = input_shape
out_channels_, in_channels_, kw = weight_shape
assert (
batch == batch_
and in_channels == in_channels_
and out_channels == out_channels_
)
for p in prange(out_size):
out_index = np.zeros(3, np.int32)
pos1 = np.zeros(3, np.int32)
pos2 = np.zeros(3, np.int32)
count(p, out_shape, out_index)
cur_batch, cur_out_channels, cur_n = out_index
# input matrix - out_tensor[cur_batch, cur_out_channels: cur_out_channels + in_channels, cur_n: cur_n + kw]
# shape of input matrix: (1, in_channels, kw)
# weight matrix - (1, in_channels, kw)
v = 0
for i in range(in_channels):
for j in range(kw):
if not reverse:
pos1[0] = cur_out_channels
pos1[1] = i
pos1[2] = j
pos2[0] = cur_batch
pos2[1] = i
pos2[2] = cur_n + j
if cur_n + j >= width:
v += 0
else:
v += weight[index_to_position(pos1, weight_strides)] * input[index_to_position(
pos2, input_strides)]
else:
if cur_n - j < 0:
v += 0
else:
pos1[0] = cur_out_channels
pos1[1] = i
pos1[2] = j
pos2[0] = cur_batch
pos2[1] = i
pos2[2] = cur_n - j
v += weight[index_to_position(pos1, weight_strides)] * input[index_to_position(
pos2, input_strides)]
out[index_to_position(out_index, out_strides)] = v
def _astype_numba(arr, result):
for i in nb.prange(len(arr)):
# conversion occurs implicitly, and numba only supports conversion
# between arrays of numeric types.
result[i] = arr[i]
def find_node_split(context, sample_indices):
"""For each feature, find the best bin to split on at a given node.
Returns the best split info among all features, and the histograms of
all the features. The histograms are computed by scanning the whole
data.
Parameters
----------
context : SplittingContext
The splitting context
sample_indices : array of int
The indices of the samples at the node to split.
Returns
-------
best_split_info : SplitInfo
The info about the best possible split among all features.
histograms : array of HISTOGRAM_DTYPE, shape=(n_features, max_bins)
The histograms of each feature. A histogram is an array of
HISTOGRAM_DTYPE of size ``max_bins`` (only
``n_bins_per_features[feature]`` entries are relevant).
"""
ctx = context # shorter name to avoid various line breaks
n_samples = sample_indices.shape[0]
# Need to declare local variables, else they're not updated
# (see numba issue 3459)
ordered_gradients = ctx.ordered_gradients
ordered_hessians = ctx.ordered_hessians
# Populate ordered_gradients and ordered_hessians. (Already done for root)
# Ordering the gradients and hessians helps to improve cache hit.
# This is a parallelized version of the following vanilla code:
# for i range(n_samples):
# ctx.ordered_gradients[i] = ctx.gradients[samples_indices[i]]
if sample_indices.shape[0] != ctx.gradients.shape[0]:
starts, ends, n_threads = get_threads_chunks(n_samples)
if ctx.constant_hessian:
for thread_idx in prange(n_threads):
for i in range(starts[thread_idx], ends[thread_idx]):
ordered_gradients[i] = ctx.gradients[sample_indices[i]]
else:
for thread_idx in prange(n_threads):
for i in range(starts[thread_idx], ends[thread_idx]):
ordered_gradients[i] = ctx.gradients[sample_indices[i]]
ordered_hessians[i] = ctx.hessians[sample_indices[i]]
ctx.sum_gradients = ctx.ordered_gradients[:n_samples].sum()
if ctx.constant_hessian:
ctx.sum_hessians = ctx.constant_hessian_value * float32(n_samples)
else:
ctx.sum_hessians = ctx.ordered_hessians[:n_samples].sum()
# Pre-allocate the results datastructure to be able to use prange:
# numba jitclass do not seem to properly support default values for kwargs.
split_infos = [
SplitInfo(-1., 0, 0, 0., 0., 0., 0., 0, 0)
for i in range(context.n_features)
]
histograms = np.empty(shape=(np.int64(context.n_features),
np.int64(context.max_bins)),
dtype=HISTOGRAM_DTYPE)
for feature_idx in prange(context.n_features):
split_info, histogram = _find_histogram_split(context, feature_idx,
sample_indices)
split_infos[feature_idx] = split_info
histograms[feature_idx, :] = histogram
split_info = _find_best_feature_to_split_helper(split_infos)
return split_info, histograms
def find_node_split_subtraction(context, sample_indices, parent_histograms,
sibling_histograms):
"""For each feature, find the best bin to split on at a given node.
Returns the best split info among all features, and the histograms of
all the features.
This does the same job as ``find_node_split()`` but uses the histograms
of the parent and sibling of the node to split. This allows to use the
identity: ``histogram(parent) = histogram(node) - histogram(sibling)``,
which is significantly faster than computing the histograms from data.
Returns the best SplitInfo among all features, along with all the feature
histograms that can be latter used to compute the sibling or children
histograms by substraction.
Parameters
----------
context : SplittingContext
The splitting context
sample_indices : array of int
The indices of the samples at the node to split.
parent_histograms : array of HISTOGRAM_DTYPE of shape(n_features, max_bins)
The histograms of the parent
sibling_histograms : array of HISTOGRAM_DTYPE of \
shape(n_features, max_bins)
The histograms of the sibling
Returns
-------
best_split_info : SplitInfo
The info about the best possible split among all features.
histograms : array of HISTOGRAM_DTYPE, shape=(n_features, max_bins)
The histograms of each feature. A histogram is an array of
HISTOGRAM_DTYPE of size ``max_bins`` (only
``n_bins_per_features[feature]`` entries are relevant).
