def _slogdet_one(a):
util._assert_rank2(a)
util._assert_nd_squareness(a)
dtype = a.dtype
handle = device.get_cusolver_handle()
m = len(a)
ipiv = cupy.empty(m, 'i')
info = cupy.empty((), 'i')
# Need to make a copy because getrf works inplace
a_copy = a.copy(order='F')
if dtype == 'f':
getrf_bufferSize = cusolver.sgetrf_bufferSize
getrf = cusolver.sgetrf
else:
getrf_bufferSize = cusolver.dgetrf_bufferSize
getrf = cusolver.dgetrf
buffersize = getrf_bufferSize(handle, m, m, a_copy.data.ptr, m)
workspace = cupy.empty(buffersize, dtype=dtype)
getrf(handle, m, m, a_copy.data.ptr, m, workspace.data.ptr,
ipiv.data.ptr, info.data.ptr)
if info[()] == 0:
diag = cupy.diag(a_copy)
# ipiv is 1-origin
non_zero = (cupy.count_nonzero(ipiv != cupy.arange(1, m + 1)) +
cupy.count_nonzero(diag < 0))
# Note: sign == -1 ** (non_zero % 2)
sign = (non_zero % 2) * -2 + 1
logdet = cupy.log(abs(diag)).sum()
else:
sign = cupy.array(0.0, dtype=dtype)
logdet = cupy.array(float('-inf'), dtype)
return sign, logdet
def liquidize(self, intens, sigma_A, gamma_A):
'''Apply liquidization transform on given intensity'''
s_sq = (2. * cp.pi * sigma_A * self.dgen.qrad)**2
patt = cp.fft.fftshift(cp.fft.fftn(cp.fft.ifftshift(intens)))
if self.slimits.max() > 2. * np.pi * sigma_A / self.res_max:
n_max = np.where(
self.slimits > 2. * np.pi * sigma_A / self.res_max)[0][0] + 1
else:
print('No effect of liquid-like motions with these parameters')
return intens
liq = cp.zeros_like(intens)
for n in range(n_max):
kernel = cp.exp(-n * self.urad / gamma_A)
weight = cp.exp(-s_sq + n * cp.log(s_sq) -
float(special.loggamma(n + 1)))
liq += weight * cp.abs(cp.fft.fftshift(cp.fft.ifftn(
patt * kernel)))
sys.stderr.write('\rLiquidizing: %d/%d' % (n + 1, n_max))
sys.stderr.write('\n')
return liq
def forward(self, x, t):
if x.ndim == 2: # ミニバッチ使用時
x = x - x.max(axis=1, keepdims=True)
x = cp.exp(x)
y = x / x.sum(axis=1, keepdims=True)
elif x.ndim == 1:
x = x - cp.max(x)
y = cp.exp(x) / cp.sum(cp.exp(x))
if y.ndim == 1:
t = t.reshape(1, t.size)
y = y.reshape(1, y.size)
# 教師ラベルがone-hotベクトルの場合、正解のインデックスに変換
if t.size == y.size:
t = t.argmax(axis=1)
batch_size = y.shape[0]
loss = -1.0 * cp.sum(
t * cp.log(y[cp.arange(batch_size), t] + 1e-7)) / batch_size
self.y = y
self.t = t
return loss
def rayleigh(self, scale=1.0, size=None, dtype=float):
"""Returns an array of samples drawn from a rayleigh distribution.
.. warning::
This function may synchronize the device.
.. seealso::
:func:`cupy.random.rayleigh` for full documentation,
:meth:`numpy.random.RandomState.rayleigh
<numpy.random.mtrand.RandomState.rayleigh>`
"""
scale = cupy.asarray(scale)
if size is None:
size = scale.shape
if cupy.any(scale < 0): # synchronize!
