def sortBatches2(ccb0):
# takes as input a matrix of nBatches by nBatches containing
# dissimilarities.
# outputs a matrix of sorted batches, and the sorting order, such that
# ccb1 = ccb0(isort, isort)
# put this matrix on the GPU
ccb0 = cp.asarray(ccb0, order='F')
# compute its svd on the GPU (this might also be fast enough on CPU)
u, s, v = svdecon(ccb0)
# HACK: consistency with MATLAB
u = u * cp.sign(u[0, 0])
v = v * cp.sign(u[0, 0])
# initialize the positions xs of the batch embeddings to be very small but proportional to
# the first PC
xs = .01 * u[:, 0] / cp.std(u[:, 0], ddof=1)
# 200 iterations of gradient descent should be enough
# TODO: move_to_config
niB = 200
# this learning rate should usually work fine, since it scales with the average gradient
# and ccb0 is z-scored
# TODO: move_to_config
eta = 1
for k in tqdm(range(niB), desc="Sorting %d batches" % ccb0.shape[0]):
# euclidian distances between 1D embedding positions
ds = (xs - xs[:, np.newaxis])**2
# the transformed distances go through this function
W = cp.log(1 + ds)
# the error is the difference between ccb0 and W
err = ccb0 - W
# ignore the mean value of ccb0
err = err - cp.mean(err, axis=0)
# backpropagate the gradients
err = err / (1 + ds)
err2 = err * (xs[:, np.newaxis] - xs)
D = cp.mean(
err2, axis=1) # one half of the gradients is along this direction
E = cp.mean(err2, axis=0) # the other half is along this direction
# we don't need to worry about the gradients for the diagonal because those are 0
# final gradients for the embedding variable
dx = -D + E.T
# take a gradient step
xs = xs - eta * dx
# sort the embedding positions xs
isort = cp.argsort(xs, axis=0)
# sort the matrix of dissimilarities
ccb1 = ccb0[isort, :][:, isort]
return ccb1, isort
def _svd_flip(u, v, u_based_decision=True):
"""Sign correction to ensure deterministic output from SVD.
Adjusts the columns of u and the rows of v such that the loadings in the
columns in u that are largest in absolute value are always positive.
Parameters
----------
u : cupy.ndarray
u and v are the output of `cupy.linalg.svd`
v : cupy.ndarray
u and v are the output of `cupy.linalg.svd`
u_based_decision : boolean, (default=True)
If True, use the columns of u as the basis for sign flipping.
Otherwise, use the rows of v. The choice of which variable to base the
decision on is generally algorithm dependent.
Returns
-------
u_adjusted, v_adjusted : arrays with the same dimensions as the input.
"""
if u_based_decision:
# columns of u, rows of v
max_abs_cols = cp.argmax(cp.abs(u), axis=0)
signs = cp.sign(u[max_abs_cols, range(u.shape[1])])
u *= signs
v *= signs[:, cp.newaxis]
else:
# rows of v, columns of u
max_abs_rows = cp.argmax(cp.abs(v), axis=1)
signs = cp.sign(v[list(range(v.shape[0])), max_abs_rows])
u *= signs
v *= signs[:, cp.newaxis]
return u, v
def iterative_least_likely(self, eps, alpha =1, n_iter = None, index = None):
xp = AdvImage.xp
if n_iter is None:
n_iter = int(min(eps + 4, 1.25 * eps))
if index is None:
probs = AdvImage.model.predict([self.org_image], oversample=False).data[0]
probs = cuda.to_cpu(probs)
least_index = np.argmin(probs)
t = xp.array([least_index], dtype=xp.int32)
out_layer = AdvImage.last_layer
target_org = self.target.data.copy()
for _ in range(n_iter):
x = AdvImage.model(self.target, layers=[out_layer])[out_layer]
loss = F.softmax_cross_entropy(x, t)
self.target.cleargrad()
AdvImage.model.cleargrads()
loss.backward()
perturb = xp.sign(self.target.grad)
updated_data = self.target.data - alpha * perturb
clipped_data = xp.clip(updated_data, target_org - eps, target_org + eps)
self.target = Variable(clipped_data)
self.adv_image = self._restore_image(self.target)
def signedpower(x, a):
#将inf和-inf处理为nan
#x[cp.isinf(x)] = cp.nan
# 经测试where替换更快一些
x = cp.where(cp.isinf(x), cp.nan, x)
return cp.sign(x) * cp.abs(x)**a
def constrain_variable_probe(variable_probe, weights):
"""Add the following constraints to variable probe weights
1. Remove outliars from weights
2. Enforce orthogonality once per epoch
"""
logger.info('Orthogonalize variable probes')
variable_probe = tike.linalg.orthogonalize_gs(
variable_probe,
axis=(-3, -2, -1),
)
logger.info('Remove outliars from variable probe weights')
aevol = cp.abs(weights)
weights = cp.minimum(
aevol,
1.5 * cp.percentile(
aevol,
[95],
axis=[-3],
keepdims=True,
).astype(weights.dtype),
) * cp.sign(weights)
# TODO: Smooth the weights as a function of the frame index.
return variable_probe, weights
def natural_compression(x):
dim = x.shape[0]
logx = xp.ma.log2(xp.abs(x)).filled(-15)
logx_floor = xp.floor(logx)
noise = xp.random.uniform(0.0, 1.0, dim)
leftx = xp.exp2(logx_floor)
rounded = xp.floor(xp.ma.log2(xp.abs(x) + leftx * noise).filled(-15))
compressed = xp.sign(x) * xp.exp2(rounded)
return compressed
def sign(x: Array, /) -> Array:
"""
Array API compatible wrapper for :py:func:`np.sign <numpy.sign>`.