"""
# We can pick any feature (here the first) in the histograms to
# compute the gradients: they must be the same across all features
# anyway, we have tests ensuring this. Maybe a more robust way would
# be to compute an average but it's probably not worth it.
context.sum_gradients = (parent_histograms[0]['sum_gradients'].sum() -
sibling_histograms[0]['sum_gradients'].sum())
n_samples = sample_indices.shape[0]
if context.constant_hessian:
context.sum_hessians = \
context.constant_hessian_value * float32(n_samples)
else:
context.sum_hessians = (parent_histograms[0]['sum_hessians'].sum() -
sibling_histograms[0]['sum_hessians'].sum())
# Pre-allocate the results datastructure to be able to use prange
split_infos = [
SplitInfo(-1., 0, 0, 0., 0., 0., 0., 0, 0)
for i in range(context.n_features)
]
histograms = np.empty(shape=(np.int64(context.n_features),
np.int64(context.max_bins)),
dtype=HISTOGRAM_DTYPE)
for feature_idx in prange(context.n_features):
split_info, histogram = _find_histogram_split_subtraction(
context, feature_idx, parent_histograms, sibling_histograms,
n_samples)
split_infos[feature_idx] = split_info
histograms[feature_idx, :] = histogram
split_info = _find_best_feature_to_split_helper(split_infos)
return split_info, histograms
def test_impl(t_obj, X):
for i in prange(t_obj.T):
X[i] = i
return X.sum()
def sdc_take_array_impl(data, indices):
res_size = len(indices)
res_arr = numpy.empty(res_size, dtype=data_dtype)
for i in numba.prange(res_size):
res_arr[i] = data[indices[i]]
return res_arr
def etimated_cov_mat(vert, eta, eigenvalue_nk, alphalapl, u_nkl_norm):
expressionAcc = np.zeros((vert, vert))
for m in prange(0, vert):
for n in range(0, vert):
expressionAcc[m][n] = expressionAcc[m][n] + (eta ** 2) * (eigenvalue_nk ** (-alphalapl)) * u_nkl_norm[m] * u_nkl_norm[n]
return expressionAcc
def test_impl(data):
N = data.shape[0]
sums = np.zeros(N)
for i in prange(N):
sums[i] = np.sum(data[np.int32(0):np.int32(1)])
return sums
def make1Dtf(tmp, scale1d, scale2d, nx, ny, nz, size, lenwt1d, mode='haar'):
# 1d wavelet decomposition
scale = scale1d
if mode == 'haar':
print "Using Haar decomp"
F = np.array([0.5, 0.5])
p = 1
Lo_R = np.sqrt(2) * F / np.sum(F)
Hi_R = Lo_R[::-1].copy()
first = 2 - p % 2
Hi_R[first::2] = -Hi_R[first::2]
Hi_D = Hi_R[::-1].copy()
Lo_D = Lo_R[::-1].copy()
if mode == '9/7':
print "Using 9/7 decomp"
Df = np.array([0.0267487574110000,-0.0168641184430000,-0.0782232665290000,0.266864118443000,\
0.602949018236000,0.266864118443000,-0.0782232665290000,-0.0168641184430000,\
0.0267487574110000])
Rf = np.array([-0.0456358815570000,-0.0287717631140000,0.295635881557000,0.557543526229000,\
0.295635881557000,-0.0287717631140000,-0.0456358815570000])
lr = len(Rf)
ld = len(Df)
lmax = max(lr, ld)
if lmax % 2:
lmax += 1
Rf = np.hstack(
[np.zeros((lmax - lr) / 2), Rf,
np.zeros((lmax - lr + 1) / 2)])
Df = np.hstack(
[np.zeros((lmax - ld) / 2), Df,
np.zeros((lmax - ld + 1) / 2)])
p = 1
first = 2 - p % 2
Lo_R1 = np.sqrt(2) * Df / np.sum(Df)
Hi_R1 = Lo_R1[::-1].copy()
Hi_R1[first::2] = -Hi_R1[first::2]
Hi_D1 = Hi_R1[::-1].copy()
Lo_D1 = Lo_R1[::-1].copy()
Lo_R2 = np.sqrt(2) * Rf / np.sum(Rf)
Hi_R2 = Lo_R2[::-1].copy()
Hi_R2[first::2] = -Hi_R2[first::2]
Hi_D2 = Hi_R2[::-1].copy()
Lo_D2 = Lo_R2[::-1].copy()
Lo_D = Lo_D1
Hi_D = Hi_D2
(h0, g0) = (Lo_D, Hi_D)
lf = h0.size
band = np.zeros(scale + 1)
band[-1] = size
end = size
start = 1
wt2d1d = np.zeros((scale2d * lenwt1d, nx, ny), dtype=np.float64)
#print wt2d1d.shape
wt = np.array([], dtype=np.float64)
for bd in prange(scale2d):
for corx in prange(nx):
for cory in prange(ny):
x = tmp[bd * nz:(bd + 1) * nz, corx, cory].copy()
#pdb.set_trace()
wt = np.array([], dtype=np.float64)
#print "HELLO"
for sc in np.arange(scale - 1):
#print "Hello in scale",sc
lsig = x.size
end = lsig + lf - 1
lenExt = lf - 1
#xExt=np.array([1,2,3],dtype=np.float64)
# xExt = np.lib.pad(x, (lenExt,lenExt), 'symmetric')
#xExt=np.concatenate((x[lenExt-1::-1],x,x[-lenExt+1:-lenExt-1:-1]),axis=0)
xExt = np.concatenate(
(x[0:lenExt][::-1], x, x[::-1][0:lenExt]), axis=0)
#print x
#print xExt
#pdb.set_trace()
app = fast_convolution(xExt, h0,
0) #np.convolve(xExt,h0,'valid')
x = app[start:end:2].copy()
detail = fast_convolution(xExt, g0,
0) #np.convolve(xExt,g0,'valid')
#print app
#print detail
#pdb.set_trace()
#detail=np.array([1,2,3,4,5])
wt = np.hstack((detail[start:end:2], wt))
band[-2 - sc] = len(detail[start:end:2])
#print wt
#print wt
#pdb.set_trace()
wt = np.hstack((x, wt))
band[0] = len(x)
#print "wt=",wt.shape
#pdb.set_trace()
index1 = lenwt1d * bd
index2 = lenwt1d * (bd + 1)
a = wt.copy()
wt2d1d[index1:index2, corx, cory] = a
#print index1,index2,corx,cory,wt2d1d[index1:index2,corx,cory]
#print "Before return in make1Dtf"
return (wt2d1d, band)
def base_calc_semblance(
calc_nmo_func,
seismogram,
times,
offsets,
velocity,
sample_rate, # pylint: disable=too-many-arguments
win_size,
t_min,
t_max):
""" Calculate semblance for specified velocity in the preset time window from `t_min` to `t_max`.