raise ValueError('scale < 0')
x = self._random_sample_raw(size, dtype)
x = cupy.log(x, out=x)
x = cupy.multiply(x, -2., out=x)
x = cupy.sqrt(x, out=x)
x = cupy.multiply(x, scale, out=x)
return x
def o_f2(imgs, hres_size, row=None, do_fil=False, show=False):
"""
Fourier Spectrum Initialization #2
----------------------------------
Mean Image > pad with reflect padding > sqrt > Ft > (optional) filter Ft with 2 * cutoff_freq > return Ft
"""
im = cp.array(imgs).mean(0)
pad = (hres_size[0] - imgs[0].shape[0]) // 2
im = cp.pad(cp.array(im), [(pad, pad), (pad, pad)], mode='reflect')
f = Ft(cp.sqrt(im))
if do_fil:
_orig = hres_size[0] // 2 - 1
CUTOFF_FREQ_px = get_cutoff(row)
fil = np.zeros(hres_size)
fil = cp.array(
cv2.circle(fil, (_orig, _orig), 2 * CUTOFF_FREQ_px, 1, -1))
f = f * fil
if show:
plt.imshow(cp.asnumpy(cp.log(abs(f) + 1e-7)))
plt.title(f'o_f2 {"without" if not do_fil else "with"} filtering')
plt.show()
return f
def __extended_likelihood(self, params):
data = self.__data_amplitude.calculate(params)
mcdata = self.__monte_carlo_amplitude.calculate(params)
if self.__data_amplitude.USE_GPU:
likelihood_data = cp.sum(self.sweight * cp.log(data))
likelihood_mc = cp.sum(self.mcweight * mcdata)
return cp.asnumpy(
likelihood_data - self.__generated * likelihood_mc
)
else:
likelihood_data = ne.evaluate(
"sum(sw * log(data))", local_dict={
"sw": self.sweight,
"data": data
}
)
likelihood_mc = ne.evaluate(
"sum(mcw * mcdata)", local_dict={
"mcw": self.mcweight,
"mcdata": mcdata
}
)
return likelihood_data - self.__generated * likelihood_mc
def _acausal_classifier_gpu(filter_posterior, movement_state_transition,
discrete_state_transition, observed_position_bin,
uniform):
'''
Parameters
----------
filter_posterior : ndarray, shape (n_time, 2, n_position_bins)
movement_state_transition : ndarray, shape (n_position_bins,
n_position_bins)
discrete_state_transition : ndarray, shape (n_time, 2)
discrete_state_transition[k, 0] = Pr(I_{k} = 1 | I_{k-1} = 0, v_{k})
discrete_state_transition[k, 1] = Pr(I_{k} = 1 | I_{k-1} = 1, v_{k})
observed_position_bin : ndarray, shape (n_time,)
Which position bin is the animal in.
position_bin_size : float
Returns
-------
smoother_posterior : ndarray, shape (n_time, 2, n_position_bins)
p(x_{k + 1}, I_{k + 1} \vert H_{1:T})
smoother_probability : ndarray, shape (n_time, 2)
smoother_probability[:, 0] = Pr(I_{1:T} = 0)
smoother_probability[:, 1] = Pr(I_{1:T} = 1)
smoother_prior : ndarray, shape (n_time, 2, n_position_bins)
p(x_{k + 1}, I_{k + 1} \vert H_{1:k})
weights : ndarray, shape (n_time, 2, n_position_bins)
\sum_{I_{k+1}} \int \Big[ \frac{p(x_{k+1} \mid x_{k}, I_{k}, I_{k+1}) *
Pr(I_{k + 1} \mid I_{k}, v_{k}) * p(x_{k+1}, I_{k+1} \mid H_{1:T})}
{p(x_{k + 1}, I_{k + 1} \mid H_{1:k})} \Big] dx_{k+1}
''' # noqa
filter_posterior = cp.asarray(filter_posterior, dtype=cp.float32)
movement_state_transition = cp.asarray(
movement_state_transition, dtype=cp.float32)
discrete_state_transition = cp.asarray(
discrete_state_transition, dtype=cp.float32)
observed_position_bin = cp.asarray(observed_position_bin)
uniform = cp.asarray(uniform, dtype=cp.float32)
EPS = cp.asarray(np.spacing(1), dtype=cp.float32)
filter_probability = cp.sum(filter_posterior, axis=2)
smoother_posterior = cp.zeros_like(filter_posterior)
n_time, _, n_position_bins = filter_posterior.shape
smoother_posterior[-1] = filter_posterior[-1].copy()
for k in cp.arange(n_time - 2, -1, -1):
smoother_prior = cp.zeros((2, n_position_bins), dtype=cp.float32)
weights = cp.zeros((2, n_position_bins), dtype=cp.float32)
position_ind = observed_position_bin[k + 1]
# Predict p(x_{k + 1}, I_{k + 1} \vert H_{1:k})
# I_{k} = 0, I_{k + 1} = 0
smoother_prior[0, position_ind] = (
(1 - discrete_state_transition[k + 1, 0]) * filter_probability[k, 0])
# I_{k} = 1, I_{k + 1} = 0
smoother_prior[0, position_ind] += (
(1 - discrete_state_transition[k + 1, 1]) * filter_probability[k, 1])
# I_{k} = 0, I_{k + 1} = 1
smoother_prior[1] = (
discrete_state_transition[k + 1, 0] * uniform *
filter_probability[k, 0])
# I_{k} = 1, I_{k + 1} = 1
smoother_prior[1] += (
discrete_state_transition[k + 1, 1] *
(movement_state_transition.T @ filter_posterior[k, 1]))
# Update p(x_{k}, I_{k} \vert H_{1:k})
ratio = cp.exp(
cp.log(smoother_posterior[k + 1]) -
cp.log(smoother_prior + EPS))