See its docstring for more information.
"""
if x.dtype not in _numeric_dtypes:
raise TypeError("Only numeric dtypes are allowed in sign")
return Array._new(np.sign(x._array))
def random_quantization(x, s):
dim = x.shape[0]
xnorm = xp.linalg.norm(x)
if s == 0 or xnorm == 0:
return xp.zeros(dim, dtype=int)
noise = xp.random.uniform(0, 1, dim)
rounded = xp.floor(s * xp.abs(x) / xnorm + noise)
compressed = (xnorm / s) * xp.sign(x) * rounded
return compressed
def tvl1(lp, init_recon, tomo0, num_iter, reg_par, gpu):
"""
Reconstruction with the total variation method
with the regularization parameter reg_par.
Solving the problem ||R(recon)-tomo||_1 + reg_par*TV(recon) -> min
"""
# choose device
cp.cuda.Device(gpu).use()
# Allocating necessary gpu arrays
recon = cp.array(init_recon)
tomo = cp.array(tomo0)
g = tomo * 0
p = recon * 0
recon0 = recon
prox0x = recon * 0
prox0y = recon * 0
div0 = recon * 0
prox1 = tomo * 0
lam = reg_par
c = 0.35 # 1/power_method(lp,tomo,num_iter)
# tv iterations
for i in range(0, num_iter):
# forward step
# compute proximal prox0
prox0x[:, :, :-1] += c * (recon[:, :, 1:] - recon[:, :, :-1])
prox0y[:, :-1, :] += c * (recon[:, 1:, :] - recon[:, :-1, :])
nprox = cp.array(
cp.maximum(1, (cp.sqrt(prox0x * prox0x + prox0y * prox0y) / lam)))
prox0x = prox0x / nprox
prox0y = prox0y / nprox
# compute proximal prox1
lp.fwdp(g, recon, gpu)
tmp = prox1 + c * g - c * tomo
tmp2 = cp.abs(tmp) - 1
tmp2[tmp2 < 0] = 0
prox1 = tmp - tmp2 * cp.sign(tmp)
# backward step
recon = recon0
div0[:, :, 1:] = (prox0x[:, :, 1:] - prox0x[:, :, :-1])
div0[:, :, 0] = prox0x[:, :, 0]
div0[:, 1:, :] += (prox0y[:, 1:, :] - prox0y[:, :-1, :])
div0[:, 0, :] += prox0y[:, 0, :]
lp.adjp(p, prox1, gpu)
recon0 = recon0 - c * p + c * div0
# update recon
recon = 2 * recon0 - recon
return recon.get()
def FW(W, n):
## Initially the distance between all pairs are set to infinity
d = cp.ones([n, n]) * myInf
all_ones = cp.ones([n, n])
for k in range(n):
## Calculating distance in iteration k
dk = cp.log(
cp.dot(cp.exp(W[:, k].reshape(n, 1)), cp.exp(W[k, :].reshape(1,
n))))
a = cp.sign(d - dk)
where_to = cp.where(a > 0)
## Updating distances where necessary
d[where_to] = dk[where_to]
return d
def fast_gradient(self, eps):
xp = AdvImage.xp
out_layer = AdvImage.last_layer
x = AdvImage.model(self.target, layers=[out_layer])[out_layer]
t = xp.array([self.index], dtype=xp.int32)
loss = F.softmax_cross_entropy(x, t)
self.target.cleargrad()
AdvImage.model.cleargrads()
loss.backward()
perturb = xp.sign(self.target.grad)
self.target = Variable(self.target.data + eps * perturb)
self.adv_image = self._restore_image(self.target)
def itkrm(data,K,S,maxitr,startD=np.array([1])):
M, N = data.shape
if startD.all()==1:
D_init = np.random.randn(M, K)
else:
D_init = startD
#Algorithm
GPU_D_old = cp.asarray(D_init)
GPU_Y = cp.asarray(data)
GPU_M = int(cp.asarray(M))
GPU_N = int(cp.asarray(N))
GPU_S = int(cp.asarray(S))
GPU_maxitr = int(cp.asarray(maxitr))
GPU_I_D = cp.zeros((S,N),dtype=cp.int32)
for i in range(GPU_maxitr):
start_time = N_timer.cont_timer(0,0)
N_timer.Timer(i,maxitr)
for n in range(GPU_N):
GPU_I_D[:,n] = max_atoms(GPU_D_old,GPU_Y[:,n],GPU_S)
GPU_D_new = cp.zeros((M,K))
GPU_DtD = GPU_D_old.T @ GPU_D_old
for n in range(GPU_N):
GPU_DtY = GPU_D_old[:,GPU_I_D[:,n]].T @ GPU_Y[:,n]
GPU_matproj = cp.repeat((GPU_D_old[:,GPU_I_D[:,n]] @ cp.linalg.inv(GPU_DtD[GPU_I_D[:,n,None], GPU_I_D[:,n]]) @ GPU_DtY)[:,None],GPU_S,axis=1)
GPU_vecproj = GPU_D_old[:,GPU_I_D[:,n]] @ cp.diag(cp.diag( GPU_DtD[GPU_I_D[:,n,None], GPU_I_D[:,n]] )**-1*( GPU_DtY ))
GPU_signer = cp.sign( GPU_DtY )
GPU_D_new[:,GPU_I_D[:,n]] = GPU_D_new[:,GPU_I_D[:,n]] + (cp.repeat(GPU_Y[:,n,None], S, axis=1) - GPU_matproj + GPU_vecproj)*GPU_signer
GPU_scale = cp.sum(GPU_D_new*GPU_D_new, axis=0)
GPU_iszero = cp.where(GPU_scale < 0.00001)[0]
# GPU_D_new[:,GPU_iszero] = np.random.randn(GPU_M, len(GPU_iszero)) # generate random with GPU
GPU_D_new[:,GPU_iszero] = cp.asarray(np.random.randn(M, len(GPU_iszero))) # generate random with CPU
#end hugget
GPU_D_new = normalize_mat_col(GPU_D_new)
GPU_D_old = 1*GPU_D_new
return cp.asnumpy(GPU_D_old)