Parameters
----------
calc_nmo_func : njitted callable
Callable that calculates normal moveout corrected seismogram for specified time and velocity values
and range of offsets.
seismogram : np.ndarray
Data for calculating semblance.
times : array-like
An array containing the recording time for each trace value.
offsets : array-like
The distance from the source to the receiver for each trace.
velocity : array-like
Velocity value for semblance computation.
sample_rate : int
Step in milliseconds between signal amplitude measurements during shooting.
_win_size : int
Window size for smoothing the semblance.
Measured in samples.
t_min : int
Time value to start compute semblance from.
Measured in samples.
t_max : int
The last time value for semblance computation.
Measured in samples.
Returns
-------
slice_semblance : 1d array
Semblance values for a specified `veloicty` in time range from `t_min` to `t_max`.
"""
t_win_size_min = max(0, t_min - win_size)
t_win_size_max = min(len(times) - 1, t_max + win_size)
nmo = np.empty(
(t_win_size_max - t_win_size_min + 1, seismogram.shape[1]))
for i in prange(t_win_size_min, t_win_size_max):
nmo[i - t_win_size_min] = calc_nmo_func(seismogram, times[i],
offsets, velocity,
sample_rate)
numerator = np.sum(nmo, axis=1)**2
denominator = np.sum(nmo**2, axis=1)
slice_semblance = np.zeros(t_max - t_min)
for t in prange(t_min, t_max):
t_rel = t - t_win_size_min
ix_from = max(0, t_rel - win_size)
ix_to = min(len(nmo) - 1, t_rel + win_size)
slice_semblance[t - t_min] = (
np.sum(numerator[ix_from:ix_to]) /
(len(offsets) * np.sum(denominator[ix_from:ix_to]) + 1e-6))
return slice_semblance
def tensor_conv2d(
out,
out_shape,
out_strides,
out_size,
input,
input_shape,
input_strides,
weight,
weight_shape,
weight_strides,
reverse,
):
"""
2D Convolution implementation.
Given input tensor of
`batch, in_channels, height, width`
and weight tensor
`out_channels, in_channels, k_height, k_width`
Computes padded output of
`batch, out_channels, height, width`
`Reverse` decides if weight is anchored top-left (False) or bottom-right.
(See diagrams)
Args:
out (array): storage for `out` tensor.
out_shape (array): shape for `out` tensor.
out_strides (array): strides for `out` tensor.
out_size (int): size of the `out` tensor.
input (array): storage for `input` tensor.
input_shape (array): shape for `input` tensor.
input_strides (array): strides for `input` tensor.
weight (array): storage for `input` tensor.
weight_shape (array): shape for `input` tensor.
weight_strides (array): strides for `input` tensor.
reverse (bool): anchor weight at top-left or bottom-right
"""
batch_, out_channels, _, _ = out_shape
batch, in_channels, height, width = input_shape
out_channels_, in_channels_, kh, kw = weight_shape
assert (
batch == batch_
and in_channels == in_channels_
and out_channels == out_channels_
)
for p in prange(out_size):
out_index = np.zeros(4, np.int32)
pos1 = np.zeros(4, np.int32)
pos2 = np.zeros(4, np.int32)
count(p, out_shape, out_index)
cur_batch, cur_out_channels, cur_h, cur_w = out_index
v = 0
for i in range(in_channels):
for j in range(kh):
for k in range(kw):
if not reverse:
pos1[0] = cur_out_channels
pos1[1] = i
pos1[2] = j
pos1[3] = k
pos2[0] = cur_batch
pos2[1] = i
pos2[2] = cur_h + j
pos2[3] = cur_w + k
if cur_h + j >= height or cur_w + k >= width:
v += 0
else:
v += weight[index_to_position(pos1, weight_strides)] * input[index_to_position(
pos2, input_strides)]
else:
pos1[0] = cur_out_channels
pos1[1] = i
pos1[2] = j
pos1[3] = k
pos2[0] = cur_batch
pos2[1] = i
pos2[2] = cur_h - j
pos2[3] = cur_w - k
if cur_h - j < 0 or cur_w - k < 0:
v += 0
else:
v += weight[index_to_position(pos1, weight_strides)] * input[index_to_position(
pos2, input_strides)]
out[index_to_position(out_index, out_strides)] = v
def split_indices(context, split_info, sample_indices):
"""Split samples into left and right arrays.
Parameters
----------
context : SplittingContext
The splitting context
split_ingo : SplitInfo
The SplitInfo of the node to split
sample_indices : array of int
The indices of the samples at the node to split. This is a view on
context.partition, and it is modified inplace by placing the indices
of the left child at the beginning, and the indices of the right child
at the end.
Returns
-------
left_indices : array of int
The indices of the samples in the left child. This is a view on
context.partition.
right_indices : array of int
The indices of the samples in the right child. This is a view on
context.partition.
"""