integrated_ratio = cp.sum(ratio, axis=1)
# I_{k} = 0, I_{k + 1} = 0
weights[0] = (
(1 - discrete_state_transition[k + 1, 0]) * ratio[0, position_ind])
# I_{k} = 0, I_{k + 1} = 1
weights[0] += (
uniform * discrete_state_transition[k + 1, 0] * integrated_ratio[1])
# I_{k} = 1, I_{k + 1} = 0
weights[1] = (
(1 - discrete_state_transition[k + 1, 1]) * ratio[0, position_ind])
# I_{k} = 1, I_{k + 1} = 1
weights[1] += (
discrete_state_transition[k + 1, 1] *
ratio[1] @ movement_state_transition)
smoother_posterior[k] = weights * filter_posterior[k]
smoother_posterior[k] /= cp.nansum(smoother_posterior[k])
smoother_probability = cp.sum(smoother_posterior, axis=2)
return (cp.asnumpy(smoother_posterior),
cp.asnumpy(smoother_probability))
def _causal_classifier_gpu(likelihood, movement_state_transition, discrete_state_transition,
observed_position_bin, uniform):
'''
Parameters
----------
likelihood : ndarray, shape (n_time, ...)
movement_state_transition : ndarray, shape (n_position_bins,
n_position_bins)
discrete_state_transition : ndarray, shape (n_time, 2)
discrete_state_transition[k, 0] = Pr(I_{k} = 1 | I_{k-1} = 0, v_{k})
discrete_state_transition[k, 1] = Pr(I_{k} = 1 | I_{k-1} = 1, v_{k})
observed_position_bin : ndarray, shape (n_time,)
Which position bin is the animal in.
position_bin_size : float
Returns
-------
posterior : ndarray, shape (n_time, 2, n_position_bins)
state_probability : ndarray, shape (n_time, 2)
state_probability[:, 0] = Pr(I_{1:T} = 0)
state_probability[:, 1] = Pr(I_{1:T} = 1)
prior : ndarray, shape (n_time, 2, n_position_bins)
'''
likelihood = cp.asarray(likelihood, dtype=cp.float32)
movement_state_transition = cp.asarray(
movement_state_transition, dtype=cp.float32)
discrete_state_transition = cp.asarray(
discrete_state_transition, dtype=cp.float32)
observed_position_bin = cp.asarray(observed_position_bin)
uniform = cp.asarray(uniform, dtype=cp.float32)
n_position_bins = movement_state_transition.shape[0]
n_time = likelihood.shape[0]
n_states = 2
posterior = cp.zeros(
(n_time, n_states, n_position_bins), dtype=cp.float32)
state_probability = cp.zeros((n_time, n_states), dtype=cp.float32)
# Initial Conditions
posterior[0, 0, observed_position_bin[0]] = likelihood[0, 0, 0]
norm = cp.nansum(posterior[0])
data_log_likelihood = cp.log(norm)
posterior[0] /= norm
state_probability[0] = cp.sum(posterior[0], axis=1)
for k in np.arange(1, n_time):
prior = cp.zeros((n_states, n_position_bins), dtype=cp.float32)
position_ind = observed_position_bin[k]
# I_{k - 1} = 0, I_{k} = 0
prior[0, position_ind] = (
(1 - discrete_state_transition[k, 0]) * state_probability[k - 1, 0])
# I_{k - 1} = 1, I_{k} = 0
prior[0, position_ind] += (
(1 - discrete_state_transition[k, 1]) * state_probability[k - 1, 1])
# I_{k - 1} = 0, I_{k} = 1
prior[1] = (discrete_state_transition[k, 0] * uniform *
state_probability[k - 1, 0])
# I_{k - 1} = 1, I_{k} = 1
prior[1] += (
discrete_state_transition[k, 1] *
(movement_state_transition.T @ posterior[k - 1, 1]))
posterior[k] = prior * likelihood[k]
norm = cp.nansum(posterior[k])
data_log_likelihood += cp.log(norm)
posterior[k] /= norm
state_probability[k] = cp.sum(posterior[k], axis=1)
return (cp.asnumpy(posterior),
cp.asnumpy(state_probability),
data_log_likelihood)
def convolutional_barycenter_gpu(Hv,
reg,
alpha,
stabThresh=1e-30,
niter=1500,
tol=1e-9,
sharpening=False,
verbose=False):
"""Main function solving wasserstein barycenter problem using gpu
Arguments:
Hv {Set of distributions (cparray)} --
reg {regularization term "gamma"} -- float superior to 0, generally equals size of space/40
alpha {list} -- set of weights
Keyword Arguments:
stabThresh {float} -- Stabilization threshold to prevent division by 0 (default: {1e-30})
niter {int} -- Maximum number of loop iteration (default: {1500})
tol {float} -- convergence tolerance at which point iterations stop (default: {1e-9})
sharpening {bool} -- Whether or not entropic sharpening is used (default: {False})
verbose {bool} -- verbose option
Returns:
cparray -- solution of weighted wassertein barycenter problem
"""
def K(x):
return cp.array(gaussian_filter(cp.asnumpy(x), sigma=reg))
def to_find_root(barycenter, H0, beta):
return entropy(barycenter**beta) - H0
alpha = cp.array(alpha)
alpha = alpha / alpha.sum()
Hv = cp.array(Hv)
mean_weights = (Hv[0].sum() + Hv[1].sum()) / 2.