def extractPCfromSnippets(proc, probe=None, params=None, Nbatch=None):
# extracts principal components for 1D snippets of spikes from all channels
# loads a subset of batches to find these snippets
NT = params.NT
nPCs = params.nPCs
Nchan = probe.Nchan
batchstart = np.arange(0, NT * Nbatch + 1, NT).astype(np.int64)
# extract the PCA projections
# initialize the covariance of single-channel spike waveforms
CC = cp.zeros(params.nt0, dtype=np.float32)
# from every 100th batch
for ibatch in range(0, Nbatch, 100):
offset = Nchan * batchstart[ibatch]
dat = proc.flat[offset:offset + NT * Nchan].reshape((-1, Nchan),
order='F')
if dat.shape[0] == 0:
continue
# move data to GPU and scale it back to unit variance
dataRAW = cp.asarray(dat, dtype=np.float32) / params.scaleproc
# find isolated spikes from each batch
row, col, mu = isolated_peaks_new(dataRAW, params)
# for each peak, get the voltage snippet from that channel
c = get_SpikeSample(dataRAW, row, col, params)
# scale covariance down by 1,000 to maintain a good dynamic range
CC = CC + cp.dot(c, c.T) / 1e3
# the singular vectors of the covariance matrix are the PCs of the waveforms
U, Sv, V = svdecon(CC)
wPCA = U[:, :nPCs] # take as many as needed
# adjust the arbitrary sign of the first PC so its negativity is downward
# TODO: unclear - is 20 here the index into the spike waveform? Should this be hardcoded?
# - should it be nt0min instead?
wPCA[:, 0] = -wPCA[:, 0] * cp.sign(wPCA[20, 0])
return wPCA
def inverse(
data,
impulse_response=None,
filter_params={},
max_gain=2,
predefined_filter=None,
):
"""Apply the filter in reverse to the given data.
Parameters
----------
data : (M,N) ndarray
Input data.
impulse_response : callable `f(r, c, **filter_params)`
Impulse response of the filter. See LPIFilter2D.__init__.
filter_params : dict
Additional keyword parameters to the impulse_response function.
max_gain : float
Limit the filter gain. Often, the filter contains zeros, which would
cause the inverse filter to have infinite gain. High gain causes
amplification of artefacts, so a conservative limit is recommended.
Other Parameters
----------------
predefined_filter : LPIFilter2D
If you need to apply the same filter multiple times over different
images, construct the LPIFilter2D and specify it here.
"""
check_nD(data, 2, "data")
if predefined_filter is None:
filt = LPIFilter2D(impulse_response, **filter_params)
else:
filt = predefined_filter
F, G = filt._prepare(data)
_min_limit(F)
F = 1 / F
mask = cp.abs(F) > max_gain
F[mask] = cp.sign(F[mask]) * max_gain
return _centre(cp.abs(fft.ifftshift(fft.ifftn(G * F))), data.shape)
def iterative_gradient(self, eps, alpha =1, n_iter = None):
xp = AdvImage.xp
if n_iter is None:
n_iter = int(min(eps + 4, 1.25 * eps))
t = xp.array([self.index], dtype=xp.int32)
out_layer = AdvImage.last_layer
target_org = self.target.data.copy()
for _ in range(n_iter):
x = AdvImage.model(self.target, layers=[out_layer])[out_layer]
loss = F.softmax_cross_entropy(x, t)
self.target.cleargrad()
AdvImage.model.cleargrads()
loss.backward()
perturb = xp.sign(self.target.grad)
updated_data = self.target.data + alpha * perturb
clipped_data = xp.clip(updated_data, target_org - eps, target_org + eps)
self.target = Variable(clipped_data)
self.adv_image = self._restore_image(self.target)
def FISTA(grad_f, x_0, L, LAMBDA, n_iters=100, eps=1e-10):
'''FISTA'''
r = xp.zeros(n_iters + 1)
for t in range(1, n_iters + 1):
r[t] = 0.5 + xp.sqrt(1 + 4 * r[t - 1]**2) / 2
gamma = (1 - r[:n_iters]) / r[1:]
x = x_0.copy()
y = x_0.copy()
for t in range(1, n_iters):
_grad = grad_f(x)
if xp.linalg.norm(_grad) < eps:
break
x -= _grad / L
y_new = xp.sign(x) * xp.maximum(xp.abs(x) - LAMBDA / L, 0)
x = (1 - gamma[t]) * y_new + gamma[t] * y
y = y_new
return y, t + 1
def fixpix(data, mask, out=None, dtype=None, fix_NaN=False):
'''
fill the bad pixel with mean of surrounding pixels
Parameters
----------
data : ndarray
An array of image
If a 3D array containing multiple images,
the images must be stacked along the 1st dimension (axis=0).
mask : ndarray
An array indicates bad pixel positions
The shape must be same as image.