# This is a multi-threaded implementation inspired by lightgbm.
# Here is a quick break down. Let's suppose we want to split a node with
# 24 samples named from a to x. context.partition looks like this (the *
# are indices in other leaves that we don't care about):
# partition = [*************abcdefghijklmnopqrstuvwx****************]
# ^ ^
# node_position node_position + node.n_samples
# Ultimately, we want to reorder the samples inside the boundaries of the
# leaf (which becomes a node) to now represent the samples in its left and
# right child. For example:
# partition = [*************abefilmnopqrtuxcdghjksvw*****************]
# ^ ^
# left_child_pos right_child_pos
# Note that left_child_pos always takes the value of node_position, and
# right_child_pos = left_child_pos + left_child.n_samples. The order of
# the samples inside a leaf is irrelevant.
# 1. samples_indices is a view on this region a..x. We conceptually
# divide it into n_threads regions. Each thread will be responsible for
# its own region. Here is an example with 4 threads:
# samples_indices = [abcdef|ghijkl|mnopqr|stuvwx]
# 2. Each thread processes 6 = 24 // 4 entries and maps them into
# left_indices_buffer or right_indices_buffer. For example, we could
# have the following mapping ('.' denotes an undefined entry):
# - left_indices_buffer = [abef..|il....|mnopqr|tux...]
# - right_indices_buffer = [cd....|ghjk..|......|svw...]
# 3. We keep track of the start positions of the regions (the '|') in
# ``offset_in_buffers`` as well as the size of each region. We also keep
# track of the number of samples put into the left/right child by each
# thread. Concretely:
# - left_counts = [4, 2, 6, 3]
# - right_counts = [2, 4, 0, 3]
# 4. Finally, we put left/right_indices_buffer back into the
# samples_indices, without any undefined entries and the partition looks
# as expected
# partition = [*************abefilmnopqrtuxcdghjksvw*****************]
# Note: We here show left/right_indices_buffer as being the same size as
# sample_indices for simplicity, but in reality they are of the same size
# as partition.
X_binned = context.X_binned.T[split_info.feature_idx]
n_threads = numba.config.NUMBA_DEFAULT_NUM_THREADS
n_samples = sample_indices.shape[0]
# Note: we could probably allocate all the arrays of size n_threads in the
# splitting context as well, but gains are probably going to be minimal
sizes = np.full(n_threads, n_samples // n_threads, dtype=np.int32)
if n_samples % n_threads > 0:
# array[:0] will cause a bug in numba 0.41 so we need the if. Remove
# once issue numba 3554 is fixed.
sizes[:n_samples % n_threads] += 1
offset_in_buffers = np.zeros(n_threads, dtype=np.int32)
offset_in_buffers[1:] = np.cumsum(sizes[:-1])
left_counts = np.empty(n_threads, dtype=np.int32)
right_counts = np.empty(n_threads, dtype=np.int32)
# Need to declare local variables, else they're not updated :/
# (see numba issue 3459)
left_indices_buffer = context.left_indices_buffer
right_indices_buffer = context.right_indices_buffer
# map indices from samples_indices to left/right_indices_buffer
for thread_idx in prange(n_threads):
left_count = 0
right_count = 0
start = offset_in_buffers[thread_idx]
stop = start + sizes[thread_idx]
for i in range(start, stop):
sample_idx = sample_indices[i]
if X_binned[sample_idx] <= split_info.bin_idx:
left_indices_buffer[start + left_count] = sample_idx
left_count += 1
else:
right_indices_buffer[start + right_count] = sample_idx
right_count += 1
left_counts[thread_idx] = left_count
right_counts[thread_idx] = right_count
# position of right child = just after the left child
right_child_position = left_counts.sum()
# offset of each thread in samples_indices for left and right child, i.e.
# where each thread will start to write.
left_offset = np.zeros(n_threads, dtype=np.int32)
left_offset[1:] = np.cumsum(left_counts[:-1])
right_offset = np.full(n_threads, right_child_position, dtype=np.int32)
right_offset[1:] += np.cumsum(right_counts[:-1])
# map indices in left/right_indices_buffer back into samples_indices. This
# also updates context.partition since samples_indice is a view.
for thread_idx in prange(n_threads):
for i in range(left_counts[thread_idx]):
sample_indices[left_offset[thread_idx] + i] = \
left_indices_buffer[offset_in_buffers[thread_idx] + i]
for i in range(right_counts[thread_idx]):
sample_indices[right_offset[thread_idx] + i] = \
right_indices_buffer[offset_in_buffers[thread_idx] + i]
return (sample_indices[:right_child_position],
sample_indices[right_child_position:])
def fastfield(coordinates, r_p, k, phase, ab, result, bohren, cartesian):
'''
Returns the field scattered by the particle at each coordinate
Arguments
----------
coordinates : numpy.ndarray of dtype numpy.complex128
[3, npts] coordinate system for scattered field calculation
r_p : numpy.ndarray
[3] position of scatterer
k : float
Wavenumber of the light in medium of refractive index n_m
phase : np.complex128
Complex exponential phase to attach to Lorenz-Mie scattering
function. See equation XXX
ab : numpy.ndarray of dtype numpy.complex128
[2, norders] Mie scattering coefficients
result : numpy.ndarray of dtype numpy.complex128
[3, npts] buffer for final scattered field
cartesian : bool
If set, return field projected onto Cartesian coordinates.
Otherwise, return polar projection.
bohren : bool
If set, use sign convention from Bohren and Huffman.
Otherwise, use opposite sign convention.
'''
length = coordinates.shape[1]
norders = ab.shape[0] # number of partial waves in sum
# GEOMETRY
# 1. particle displacement [pixel]
# Note: The sign convention used here is appropriate
# for illumination propagating in the -z direction.
# This means that a particle forming an image in the
# focal plane (z = 0) is located at positive z.
# Accounting for this by flipping the axial coordinate
# is equivalent to using a mirrored (left-handed)
# coordinate system.
for idx in prange(length):
kx = k * (coordinates[0, idx] - r_p[0])
ky = k * (coordinates[1, idx] - r_p[1])
kz = k * (coordinates[2, idx] - r_p[2])
# 2. geometric factors
kz *= -1. # z convention
krho = math.sqrt(kx**2 + ky**2)
kr = math.sqrt(krho**2 + kz**2)
theta = math.atan2(krho, kz)
phi = math.atan2(ky, kx)
sintheta = math.sin(theta)
costheta = math.cos(theta)
sinphi = math.sin(phi)
cosphi = math.cos(phi)
sinkr = math.sin(kr)
coskr = math.cos(kr)
# SPECIAL FUNCTIONS
# starting points for recursive function evaluation ...
# 1. Riccati-Bessel radial functions, page 478.
# Particles above the focal plane create diverging waves
# described by Eq. (4.13) for $h_n^{(1)}(kr)$. These have z > 0.
# Those below the focal plane appear to be converging from the
# perspective of the camera. They are descrinbed by Eq. (4.14)
# for $h_n^{(2)}(kr)$, and have z < 0. We can select the
# appropriate case by applying the correct sign of the imaginary
# part of the starting functions...
if kz > 0:
factor = 1. * 1.j
elif kz < 0:
factor = -1. * 1.j
else:
factor = 0. * 1.j
if not bohren:
factor = -1. * factor
xi_nm2 = coskr + factor * sinkr # \xi_{-1}(kr)
xi_nm1 = sinkr - factor * coskr # \xi_0(kr)
# 2. Angular functions (4.47), page 95
pi_nm1 = 0. # \pi_0(\cos\theta)
pi_n = 1. # \pi_1(\cos\theta)
# 3. Vector spherical harmonics: [r,theta,phi]
mo1nr = 0.j
mo1nt = 0.j
mo1np = 0.j
ne1nr = 0.j
ne1nt = 0.j
ne1np = 0.j
# storage for scattered field
esr = 0.j
est = 0.j
esp = 0.j
# COMPUTE field by summing partial waves
for n in range(1, norders):
n = np.float64(n)