#print('mean weights', mean_weights)
for i in range(len(Hv)):
Hv[i] = Hv[i] / Hv[i].sum()
v = cp.ones(Hv.shape)
Kw = cp.ones(Hv.shape)
entropy_max = max_entropy(Hv)
barycenter = cp.zeros(Hv[0].shape)
change = 1
for j in range(niter):
t0 = time.time()
barycenterOld = barycenter
barycenter = cp.zeros_like(Hv[0, :, :])
for i in range(Hv.shape[0]):
Kw[i, :, :] = K(Hv[i, :, :] /
cp.maximum(stabThresh, K(v[i, :, :])))
barycenter += alpha[i] * cp.log(
cp.maximum(stabThresh, v[i, :, :] * Kw[i, :, :]))
barycenter = cp.exp(barycenter)
change = cp.sum(cp.abs(barycenter - barycenterOld))
if sharpening:
if (entropy(barycenter)) > (entropy_max):
beta = newton(
lambda beta: to_find_root(barycenter, entropy_max, beta),
1,
tol=1e-6)
if beta < 0:
beta = 1
else:
beta = 1
barycenter = barycenter**beta
for i in range(Hv.shape[0]):
v[i, :, :] = barycenter / cp.maximum(stabThresh, Kw[i, :, :])
if verbose:
#sys.stdout('output.log','a')
print("iter : ", j, "change : ", change, 'time :',
time.time() - t0)
if change < tol:
break
return cp.asnumpy(barycenter)
def choice(self, a, size=None, replace=True, p=None):
"""Returns an array of random values from a given 1-D array.
.. seealso::
:func:`cupy.random.choice` for full document,
:meth:`numpy.random.choice`
"""
if a is None:
raise ValueError('a must be 1-dimensional or an integer')
if isinstance(a, cupy.ndarray) and a.ndim == 0:
raise NotImplementedError
if isinstance(a, six.integer_types):
a_size = a
if a_size <= 0:
raise ValueError('a must be greater than 0')
else:
a = cupy.array(a, copy=False)
if a.ndim != 1:
raise ValueError('a must be 1-dimensional or an integer')
else:
a_size = len(a)
if a_size == 0:
raise ValueError('a must be non-empty')
if p is not None:
p = cupy.array(p)
if p.ndim != 1:
raise ValueError('p must be 1-dimensional')
if len(p) != a_size:
raise ValueError('a and p must have same size')
if not (p >= 0).all():
raise ValueError('probabilities are not non-negative')
p_sum = cupy.sum(p).get()
if not numpy.allclose(p_sum, 1):
raise ValueError('probabilities do not sum to 1')
if not replace:
raise NotImplementedError
if size is None:
raise NotImplementedError
shape = size
size = numpy.prod(shape)
if p is not None:
p = cupy.broadcast_to(p, (size, a_size))
index = cupy.argmax(cupy.log(p) -
cupy.random.gumbel(size=(size, a_size)),
axis=1)
if not isinstance(shape, six.integer_types):
index = cupy.reshape(index, shape)
else:
index = cupy.random.randint(0, a_size, size=shape)
# Align the dtype with NumPy
index = index.astype(cupy.int64, copy=False)
if isinstance(a, six.integer_types):
return index
if index.ndim == 0:
return cupy.array(a[index], dtype=a.dtype)
return a[index]
def compute_cost(AL, Y):
m = Y.shape[1]
cost = (1./m) * (-np.dot(Y, np.log(AL).T) - np.dot(1-Y, np.log(1 - AL).T))
cost = np.squeeze(cost)
return cost
def mi_model_1d_gpu_gd(x, y, biascorrect=False, demeaned=False):
"""Mutual information between a Gaussian and a discrete variable in bits.