The value of bad pixel is nonzero, and the others is 0.
If all pixels are bad, raise ValueError.
out : cupy.ndarray, default None
Alternate output array in which to place the result. The default
is ``None``; if provided, it must have the same shape as the
expected output, but the type will be cast if necessary.
dtype : str or dtype, default 'float32'
dtype of array used internally
If None, this value will be usually "float32",
but this can be changed with eclair.set_dtype.
If the input dtype is different, use a casted copy.
fix_NaN : bool, default False
If true, fix NaN pixel even if it's not bad pixel in the mask.
Returns
-------
fixed : ndarray
An array of images fixed bad pixel
Notes
-----
NaN is ignored in interpolation calculations,
but is not fixed if fix_NaN is False.
'''
dtype = judge_dtype(dtype)
if data.shape[-2:] != mask.shape[-2:]:
raise ValueError('shape differs between data and mask')
elif mask.all():
raise ValueError('No available pixel')
data = cp.array(data, dtype=dtype, copy=False, ndmin=3)
mask = cp.array(mask, dtype=dtype, copy=False, ndmin=3)
convolution = lambda data, out: conv_kernel(data, out)
if out is None:
out = cp.empty_like(data)
cp.copyto(out, data)
filt = mask2filter(mask)
if fix_NaN:
filt = checkfinite(data, filt)
out *= filt
dconv = cp.empty_like(out)
nconv = cp.empty_like(filt)
while not filt.all():
convolution(out, dconv)
convolution(filt, nconv)
fix_kernel(out, filt, dconv, nconv, out)
cp.sign(nconv, out=filt)
return cp.squeeze(out)
def gsgd(x, b):
norm = xp.linalg.norm(x, axis=0)
return norm / (2**(b - 1)) * xp.sign(x) * xp.floor(
(2**(b - 1)) / norm * xp.abs(x) + xp.random.uniform(0, 1, x.shape))
def extractTemplatesfromSnippets(proc=None,
probe=None,
params=None,
Nbatch=None,
nPCs=None):
# this function is very similar to extractPCfromSnippets.
# outputs not just the PC waveforms, but also the template "prototype",
# basically k-means clustering of 1D waveforms.
NT = params.NT
# skip every this many batches
nskip = params.nskip
nPCs = nPCs or params.nPCs
nt0min = params.nt0min
Nchan = probe.Nchan
batchstart = np.arange(0, NT * Nbatch + 1, NT).astype(np.int64)
k = 0
# preallocate matrix to hold 1D spike snippets
# dd = cp.zeros((params.nt0, int(5e4)), dtype=np.float32, order='F')
dds = []
for ibatch in tqdm(range(0, Nbatch, nskip), desc="Extracting templates"):
offset = Nchan * batchstart[ibatch]
dat = proc.flat[offset:offset + NT * Nchan].reshape((-1, Nchan),
order='F')
# move data to GPU and scale it back to unit variance
dataRAW = cp.asarray(dat, dtype=np.float32) / params.scaleproc
# find isolated spikes from each batch
row, col, mu = isolated_peaks_new(dataRAW, params)
# for each peak, get the voltage snippet from that channel
c = get_SpikeSample(dataRAW, row, col, params)
# if k + c.shape[1] > dd.shape[1]:
# dd = cp.pad(dd, (0, dd.shape[1]), mode='constant')
# dd[:, k:k + c.shape[1]] = c
dds.append(c)
k = k + c.shape[1]
if k > 1e5:
break
# discard empty samples
# dd = dd[:, :k]
dd = cp.asfortranarray(cp.concatenate(dds, axis=1).astype(np.float32))
# initialize the template clustering with random waveforms
uu = np.random.permutation(dd.shape[1])[:nPCs]
wTEMP = dd[:, uu]
wTEMP = wTEMP / cp.sum(wTEMP**2, axis=0)**.5 # normalize them
for i in range(10):
# at each iteration, assign the waveform to its most correlated cluster
cc = cp.dot(wTEMP.T, dd)
imax = cp.argmax(cc, axis=0)
amax = cc[imax, np.arange(cc.shape[1])]
for j in range(nPCs):
# weighted average to get new cluster means
wTEMP[:, j] = cp.dot(dd[:, imax == j], amax[imax == j].T)
wTEMP = wTEMP / cp.sum(wTEMP**2, axis=0)**.5 # unit normalize
# the PCs are just the left singular vectors of the waveforms
U, Sv, V = svdecon(dd)
# take as many as needed
wPCA = U[:, :nPCs]
# adjust the arbitrary sign of the first PC so its negativity is downward
wPCA[:, 0] = -wPCA[:, 0] * cp.sign(wPCA[nt0min, 0])
return wTEMP, wPCA
def _tvl1(
reference_image,
moving_image,
flow0,
attachment,
tightness,
num_warp,
num_iter,
tol,
prefilter,
):
"""TV-L1 solver for optical flow estimation.