# upward recurrences ...
# 4. Legendre factor (4.47)
# Method described by Wiscombe (1980)
swisc = pi_n * costheta
twisc = swisc - pi_nm1
tau_n = pi_nm1 - n * twisc # -\tau_n(\cos\theta)
# ... Riccati-Bessel function, page 478
xi_n = (2. * n - 1.) * \
(xi_nm1 / kr) - xi_nm2 # \xi_n(kr)
# ... Deirmendjian's derivative
dn = (n * xi_n) / kr - xi_nm1
# vector spherical harmonics (4.50)
mo1nt = pi_n * xi_n # ... divided by cosphi/kr
mo1np = tau_n * xi_n # ... divided by sinphi/kr
# ... divided by cosphi sintheta/kr^2
ne1nr = n * (n + 1.) * pi_n * xi_n
ne1nt = tau_n * dn # ... divided by cosphi/kr
ne1np = pi_n * dn # ... divided by sinphi/kr
mod = n % 4
if mod == 1:
fac = 1.j
elif mod == 2:
fac = -1. + 0.j
elif mod == 3:
fac = -0. - 1.j
else:
fac = 1. + 0.j
# prefactor, page 93
en = fac * (2. * n + 1.) / \
n / (n + 1.)
# the scattered field in spherical coordinates (4.45)
esr += (1.j * en * ab[int(n), 0]) * ne1nr
est += (1.j * en * ab[int(n), 0]) * ne1nt
esp += (1.j * en * ab[int(n), 0]) * ne1np
esr -= (en * ab[int(n), 1]) * mo1nr
est -= (en * ab[int(n), 1]) * mo1nt
esp -= (en * ab[int(n), 1]) * mo1np
# upward recurrences ...
# ... angular functions (4.47)
# Method described by Wiscombe (1980)
pi_nm1 = pi_n
pi_n = swisc + ((n + 1.) / n) * twisc
# ... Riccati-Bessel function
xi_nm2 = xi_nm1
xi_nm1 = xi_n
# n: multipole sum
# geometric factors were divided out of the vector
# spherical harmonics for accuracy and efficiency ...
# ... put them back at the end.
radialfactor = 1. / kr
esr *= cosphi * sintheta * radialfactor**2
est *= cosphi * radialfactor
esp *= sinphi * radialfactor
# By default, the scattered wave is returned in spherical
# coordinates. Project components onto Cartesian coordinates.
# Assumes that the incident wave propagates along z and
# is linearly polarized along x
if cartesian:
ecx = esr * sintheta * cosphi
ecx += est * costheta * cosphi
ecx -= esp * sinphi
ecy = esr * sintheta * sinphi
ecy += est * costheta * sinphi
ecy += esp * cosphi
ecz = (esr * costheta - est * sintheta)
result[0, idx] += ecx * phase
result[1, idx] += ecy * phase
result[2, idx] += ecz * phase
else:
result[0, idx] += esr * phase
result[1, idx] += est * phase
result[2, idx] += esp * phase
def inner(u, v):
sumUV = 0
for i in nb.prange(len(u)):
sumUV += u[i] * v[i]
return sumUV
def monte_carlo_slip_tendency(pole, pole_unc, stress_tensor, stress_unc, axis, axis_unc, pf, mu,
mu_unc, pf_l=0.5, pf_u=2.5):
"""
Computes fault slip tendency for a specific planar node defined by a pole.
Parameters
----------
pole : numpy.ndarray
Pole to fault plane (1x3)
pole_unc : numpy.ndarray
Pole uncertainty (1x3)
stress_tensor : numpy.ndarray
3x3 array with principal stresses
stress_unc : numpy.ndarray
3x3 array with stress uncertainty
axis : numpy.ndarray
1x3 vector with sigma-1 orientation
axis_unc : numpy.ndarray
uncertainty for sigma-1 orientation
pf : float
Best Guess pore fluid pressure at depth
mu : float
Coeffecient of static friction for fault
mu_unc : float
Uncertainty for mu
Returns
-------
out_data : numpy.ndarray
nx3 array [fluid pressure, mu, slip tendency]
"""
# n_sims = 10000
# pf_range = 25.
# lower_pf = -0.2
# upper_pf = 5.9
# lower_pf = 0.
# upper_pf = 5.
upper_pf = pf_u
lower_pf = pf_l
# initialize uncertainty bounds
princ_stress_vec = np.array([stress_tensor[0, 0], stress_tensor[1, 1], stress_tensor[2, 2]])
princ_stress_unc = np.array([stress_unc[0, 0], stress_unc[1, 1], stress_unc[2, 2]])
# main simulation loop
out_data = np.empty((nsims, 3))
for i in numba.prange(nsims):
pole_rand = np.random.randn(3)
pole1 = (pole_unc * pole_rand) + pole
pole1 = pole1 / np.linalg.norm(pole1)
stress_rand = np.random.randn(3)
stress_rand = (princ_stress_unc * stress_rand) + princ_stress_vec
stress1 = stress_rand * np.identity(3)
axis_rand = np.random.randn(3)
axis1 = (axis_unc * axis_rand) + axis
axis1 = axis1 / np.linalg.norm(axis1)
# pf1 = np.random.random() * (pf + pf_range)
hydro1 = pf - lower_pf
hydro2 = pf + upper_pf
pf1 = (hydro2 - hydro1) * np.random.random() + hydro1
mu1 = (np.random.randn() * mu_unc) + mu
plane_stress = rotate_plane_stress(axis1, pole1, stress1)
pole_N_centered = np.array([1., 0., 0.])