This method is based on ANOVA style model comparison.
I = mi_model_gd(x,y) returns the MI between the (possibly multidimensional)
Gaussian variable x and the discrete variable y.
Parameters
----------
x, y : array_like
Gaussian arrays of shape (n_epochs,) or (n_dimensions, n_epochs). y
must be an array of integers
biascorrect : bool | True
Specifies whether bias correction should be applied to the estimated MI
demeaned : bool | False
Specifies whether the input data already has zero mean (true if it has
been copula-normalized)
Returns
-------
i : float
Information shared by x and y (in bits)
"""
# Converting to cupy array
#x, y = cp.array(x), cp.array(y)
x, y = cp.atleast_2d(x), cp.squeeze(y)
if x.ndim > 2:
raise ValueError("x must be at most 2d")
if y.ndim > 1:
raise ValueError("only univariate discrete variables supported")
if not cp.issubdtype(y.dtype, cp.integer):
raise ValueError("y should be an integer array")
nvarx, ntrl = x.shape
ym = cp.unique(y)
if y.size != ntrl:
raise ValueError("number of trials do not match")
if not demeaned:
x = x - x.mean(axis=1)[:, cp.newaxis]
# class-conditional entropies
ntrl_y = cp.zeros(len(ym))
hcond = cp.zeros(len(ym))
for n_yi, yi in enumerate(ym):
idx = y == yi
xm = x[:, idx]
ntrl_y[n_yi] = xm.shape[1]
xm = xm - xm.mean(axis=1)[:, cp.newaxis]
cm = cp.dot(xm, xm.T) / float(ntrl_y[n_yi] - 1)
chcm = cp.linalg.cholesky(cm)
hcond[n_yi] = cp.sum(cp.log(cp.diagonal(chcm)))
# class weights
w = ntrl_y / float(ntrl)
# unconditional entropy from unconditional Gaussian fit
cx = cp.dot(x, x.T) / float(ntrl - 1)
chc = cp.linalg.cholesky(cx)
hunc = cp.sum(cp.log(cp.diagonal(chc))) # + c*nvarx
ln2 = cp.log(2)
if biascorrect:
vars = cp.arange(1, nvarx + 1)
psiterms = psi((ntrl - vars).astype(cp.float) / 2.) / 2.
dterm = (ln2 - cp.log(float(ntrl - 1))) / 2.
hunc = hunc - nvarx * dterm - psiterms.sum()
dterm = (ln2 - cp.log((ntrl_y - 1).astype(cp.float))) / 2.0
psiterms = cp.zeros(len(ym))
for vi in vars:
idx = ntrl_y - vi
psiterms = psiterms + psi(idx.astype(cp.float) / 2.)
hcond = hcond - nvarx * dterm - (psiterms / 2.)
# MI in bits
i = (hunc - cp.sum(w * hcond)) / ln2
return i
path = '/export/scratch2/kostenko/archive/OwnProjects/al_tests/new/90KV_no_filt/' dark = flex.data.read_raw(path, 'di') flat = flex.data.read_raw(path, 'io') proj = flex.data.read_raw(path, 'scan_') meta = flex.data.read_log(path, 'flexray') #%% Prepro: # Convert to CUPY: proj = cupy.array(proj) flat = cupy.array(flat) dark = cupy.array(dark) # Use CUDA to compute stuff: proj = (proj - dark) / (flat.mean(0) - dark) proj = -cupy.log(proj) proj = flex.data.raw2astra(proj) flex.util.display_slice(proj, title='Sinogram') #%% Recon vol = numpy.zeros([1, 2000, 2000], dtype='float32') flex.project.FDK(proj, vol, meta['geometry']) flex.util.display_slice(vol, bounds=[], title='FDK')
def mlog(self, psi):
res = psi.copy()
res[cp.abs(psi) < 1e-32] = 1e-32
res = cp.log(res)
return res
def _support_choice(dist, rand):
return cp.log(dist) + rand
def __call__(self, input_x, t):
output = self.predictor(input_x)
batch_size, _, grid_h, grid_w = output.shape
self.seen += batch_size
x, y, w, h, conf, prob = F.split_axis(F.reshape(
output, (batch_size, self.predictor.n_boxes,
self.predictor.n_classes + 5, grid_h, grid_w)),
(1, 2, 3, 4, 5),
axis=2)
x = F.sigmoid(x) # xのactivation
y = F.sigmoid(y) # yのactivation
conf = F.sigmoid(conf) # confのactivation