Parameters
----------
reference_image : ndarray, shape (M, N[, P[, ...]])
The first gray scale image of the sequence.
moving_image : ndarray, shape (M, N[, P[, ...]])
The second gray scale image of the sequence.
flow0 : ndarray, shape (image0.ndim, M, N[, P[, ...]])
Initialization for the vector field.
attachment : float
Attachment parameter. The smaller this parameter is,
the smoother is the solutions.
tightness : float
Tightness parameter. It should have a small value in order to
maintain attachement and regularization parts in
correspondence.
num_warp : int
Number of times image1 is warped.
num_iter : int
Number of fixed point iteration.
tol : float
Tolerance used as stopping criterion based on the L² distance
between two consecutive values of (u, v).
prefilter : bool
Whether to prefilter the estimated optical flow before each
image warp.
Returns
-------
flow : ndarray, shape ((image0.ndim, M, N[, P[, ...]])
The estimated optical flow components for each axis.
"""
dtype = reference_image.dtype
grid = cp.stack(
cp.meshgrid(
*[cp.arange(n, dtype=dtype) for n in reference_image.shape],
indexing="ij",
),
axis=0,
)
dt = 0.5 / reference_image.ndim
reg_num_iter = 2
f0 = attachment * tightness
f1 = dt / tightness
tol *= reference_image.size
flow_current = flow_previous = flow0
g = cp.zeros((reference_image.ndim, ) + reference_image.shape, dtype=dtype)
proj = cp.zeros(
(reference_image.ndim, reference_image.ndim) + reference_image.shape,
dtype=dtype,
)
s_g = [slice(None)] * g.ndim
s_p = [slice(None)] * proj.ndim
s_d = [slice(None)] * (proj.ndim - 2)
for _ in range(num_warp):
if prefilter:
flow_current = ndi.median_filter(flow_current,
[1] + reference_image.ndim * [3])
image1_warp = warp(moving_image, grid + flow_current, mode="nearest")
grad = cp.stack(cnp.gradient(image1_warp))
NI = (grad * grad).sum(0)
NI[NI == 0] = 1
rho_0 = image1_warp - reference_image - (grad * flow_current).sum(0)
for _ in range(num_iter):
# Data term
rho = rho_0 + (grad * flow_current).sum(0)
idx = abs(rho) <= f0 * NI
flow_auxiliary = flow_current
flow_auxiliary[:, idx] -= rho[idx] * grad[:, idx] / NI[idx]
idx = ~idx
srho = f0 * cp.sign(rho[idx])
flow_auxiliary[:, idx] -= srho * grad[:, idx]
# Regularization term
flow_current = flow_auxiliary.copy()
for idx in range(reference_image.ndim):
s_p[0] = idx
for _ in range(reg_num_iter):
for ax in range(reference_image.ndim):
s_g[0] = ax
s_g[ax + 1] = slice(0, -1)
g[tuple(s_g)] = cp.diff(flow_current[idx], axis=ax)
s_g[ax + 1] = slice(None)
norm = cp.sqrt((g * g).sum(0, keepdims=True))
norm *= f1
norm += 1.0
proj[idx] -= dt * g
proj[idx] /= norm
# d will be the (negative) divergence of proj[idx]
d = -proj[idx].sum(0)
for ax in range(reference_image.ndim):
s_p[1] = ax
s_p[ax + 2] = slice(0, -1)
s_d[ax] = slice(1, None)
d[tuple(s_d)] += proj[tuple(s_p)]
s_p[ax + 2] = slice(None)
s_d[ax] = slice(None)
flow_current[idx] = flow_auxiliary[idx] + d
flow_previous -= flow_current # The difference as stopping criteria
if (flow_previous * flow_previous).sum() < tol:
break
flow_previous = flow_current
return flow_current
def compute(self, ab, krv, cartesian=True, bohren=True):
'''Returns the field scattered by the particle at each coordinate
Parameters
----------
ab : numpy.ndarray
[2, norders] Mie scattering coefficients
krv : numpy.ndarray
Reduced vector displacements of particle from image coordinates
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.
Returns
-------
field : numpy.ndarray
[3, npts] array of complex vector values of the
scattered field at each coordinate.