pole_N_centered = pole_N_centered / np.linalg.norm(pole_N_centered)
# Leave this in for SHAME
sigma_n = plane_stress[0, 0]
sigma_tyx = plane_stress[1, 0]
sigma_tzx = plane_stress[2, 0]
sigma_t = np.sqrt((sigma_tyx ** 2) + (sigma_tzx ** 2))
# sigma_n_vec = plane_stress @ pole_N_centered
# sigma_n = np.sqrt(sigma_n_vec.dot(sigma_n_vec))
# sigma_t_mat = np.sqrt(np.abs(plane_stress @ plane_stress) - (sigma_n ** 2))
# sigma_t = np.nansum(sigma_t_mat[:])
# slip_tendency = sigma_t / sigma_n
sigma_n_eff = sigma_n - pf1
slip_tendency_eff = sigma_t / sigma_n_eff
out_data[i, 0] = pf1
out_data[i, 1] = mu1
out_data[i, 2] = slip_tendency_eff
return out_data
def transform(X, parameters):
num_examples, input_length = X.shape
dilations, num_features_per_dilation, biases = parameters
# equivalent to:
# >>> from itertools import combinations
# >>> indices = np.array([_ for _ in combinations(np.arange(9), 3)], dtype = np.int32)
indices = np.array(
(0, 1, 2, 0, 1, 3, 0, 1, 4, 0, 1, 5, 0, 1, 6, 0, 1, 7, 0, 1, 8, 0, 2,
3, 0, 2, 4, 0, 2, 5, 0, 2, 6, 0, 2, 7, 0, 2, 8, 0, 3, 4, 0, 3, 5, 0,
3, 6, 0, 3, 7, 0, 3, 8, 0, 4, 5, 0, 4, 6, 0, 4, 7, 0, 4, 8, 0, 5, 6,
0, 5, 7, 0, 5, 8, 0, 6, 7, 0, 6, 8, 0, 7, 8, 1, 2, 3, 1, 2, 4, 1, 2,
5, 1, 2, 6, 1, 2, 7, 1, 2, 8, 1, 3, 4, 1, 3, 5, 1, 3, 6, 1, 3, 7, 1,
3, 8, 1, 4, 5, 1, 4, 6, 1, 4, 7, 1, 4, 8, 1, 5, 6, 1, 5, 7, 1, 5, 8,
1, 6, 7, 1, 6, 8, 1, 7, 8, 2, 3, 4, 2, 3, 5, 2, 3, 6, 2, 3, 7, 2, 3,
8, 2, 4, 5, 2, 4, 6, 2, 4, 7, 2, 4, 8, 2, 5, 6, 2, 5, 7, 2, 5, 8, 2,
6, 7, 2, 6, 8, 2, 7, 8, 3, 4, 5, 3, 4, 6, 3, 4, 7, 3, 4, 8, 3, 5, 6,
3, 5, 7, 3, 5, 8, 3, 6, 7, 3, 6, 8, 3, 7, 8, 4, 5, 6, 4, 5, 7, 4, 5,
8, 4, 6, 7, 4, 6, 8, 4, 7, 8, 5, 6, 7, 5, 6, 8, 5, 7, 8, 6, 7, 8),
dtype=np.int32).reshape(84, 3)
num_kernels = len(indices)
num_dilations = len(dilations)
num_features = num_kernels * np.sum(num_features_per_dilation)
features = np.zeros((num_examples, num_features), dtype=np.float32)
for example_index in prange(num_examples):
_X = X[example_index]
A = -_X # A = alpha * X = -X
G = _X + _X + _X # G = gamma * X = 3X
feature_index_start = 0
for dilation_index in range(num_dilations):
_padding0 = dilation_index % 2
dilation = dilations[dilation_index]
padding = ((9 - 1) * dilation) // 2
num_features_this_dilation = num_features_per_dilation[
dilation_index]
C_alpha = np.zeros(input_length, dtype=np.float32)
C_alpha[:] = A
C_gamma = np.zeros((9, input_length), dtype=np.float32)
C_gamma[9 // 2] = G
start = dilation
end = input_length - padding
for gamma_index in range(9 // 2):
C_alpha[-end:] = C_alpha[-end:] + A[:end]
C_gamma[gamma_index, -end:] = G[:end]
end += dilation
for gamma_index in range(9 // 2 + 1, 9):
C_alpha[:-start] = C_alpha[:-start] + A[start:]
C_gamma[gamma_index, :-start] = G[start:]
start += dilation
for kernel_index in range(num_kernels):
feature_index_end = feature_index_start + num_features_this_dilation
_padding1 = (_padding0 + kernel_index) % 2
index_0, index_1, index_2 = indices[kernel_index]
C = C_alpha + C_gamma[index_0] + C_gamma[index_1] + C_gamma[
index_2]
if _padding1 == 0:
for feature_count in range(num_features_this_dilation):
features[example_index,
feature_index_start + feature_count] = _PPV(
C, biases[feature_index_start +
feature_count]).mean()
else:
for feature_count in range(num_features_this_dilation):
features[example_index,
feature_index_start + feature_count] = _PPV(
C[padding:-padding],
biases[feature_index_start +
feature_count]).mean()
feature_index_start = feature_index_end
return features
def _proximity(img, x_coords, y_coords, target_values, distance_metric):
height, width = img.shape
max_distance = x_coords[width - 1] + y_coords[height - 1]
pan_near_x = np.zeros(width, dtype=np.int64)
pan_near_y = np.zeros(width, dtype=np.int64)
# output of the function
img_proximity = np.zeros(shape=(height, width), dtype=np.float64)
# Loop from top to bottom of the image.
for i in prange(width):
pan_near_x[i] = -1
pan_near_y[i] = -1
scan_line = np.zeros(width, dtype=img.dtype)
for line in prange(height):
# Read for target values.
for i in prange(width):
scan_line[i] = img[line][i]
line_proximity = np.zeros(width, dtype=np.float64)
for i in prange(width):
line_proximity[i] = -1.0
# left to right
_process_proximity_line(scan_line, x_coords, y_coords, pan_near_x,
pan_near_y, True, line, width, max_distance,
line_proximity, target_values, distance_metric)
# right to left
_process_proximity_line(scan_line, x_coords, y_coords, pan_near_x,
pan_near_y, False, line, width, max_distance,
line_proximity, target_values, distance_metric)
for i in prange(width):
img_proximity[line][i] = line_proximity[i]
# Loop from bottom to top of the image.
for i in prange(width):
pan_near_x[i] = -1
pan_near_y[i] = -1
for line in prange(height - 1, -1, -1):
# Read first pass proximity.
for i in prange(width):
line_proximity[i] = img_proximity[line][i]