prob = F.transpose(prob, (0, 2, 1, 3, 4))
prob = F.softmax(prob) # probablitiyのacitivation
# 教師データの用意
tw = xp.zeros(
w.shape,
dtype=xp.float32) # wとhが0になるように学習(e^wとe^hは1に近づく -> 担当するbboxの倍率1)
th = xp.zeros(h.shape, dtype=xp.float32)
tx = xp.tile(0.5, x.shape).astype(xp.float32) # 活性化後のxとyが0.5になるように学習()
ty = xp.tile(0.5, y.shape).astype(xp.float32)
if self.seen < self.unstable_seen: # centerの存在しないbbox誤差学習スケールは基本0.1
box_learning_scale = xp.tile(0.1, x.shape).astype(xp.float32)
else:
box_learning_scale = xp.tile(0, x.shape).astype(xp.float32)
tconf = xp.zeros(
conf.shape, dtype=xp.float32
) # confidenceのtruthは基本0、iouがthresh以上のものは学習しない、ただしobjectの存在するgridのbest_boxのみ真のIOUに近づかせる
conf_learning_scale = xp.tile(0.1, conf.shape).astype(xp.float32)
tprob = prob.data.copy() # best_anchor以外は学習させない(自身との二乗和誤差 = 0)
# 全bboxとtruthのiouを計算(batch単位で計算する)
x_shift = Variable(
xp.broadcast_to(xp.arange(grid_w, dtype=xp.float32), x.shape[1:]))
y_shift = Variable(
xp.broadcast_to(
xp.arange(grid_h, dtype=xp.float32).reshape(grid_h, 1),
y.shape[1:]))
w_anchor = Variable(
xp.broadcast_to(
xp.reshape(
xp.array(self.anchors, dtype=xp.float32)[:, 0],
(self.predictor.n_boxes, 1, 1, 1)), w.shape[1:]))
h_anchor = Variable(
xp.broadcast_to(
xp.reshape(
xp.array(self.anchors, dtype=xp.float32)[:, 1],
(self.predictor.n_boxes, 1, 1, 1)), h.shape[1:]))
x_shift.to_gpu(), y_shift.to_gpu(), w_anchor.to_gpu(), h_anchor.to_gpu(
)
best_ious = []
for batch in range(batch_size):
n_truth_boxes = len(t[batch])
box_x = (x[batch] + x_shift) / grid_w
box_y = (y[batch] + y_shift) / grid_h
box_w = F.exp(w[batch]) * w_anchor / grid_w
box_h = F.exp(h[batch]) * h_anchor / grid_h
ious = []
for truth_index in range(n_truth_boxes):
truth_box_x = Variable(
xp.broadcast_to(
xp.array(t[batch][truth_index]["x"], dtype=xp.float32),
box_x.shape))
truth_box_y = Variable(
xp.broadcast_to(
xp.array(t[batch][truth_index]["y"], dtype=xp.float32),
box_y.shape))
truth_box_w = Variable(
xp.broadcast_to(
xp.array(t[batch][truth_index]["w"], dtype=xp.float32),
box_w.shape))
truth_box_h = Variable(
xp.broadcast_to(
xp.array(t[batch][truth_index]["h"], dtype=xp.float32),
box_h.shape))
truth_box_x.to_gpu(), truth_box_y.to_gpu(), truth_box_w.to_gpu(
), truth_box_h.to_gpu()
ious.append(
multi_box_iou(
Box(box_x, box_y, box_w, box_h),
Box(truth_box_x, truth_box_y, truth_box_w,
truth_box_h)).data.get())
ious = xp.array(ious)
best_ious.append(xp.max(ious, axis=0))
best_ious = xp.array(best_ious)
# 一定以上のiouを持つanchorに対しては、confを0に下げないようにする(truthの周りのgridはconfをそのまま維持)。
tconf[best_ious > self.thresh] = conf.data.get()[
best_ious > self.thresh]
conf_learning_scale[best_ious > self.thresh] = 0
# objectの存在するanchor boxのみ、x、y、w、h、conf、probを個別修正
abs_anchors = self.anchors / xp.array([grid_w, grid_h])
for batch in range(batch_size):
for truth_box in t[batch]:
truth_w = int(float(truth_box["x"]) * grid_w)
truth_h = int(float(truth_box["y"]) * grid_h)
truth_n = 0
best_iou = 0.0
for anchor_index, abs_anchor in enumerate(abs_anchors):
iou = box_iou(
Box(0, 0, float(truth_box["w"]),
float(truth_box["h"])),
Box(0, 0, abs_anchor[0], abs_anchor[1]))
if best_iou < iou:
best_iou = iou
truth_n = anchor_index
# objectの存在するanchorについて、centerを0.5ではなく、真の座標に近づかせる。anchorのスケールを1ではなく真のスケールに近づかせる。学習スケールを1にする。
box_learning_scale[batch, truth_n, :, truth_h, truth_w] = 1.0
tx[batch, truth_n, :, truth_h,