'''
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.
shape = krv.shape
kx = krv[0, :]
ky = krv[1, :]
kz = -krv[2, :]
# 2. geometric factors
krho = cp.sqrt(kx**2 + ky**2)
kr = cp.sqrt(krho**2 + kz**2)
self.cosphi[...] = safe_division(kx, krho, 1.)
self.sinphi[...] = safe_division(ky, krho, 0.)
self.costheta[...] = safe_division(kz, kr, 1.) # z convention
self.sintheta[...] = safe_division(krho, kr, 0.)
sinkr = cp.sin(kr)
coskr = cp.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 bohren:
factor = 1.j * cp.sign(kz)
else:
factor = -1.j * cp.sign(kz)
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]
self.mo1n[0, :] = 0.j # no radial component
# storage for scattered field
self.es.fill(0.j)
# COMPUTE field by summing partial waves
for n in range(1, norders):
# upward recurrences ...
# 4. Legendre factor (4.47)
# Method described by Wiscombe (1980)
swisc = pi_n * self.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)
self.mo1n[1, :] = pi_n * xi_n # ... divided by cosphi/kr
self.mo1n[2, :] = tau_n * xi_n # ... divided by sinphi/kr
# ... divided by cosphi sintheta/kr^2
self.ne1n[0, :] = n * (n + 1.) * pi_n * xi_n
self.ne1n[1, :] = tau_n * dn # ... divided by cosphi/kr
self.ne1n[2, :] = pi_n * dn # ... divided by sinphi/kr
# prefactor, page 93
en = 1.j**n * (2. * n + 1.) / n / (n + 1.)
# the scattered field in spherical coordinates (4.45)
self.es += (1.j * en * ab[n, 0]) * self.ne1n
self.es -= (en * ab[n, 1]) * self.mo1n
# 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
self.es[0, :] *= self.cosphi * self.sintheta * radialfactor**2
self.es[1, :] *= self.cosphi * radialfactor
self.es[2, :] *= self.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:
self.ec[0, :] = self.es[0, :] * self.sintheta * self.cosphi
self.ec[0, :] += self.es[1, :] * self.costheta * self.cosphi
self.ec[0, :] -= self.es[2, :] * self.sinphi
self.ec[1, :] = self.es[0, :] * self.sintheta * self.sinphi
self.ec[1, :] += self.es[1, :] * self.costheta * self.sinphi
self.ec[1, :] += self.es[2, :] * self.cosphi
self.ec[2, :] = (self.es[0, :] * self.costheta -
self.es[1, :] * self.sintheta)
return self.ec
else:
return self.es
def soft_py(x, tau):
threshed = np.maximum(np.abs(x) - tau, 0)
threshed = threshed * np.sign(x)
return threshed
def _min_limit(x, val=eps):
mask = cp.abs(x) < eps
x[mask] = cp.sign(x[mask]) * eps
def T(x,d,dim=2):
assert dim <= d
assert dim >= 1
assert dim == int(dim)
return x + 2*cp.sign(x)*cp.array(dim*[1]+(d-dim)*[0])
def soft_thresh(self, x, tau):
out = np.maximum(np.abs(x) - tau, 0)
out = out * np.sign(x)
return out
def fixpix(data,mask,out=None,dtype=None,fix_NaN=False):
'''
fill the bad pixel with linear interpolation using surrounding pixels
Parameters
----------
data : array-like
An array of image
If a 3D array containing multiple images,
the images must be stacked along the 1st dimension (axis=0).
mask : array-like
An array indicates bad pixel positions
The shape must be same as image.
The value of bad pixel is nonzero, and the others is 0.
If all pixels are bad, raise ValueError.
out : cupy.ndarray, default None
Alternate output array in which to place the result. The default
is None; if provided, it must have the same shape as the
expected output, but the type will be cast if necessary.
dtype : str or dtype, default None
dtype of ndarray used internally
If None, use eclair.common.default_dtype.
If the input dtype is different, use a casted copy.
fix_NaN : bool, default False
If true, fix NaN pixel even if it's not specified
as bad pixel in mask
Returns
-------
fixed : ndarray
An array of images fixed bad pixel
Notes
-----
NaN is ignored in interpolation calculations,
but is not fixed if fix_NaN is False.
'''
dtype = judge_dtype(dtype)
data = cp.asarray(data,dtype=dtype)
mask = cp.asarray(mask,dtype=dtype)
dshape = data.shape
imshape = data.shape[-2:]
if imshape != mask.shape[-2:]:
raise ValueError('shape differs between data and mask')
if mask.all(axis=(-2,-1)).any():
raise ValueError('No available pixel')
if out is None:
out = data.copy()
else:
cp.copyto(out,data)
tout = cp.array(out,copy=False,ndmin=3)
filt = cp.array(mask,ndmin=3)
elementwise_not(filt,filt)
ternary_operation(filt,tout,0,tout)
dconv = cp.empty_like(tout)
nconv = cp.empty_like(filt)
for _ in range(max(imshape)):
conv_kernel(tout,dconv)
conv_kernel(filt,nconv)
fix_core(filt,dconv,nconv,fix_NaN,tout)
if nconv.all():
break
else:
cp.sign(nconv,out=filt)
return out
def sign(self):
self.getNdArray()[:] = cp.sign(self.getNdArray())
return self
variation = proc.calc_variations('EQUALIZED_SUM')
np.save(video_folder / 'variation', variation)
variation = cp.asarray(variation)
hz = 1 / proc.fps
yfft_eq_sum = fft.fft(variation)
xfft_eq_sum = fft.fftfreq(variation.size, hz)[:variation.size // 2]
sum_energy_eq_sum = cp.cumsum(
2.0 / variation.size * cp.abs(yfft_eq_sum[0:variation.size // 2]))
eq_sum_limiar_pc_energy = sum_energy_eq_sum.max() * ENERGY_LIMIT
frequency = float(