# Read pixel target_values.
for i in prange(width):
scan_line[i] = img[line][i]
# Right to left
_process_proximity_line(scan_line, x_coords, y_coords, pan_near_x,
pan_near_y, False, line, width, max_distance,
line_proximity, target_values, distance_metric)
# Left to right
_process_proximity_line(scan_line, x_coords, y_coords, pan_near_x,
pan_near_y, True, line, width, max_distance,
line_proximity, target_values, distance_metric)
# final post processing of distances
for i in prange(width):
if line_proximity[i] < 0 or np.isnan(scan_line[i]):
# this corresponds the the nan value of input raster.
line_proximity[i] = np.nan
for i in prange(width):
img_proximity[line][i] = line_proximity[i]
return img_proximity
def _outer_dot(v, X, n_samples, window_size, n_windows):
X_new = np.empty((n_samples, window_size, window_size, n_windows))
for i in prange(n_samples):
for j in prange(window_size):
X_new[i, j] = np.dot(np.outer(v[i, :, j], v[i, :, j]), X[i])
return X_new
def calc_trans(data_i_s_):
for i in nb.prange(num_pop_assigned):
for t in range(1, sim_time):
data_i_s_[i, t] = transit(data_i_s_[i, t-1], data_rnd[i, t])
def define_principal_stresses(sv1, depth, shmin1, shmax1, hminaz, hmaxaz, sv_unc=10, shmin_unc=10, shmax_unc=10,
v_tilt_unc=10, h_az_unc=10, is_3d=False):
"""
Generates cauchy stress tensor from principal stress directions, assumes vertical stress in one direction
rotates sigma1 (maximum principal stress) direction to plane normal
Parameters
----------
sv : float
vertical stress at depth
depth : float
Depth of analysis (sv = sv_grad * depth)
shmin : float
minimum horizontal stress at depth
shmax : float
maximum horizontal stress at depth
hminaz : float
minimum horizontal stress direction
hmaxaz : float
maximum horizontal stress direction
sv_unc : float
Vertical stress uncertainty at depth
shmax_unc : float
maximum horizontal stress uncertainty
shmin_unc : float
Minimum horizontal stress uncertainty
v_tilt_unc : float
tilt uncertainty for vertical stress
h_az_unc : float
Horizontal stress orientation uncertainty
is_3d : bool
Is this model 3d? returns gradients if so
Returns
-------
princ_stress_tensor : numpy.ndarray
3x3 stress tensor aligned to principal stress orientations
princ_stress_unc : numpy.ndarray
axis : numpy.ndarray
Sigma-1 stress unit axis (which is rotated to align with plane for normal / shear stress analysis)
axis_unc : numpy.ndarray
uncertainty for sigma-1 unit axis
axis_out : numpy.ndarray
"""
nsim = nsims
h_az_unc = math.radians(h_az_unc)
v_tilt_unc = math.radians(v_tilt_unc)
if round(abs(hmaxaz - hminaz), 0) != 90.:
raise ValueError('hmin and hmax are not orthogonal')
# TODO: Truly fix azimuth issue with stress. Currently add 90 degrees to max stress direction.
# hmaxaz = math.radians(hmaxaz) + (math.pi/2)
hmaxaz = math.radians(hmaxaz)
if is_3d:
sv1 = sv1 / depth
shmax1 = shmax1 / depth
shmin1 = shmin1 / depth
DeprecationWarning("Stresses should be provided in terms of gradients. Otherwise may break things.")
axis_out = np.zeros((nsim, 3))
stress_out = np.zeros((nsim, 3))
for i in numba.prange(nsim):
sv = sv_unc * np.random.randn() + sv1
shmax = shmax_unc * np.random.randn() + shmax1
shmin = shmin_unc * np.random.randn() + shmin1
if sv1 > shmax1 > shmin1:
sigma1 = sv
sigma2 = shmax
sigma3 = shmin
az_rad = h_az_unc * np.random.randn() + hmaxaz
az_dip_rad = v_tilt_unc * np.random.randn() + math.pi / 2.
rotated_axis = np.asarray([math.sin(az_rad) * math.cos(az_dip_rad), math.cos(az_rad) * math.cos(az_dip_rad),
math.sin(az_dip_rad)])
else:
sigma1 = shmax
az_rad = h_az_unc * np.random.randn() + hmaxaz
az_dip_rad = v_tilt_unc * np.random.randn() + 0. # average dip is 0
if shmax1 > sv1 > shmin1:
sigma2 = sv
sigma3 = shmin
rotated_axis = np.asarray(
[math.sin(az_rad) * math.cos(az_dip_rad), math.cos(az_rad) * math.cos(az_dip_rad),
math.sin(az_dip_rad)])
elif shmax1 > shmin1 > sv1:
sigma2 = shmin
sigma3 = sv
rotated_axis = np.asarray(
[math.sin(az_rad) * math.cos(az_dip_rad), math.cos(az_rad) * math.cos(az_dip_rad),
math.sin(az_dip_rad)])
else:
raise ValueError('Unable to resolve principal stress orientations')
axis_out[i, :] = rotated_axis
stress_out[i, :] = np.asarray([sigma1, sigma2, sigma3])
sigma1_mean = np.mean(stress_out[:, 0])
sigma2_mean = np.mean(stress_out[:, 1])
sigma3_mean = np.mean(stress_out[:, 2])
sigma1_std = np.std(stress_out[:, 0])
sigma2_std = np.std(stress_out[:, 1])
sigma3_std = np.std(stress_out[:, 2])
rotated_axis = np.array([np.mean(axis_out[:, 0]), np.mean(axis_out[:, 1]), np.mean(axis_out[:, 2])])
rotated_axis = rotated_axis / np.linalg.norm(rotated_axis)
axis_std = np.array([np.std(axis_out[:, 0]), np.std(axis_out[:, 1]), np.std(axis_out[:, 2])])
princ_stress_tensor = np.array([[sigma1_mean, 0., 0.], [0., sigma2_mean, 0.], [0., 0., sigma3_mean]])
princ_stress_tensor_unc = np.array([[sigma1_std, 0., 0.], [0., sigma2_std, 0.], [0., 0., sigma3_std]])
# rotated_axis = np.array(rotated_axis)
return princ_stress_tensor, rotated_axis, princ_stress_tensor_unc, axis_std, axis_out
def _prescraamp(
T_A,
T_B,
m,
T_A_subseq_isfinite,
T_B_subseq_isfinite,
p,
indices,
s,
excl_zone=None,
):
"""
A Numba JIT-compiled implementation of the non-normalized (i.e., without
z-normalization) preSCRIMP algorithm.