truth_w] = float(truth_box["x"]) * grid_w - truth_w
ty[batch, truth_n, :, truth_h,
truth_w] = float(truth_box["y"]) * grid_h - truth_h
tw[batch, truth_n, :, truth_h, truth_w] = xp.log(
float(truth_box["w"]) / abs_anchors[truth_n][0])
th[batch, truth_n, :, truth_h, truth_w] = xp.log(
float(truth_box["h"]) / abs_anchors[truth_n][1])
tprob[batch, :, truth_n, truth_h, truth_w] = 0
tprob[batch,
int(truth_box["label"]), truth_n, truth_h, truth_w] = 1
# IOUの観測
full_truth_box = Box(float(truth_box["x"]),
float(truth_box["y"]),
float(truth_box["w"]),
float(truth_box["h"]))
predicted_box = Box(
(x[batch][truth_n][0][truth_h][truth_w].data.get() +
truth_w) / grid_w,
(y[batch][truth_n][0][truth_h][truth_w].data.get() +
truth_h) / grid_h,
xp.exp(w[batch][truth_n][0][truth_h][truth_w].data.get()) *
abs_anchors[truth_n][0],
xp.exp(h[batch][truth_n][0][truth_h][truth_w].data.get()) *
abs_anchors[truth_n][1])
predicted_iou = box_iou(full_truth_box, predicted_box)
tconf[batch, truth_n, :, truth_h, truth_w] = predicted_iou
conf_learning_scale[batch, truth_n, :, truth_h, truth_w] = 10.0
# debug prints
maps = F.transpose(prob[batch], (2, 3, 1, 0)).data
print(
"best confidences and best conditional probability and predicted class of each grid:"
)
for i in range(grid_h):
for j in range(grid_w):
print("%2d" %
(int(conf[batch, :, :, i, j].data.max() * 100)),
end=" ")
print(" ", end="")
for j in range(grid_w):
print("%2d" % (maps[i][j][int(
maps[i][j].max(axis=1).argmax())].argmax()),
end=" ")
print(" ", end="")
for j in range(grid_w):
print("%2d" % (maps[i][j][int(
maps[i][j].max(axis=1).argmax())].max() * 100),
end=" ")
print()
print(
"best default iou: %.2f predicted iou: %.2f confidence: %.2f class: %s"
% (best_iou, predicted_iou,
conf[batch][truth_n][0][truth_h][truth_w].data,
t[batch][0]["label"]))
print("-------------------------------")
print("seen = %d" % self.seen)
# loss計算
tx, ty, tw, th, tconf, tprob = Variable(tx), Variable(ty), Variable(
tw), Variable(th), Variable(tconf), Variable(tprob)
box_learning_scale, conf_learning_scale = Variable(
box_learning_scale), Variable(conf_learning_scale)
tx.to_gpu(), ty.to_gpu(), tw.to_gpu(), th.to_gpu(), tconf.to_gpu(
), tprob.to_gpu()
box_learning_scale.to_gpu()
conf_learning_scale.to_gpu()
x_loss = F.sum((tx - x)**2 * box_learning_scale) / 2
y_loss = F.sum((ty - y)**2 * box_learning_scale) / 2
w_loss = F.sum((tw - w)**2 * box_learning_scale) / 2
h_loss = F.sum((th - h)**2 * box_learning_scale) / 2
c_loss = F.sum((tconf - conf)**2 * conf_learning_scale) / 2
p_loss = F.sum((tprob - prob)**2) / 2
print(
"x_loss: %f y_loss: %f w_loss: %f h_loss: %f c_loss: %f p_loss: %f"
% (F.sum(x_loss).data, F.sum(y_loss).data, F.sum(w_loss).data,
F.sum(h_loss).data, F.sum(c_loss).data, F.sum(p_loss).data))
loss = x_loss + y_loss + w_loss + h_loss + c_loss + p_loss
return loss
def categorical_cross_entropy(x, y, epsilon=10**(-13)):
x = cp.clip(x, epsilon, 1. - epsilon)
N = x.shape[0]
return -cp.sum(y * cp.log(x + 0.00001)) / N
def log_loss(y_true,
y_pred,
eps=1e-15,
normalize=True,
sample_weight=None) -> float:
""" Log loss, aka logistic loss or cross-entropy loss.
This is the loss function used in (multinomial) logistic regression
and extensions of it such as neural networks, defined as the negative
log-likelihood of a logistic model that returns ``y_pred`` probabilities
for its training data ``y_true``.