xfft_eq_sum[sum_energy_eq_sum <= eq_sum_limiar_pc_energy].max())
filter_coef = firwin(NUM_TAPS, frequency, fs=2 / hz)
result = lfilter(filter_coef, 1, cp.asnumpy(variation))
np.save(video_folder / 'filtered', result)
result = cp.asarray(result)
subs = cp.diff(result)
subs2 = cp.diff(subs)
critical_points = (cp.sign(subs[1:]) != cp.sign(
subs[:-1])).astype(bool)
maxes = critical_points & (subs2 < 0).astype(bool)
mins = critical_points & (subs2 > 0).astype(bool)
cp.save(video_folder / 'maxes', maxes)
cp.save(video_folder / 'mins', mins)
cp.save(video_folder / 'diff', subs)
cp.save(video_folder / 'diff2', subs2)
with (video_folder / 'metadata.json').open('w') as fp:
dump({
'hz': hz,
'frequency': frequency,
}, fp)
def General_n_Balance_n_Collision_Eff(self,
_new_path,
length_only = True,
GPU_accelerating = False,
GPU_accelerating_data = None,
matrix_data = None):
ITC = {}
max_ITC = 1
min_ITC = sys.maxsize
total_cost = 0
max_order = 0
total_order = 0
standard_index = self.tools.GetWidth()**2 + self.tools.GetHeight()**2
# Parallelization
if GPU_accelerating and length_only:
n_AGV, population_size = GPU_accelerating_data
T_matrix, S_matrix = matrix_data
T_matrix = cp.array(T_matrix)
S_matrix = cp.array(np.array(S_matrix).astype(float))
ITC_matrix = cp.reshape(cp.dot(T_matrix, cp.array([[1],[1]])), (population_size, n_AGV))
O_matrix = cp.reshape(cp.dot(T_matrix, cp.array([[0],[1]])), (population_size, n_AGV))
TC_matrix = cp.reshape(cp.dot(ITC_matrix, cp.ones((n_AGV, 1))), (population_size))
TO_matrix = cp.reshape(cp.dot(O_matrix, cp.ones((n_AGV, 1))), (population_size))
max_ITC_matrix = cp.amax(ITC_matrix, axis=1)
min_ITC_matrix = cp.amin(ITC_matrix, axis=1)
max_order_matrix = cp.amax(O_matrix, axis=1)
_, n_order_points, _ = S_matrix.shape
t_m = cp.reshape(cp.dot(S_matrix, cp.array([[[1],[0],[0],[0],[0]]]*n_order_points)),
(population_size, n_order_points, n_order_points))
x_m = cp.reshape(cp.dot(S_matrix, cp.array([[[0],[1],[0],[0],[0]]]*n_order_points)),
(population_size, n_order_points, n_order_points))
y_m = cp.reshape(cp.dot(S_matrix, cp.array([[[0],[0],[1],[0],[0]]]*n_order_points)),
(population_size, n_order_points, n_order_points))
l_m = cp.reshape(cp.dot(S_matrix, cp.array([[[0],[0],[0],[1],[0]]]*n_order_points)),
(population_size, n_order_points, n_order_points))
o_m = cp.reshape(cp.dot(S_matrix, cp.array([[[0],[0],[0],[0],[1]]]*n_order_points)),
(population_size, n_order_points, n_order_points))
t_m_l = cp.reshape(cp.dot(S_matrix, cp.array([[[1],[0],[0],[0],[0]]])),
(population_size, n_order_points))
t_m_diff = t_m - cp.transpose(t_m, (0, 2, 1))
x_m_diff = x_m - cp.transpose(x_m, (0, 2, 1))
y_m_diff = y_m - cp.transpose(y_m, (0, 2, 1))
m_xy_diff = cp.absolute(x_m_diff) + cp.absolute(y_m_diff)
m_diff = cp.absolute(t_m_diff) + m_xy_diff
m_diff_l = m_diff - l_m * 2
m_diff_l_sign = (cp.logical_xor(cp.sign(m_diff_l) + 1, True))
m_diff_l_eff = cp.multiply(m_diff, m_diff_l_sign)
m_diff_l_sign = cp.sign(m_diff_l_eff)
m_diff_l_H = cp.multiply(cp.multiply(cp.reciprocal(m_diff_l_eff + m_diff_l_sign - 1), m_diff_l_sign),
cp.log10(m_diff_l_eff + cp.absolute(m_diff_l_sign - 1)))
d_m = cp.reciprocal(cp.sum(m_diff_l_H,
(1,2)))
# Occupancy test
"""
t_m_o = t_m + o_m - 1
m_diff_o = cp.absolute(t_m_o - cp.transpose(t_m_o, (0, 2, 1))) - o_m - 1
m_occupancy = (cp.logical_xor(cp.sign(m_diff_o) + 1, True))
m_idn = cp.identity(n_order_points)
OT = cp.prod(cp.logical_or(m_xy_diff,
cp.logical_not(m_occupancy - m_idn)),
(1,2))
"""
G1 = max_order_matrix/max_ITC_matrix
G2 = TO_matrix/TC_matrix
BU = min_ITC_matrix/max_ITC_matrix
CI = cp.multiply(d_m, BU) # d_m * 0.1
E_matrix = G1 + G2 + BU + CI
cp.cuda.Stream.null.synchronize()
return (list(E_matrix), (list(max_ITC_matrix), list(TC_matrix), list(BU), list(CI)))
# Non-Paralleization
else:
print("[Scheduling] Must be use GPU to calculate")
for each_AGV_ID in _new_path.keys():
each_AGV_len_schedule = 0
each_AGV_num_orders = 0
if length_only:
each_AGV_len_schedule, each_num_order, each_order_list = _new_path[each_AGV_ID]
each_AGV_num_orders = each_num_order
else:
each_path = _new_path[each_AGV_ID]
for each_pos_path in each_path:
if len(each_pos_path) == 3:
each_AGV_num_orders += 1
each_AGV_len_schedule += 1
cost = each_AGV_len_schedule + each_AGV_num_orders
ITC[each_AGV_ID] = cost
if each_AGV_num_orders > max_order:
max_order = each_AGV_num_orders
total_order += each_AGV_num_orders
for _, each_value in ITC.items():
if each_value > max_ITC:
max_ITC = each_value
if each_value < min_ITC:
min_ITC = each_value
total_cost += each_value
TT = max_ITC
TTC = total_cost
BU = min_ITC / max_ITC
CI = 0
G1 = max_order/TT
G2 = total_order/TTC
value = G1 + G2 + BU + CI
return (value, (TT, TTC, BU, CI))
def compute_gradients_gpu(u, v, lr, batch_size, num_negs, dimensions, indices,
negatives):
"""
This function computes and applies the gradient updates for the batch os samples, applying them inside
the function then returning the final values that the embedding should be set too.