Parameters
----------
T_A : numpy.ndarray
The time series or sequence for which to compute the matrix profile
T_B : numpy.ndarray
The time series or sequence that will be used to annotate T_A. For every
subsequence in T_A, its nearest neighbor in T_B will be recorded.
m : int
Window size
T_A_subseq_isfinite : numpy.ndarray
A boolean array that indicates whether a subsequence in `T_A` contains a
`np.nan`/`np.inf` value (False)
T_B_subseq_isfinite : numpy.ndarray
A boolean array that indicates whether a subsequence in `T_B` contains a
`np.nan`/`np.inf` value (False)
p : float, default 2.0
The p-norm to apply for computing the Minkowski distance.
i : int
The subsequence index in `T_B` that corresponds to `Q`
s : int
The sampling interval that defaults to
`int(np.ceil(m / config.STUMPY_EXCL_ZONE_DENOM))`
p_norm_profile : numpy.ndarray
A reusable array to store the computed p-norm distance profile
P : numpy.ndarray
The squared matrix profile
I : numpy.ndarray
The matrix profile indices
excl_zone : int
The half width for the exclusion zone relative to the `i`.
Notes
-----
`DOI: 10.1109/ICDM.2018.00099 \
<https://www.cs.ucr.edu/~eamonn/SCRIMP_ICDM_camera_ready_updated.pdf>`__
See Algorithm 2
"""
n_threads = numba.config.NUMBA_NUM_THREADS
l = T_A.shape[0] - m + 1
P_NORM = np.full((n_threads, l), np.inf, dtype=np.float64)
I = np.full((n_threads, l), -1, dtype=np.int64)
idx_ranges = core._get_ranges(len(indices), n_threads, truncate=False)
for thread_idx in prange(n_threads):
_compute_PI(
T_A,
T_B,
m,
T_A_subseq_isfinite,
T_B_subseq_isfinite,
p,
indices,
idx_ranges[thread_idx, 0],
idx_ranges[thread_idx, 1],
thread_idx,
s,
P_NORM,
I,
excl_zone,
)
for thread_idx in range(1, n_threads):
for i in range(l):
if P_NORM[thread_idx, i] < P_NORM[0, i]:
P_NORM[0, i] = P_NORM[thread_idx, i]
I[0, i] = I[thread_idx, i]
return np.power(P_NORM[0], 1.0 / p), I[0]
def _aamp(T_A, T_B, m, T_A_subseq_isfinite, T_B_subseq_isfinite, p, diags,
ignore_trivial):
"""
A Numba JIT-compiled version of AAMP for parallel computation of the matrix
profile and matrix profile indices.
Parameters
----------
T_A : numpy.ndarray
The time series or sequence for which to compute the matrix profile
T_B : numpy.ndarray
The time series or sequence that will be used to annotate T_A. For every
subsequence in T_A, its nearest neighbor in T_B will be recorded.
m : int
Window size
T_A_subseq_isfinite : numpy.ndarray
A boolean array that indicates whether a subsequence in `T_A` contains a
`np.nan`/`np.inf` value (False)
T_B_subseq_isfinite : numpy.ndarray
A boolean array that indicates whether a subsequence in `T_B` contains a
`np.nan`/`np.inf` value (False)
p : float
The p-norm to apply for computing the Minkowski distance.
diags : numpy.ndarray
The diag of diagonals to process and compute
ignore_trivial : bool
Set to `True` if this is a self-join. Otherwise, for AB-join, set this to
`False`. Default is `True`.
Returns
-------
P : numpy.ndarray
Matrix profile
I : numpy.ndarray
Matrix profile indices
Notes
-----
`DOI: 10.1109/ICDM.2018.00099 \
<https://www.cs.ucr.edu/~eamonn/SCRIMP_ICDM_camera_ready_updated.pdf>`__
See Algorithm 1
"""
n_A = T_A.shape[0]
n_B = T_B.shape[0]
l = n_A - m + 1
n_threads = numba.config.NUMBA_NUM_THREADS
P = np.full((n_threads, l, 3), np.inf, dtype=np.float64)
I = np.full((n_threads, l, 3), -1, dtype=np.int64)
ndist_counts = core._count_diagonal_ndist(diags, m, n_A, n_B)
diags_ranges = core._get_array_ranges(ndist_counts, n_threads, False)
for thread_idx in prange(n_threads):
# Compute and update P, I within a single thread while avoiding race conditions
_compute_diagonal(
T_A,
T_B,
m,
T_A_subseq_isfinite,
T_B_subseq_isfinite,
p,
diags,
diags_ranges[thread_idx, 0],
diags_ranges[thread_idx, 1],
thread_idx,
P,
I,
ignore_trivial,
)
# Reduction of results from all threads
for thread_idx in range(1, n_threads):
for i in prange(l):
if P[0, i, 0] > P[thread_idx, i, 0]:
P[0, i, 0] = P[thread_idx, i, 0]
I[0, i, 0] = I[thread_idx, i, 0]
# left matrix profile and left matrix profile indices
if P[0, i, 1] > P[thread_idx, i, 1]:
P[0, i, 1] = P[thread_idx, i, 1]
I[0, i, 1] = I[thread_idx, i, 1]
# right matrix profile and right matrix profile indices
if P[0, i, 2] > P[thread_idx, i, 2]:
P[0, i, 2] = P[thread_idx, i, 2]
I[0, i, 2] = I[thread_idx, i, 2]
return np.power(P[0, :, :], 1.0 / p), I[0, :, :]
def dotest(n):
x = np.random.rand(n)
y = np.random.rand(n)
z = np.empty_like(x)
for i in numba.prange(x.shape[0]):
z[i] = math.sqrt(x[i] * x[i] + y[i] * y[i])
def dists(col_map, pxl):
dists = np.empty(col_map.shape[0], dtype=np.double)
for i in prange(col_map.shape[0]):
dists[i] = col_dist(col_map[i], pxl)
return dists
def f(n):
A = empty(n)
for i in prange(n):
S = np.arange(i)
A[i] = S.sum()
return A