:param u: the embedding values for the words we are computing gradients for (batch_size, dimensions)
:param v: the embedding values for the positive sample + negatives samples (1 + num_negs, dimensions)
:param lr: learning rate
:param batch_size: size of the current batch, may vary from regular batch_size because of last remaining samples
that can't fit into an entire batch_size in an epoch
:param num_negs: number of negative samples for each positive sample
:param dimensions: dimensions of the embedding
:param indices: the indices of all the u values (batch_size)
:param negatives: the indices of all the v values (batch_size, 1 + num_negs)
:return: Returns the values that the embedding of the samples and negatives should be set too as well as the loss
Once again, calculation was adapted from R. Řehůřek and P. Sojka
RaRe-Technologies: Software Framework for Topic Modelling with Large Corpora
https://github.com/RaRe-Technologies/gensim
"""
u = cp.asarray(u)
v = cp.asarray(v)
u_orig = u
v_orig = v
u = u.T[cp.newaxis, :, :]
v = cp.swapaxes(cp.swapaxes(v, 1, 2), 0, 2)
norm_u = cp.linalg.norm(u, axis=1)
norm_v = cp.linalg.norm(v, axis=1)
distances = cp.linalg.norm(u - v, axis=1)
alpha = 1 - norm_u**2
beta = 1 - norm_v**2
gamma = 1 + 2 * ((distances**2) / (alpha * beta))
poincare_distances = cp.arccosh(gamma)
exp_negative_distances = cp.exp(-poincare_distances)
sum = cp.sum(exp_negative_distances, axis=0)
distances_squared = distances**2
c = (4 / (alpha * beta * cp.sqrt(gamma**2 - 1)))[:, cp.newaxis, :]
u_coeff = ((distances_squared + alpha) / alpha)[:, cp.newaxis, :]
distance_gradient_u = (u_coeff * u - v)
distance_gradient_u *= c
v_coeffs = ((distances_squared + beta) / beta)[:, cp.newaxis, :]
distance_gradients_v = (v_coeffs * v - u)
distance_gradients_v *= c
gradients_v = -exp_negative_distances[:,
cp.newaxis, :] * distance_gradients_v
gradients_v /= sum
gradients_v[0] += distance_gradients_v[0]
gradients_v[0] += lr * 2 * v[0]
gradients_u = -exp_negative_distances[:,
cp.newaxis, :] * distance_gradient_u
gradients_u /= sum
gradient_u = cp.sum(gradients_u, axis=0)
gradient_u += distance_gradient_u[0]
u_update = (lr * (alpha**2) / 4 * gradient_u).T
handle_duplicates(u_update, indices)
v_updates = cp.swapaxes(
cp.swapaxes((lr * (beta**2)[:, cp.newaxis, :] / 4 * gradients_v), 0,
2), 1, 2)
v_updates = cp.reshape(v_updates,
(batch_size * (num_negs + 1), dimensions))
handle_duplicates(v_updates, np.ravel(negatives))
u_orig -= u_update
v_orig = cp.reshape(v_orig, (batch_size * (num_negs + 1), dimensions))
v_orig -= v_updates
u_norms = cp.linalg.norm(u_orig, axis=1)
v_norms = cp.linalg.norm(v_orig, axis=1)
u_orig = (u_norms >= 1 - epsilon)[:, cp.newaxis] * (
u_orig / u_norms[:, cp.newaxis] - cp.sign(u_orig) * 0.00001) + (
u_norms < 1 - epsilon)[:, cp.newaxis] * u_orig
v_orig = (v_norms >= 1 - epsilon)[:, cp.newaxis] * (
v_orig / v_norms[:, cp.newaxis] - cp.sign(v_orig) * 0.00001) + (
v_norms < 1 - epsilon)[:, cp.newaxis] * v_orig
loss = cp.sum(-cp.log(exp_negative_distances[0] / sum), axis=0)
return u_orig, v_orig, loss