The log loss is only defined for two or more labels.
Parameters
----------
y_true : array-like, shape = (n_samples,)
y_pred : array-like of float,
shape = (n_samples, n_classes) or (n_samples,)
eps : float (default=1e-15)
Log loss is undefined for p=0 or p=1, so probabilities are
clipped to max(eps, min(1 - eps, p)).
normalize : bool, optional (default=True)
If true, return the mean loss per sample.
Otherwise, return the sum of the per-sample losses.
sample_weight : array-like of shape (n_samples,), default=None
Sample weights.
Returns
-------
loss : float
Examples
--------
.. code-block:: python
>>> from cuml.metrics import log_loss
>>> import cupy as cp
>>> log_loss(cp.array([1, 0, 0, 1]),
... cp.array([[.1, .9], [.9, .1], [.8, .2], [.35, .65]]))
0.21616...
References
----------
C.M. Bishop (2006). Pattern Recognition and Machine Learning. Springer,
p. 209.
Notes
-----
The logarithm used is the natural logarithm (base-e).
"""
y_true, n_rows, n_cols, ytype = \
input_to_cupy_array(y_true, check_dtype=[np.int32, np.int64,
np.float32, np.float64])
if y_true.dtype.kind == 'f' and np.any(y_true != y_true.astype(int)):
raise ValueError("'y_true' can only have integer values")
if y_true.min() < 0:
raise ValueError("'y_true' cannot have negative values")
y_pred, _, _, _ = \
input_to_cupy_array(y_pred, check_dtype=[np.int32, np.int64,
np.float32, np.float64],
check_rows=n_rows)
y_true_max = y_true.max()
if (y_pred.ndim == 1 and y_true_max > 1) \
or (y_pred.ndim > 1 and y_pred.shape[1] <= y_true_max):
raise ValueError("The shape of y_pred doesn't "
"match the number of classes")
y_true = y_true.astype('int32')
y_pred = cp.clip(y_pred, eps, 1 - eps)
if y_pred.ndim == 1:
y_pred = cp.expand_dims(y_pred, axis=1)
if y_pred.shape[1] == 1:
y_pred = cp.hstack([1 - y_pred, y_pred])
y_pred /= cp.sum(y_pred, axis=1, keepdims=True)
loss = -cp.log(y_pred)[cp.arange(y_pred.shape[0]), y_true]
return _weighted_sum(loss, sample_weight, normalize).item()
def GetStepSize(photon_state, tau_atm):
step = -cp.log(cp.random.rand(len(photon_state)))
return step
def mi_1d_gpu_gg(x, y, biascorrect=True, demeaned=False):
"""Mutual information (MI) between two Gaussian variables in bits.
This is the GPU variant of the m1_1d_gg function, using CuPy
I = mi_gg(x,y) returns the MI between two (possibly multidimensional)
Gaussian variables, x and y, with bias correction.
Parameters
----------
x, y : array_like
Gaussian arrays of shape (n_epochs,) or (n_dimensions, n_epochs)
biascorrect : bool | True
Specifies whether bias correction should be applied to the estimated MI
demeaned : bool | False
Specifies whether the input data already has zero mean (true if it has
been copula-normalized)
Returns
-------
i : float
Information shared by x and y (in bits)
"""
x, y = cp.atleast_2d(x), cp.atleast_2d(y)
if (x.ndim > 2) or (y.ndim > 2):
raise ValueError("x and y must be at most 2d")
nvarx, ntrl = x.shape
nvary = y.shape[0]
nvarxy = nvarx + nvary
if y.shape[1] != ntrl:
raise ValueError("number of trials do not match")
# joint variable
xy = cp.vstack((x, y))
if not demeaned:
xy = xy - xy.mean(axis=1)[:, cp.newaxis]
cxy = cp.dot(xy, xy.T) / float(ntrl - 1)
# submatrices of joint covariance
cx = cxy[:nvarx, :nvarx]
cy = cxy[nvarx:, nvarx:]
chcxy = cp.linalg.cholesky(cxy)
chcx = cp.linalg.cholesky(cx)
chcy = cp.linalg.cholesky(cy)
# entropies in nats
# normalizations cancel for mutual information
hx = cp.sum(cp.log(cp.diagonal(chcx)))
hy = cp.sum(cp.log(cp.diagonal(chcy)))
hxy = cp.sum(cp.log(cp.diagonal(chcxy)))
ln2 = cp.log(2)
if biascorrect:
psiterms = psi(
(ntrl - cp.arange(1, nvarxy + 1)).astype(cp.float) / 2.) / 2.
dterm = (ln2 - cp.log(ntrl - 1.)) / 2.
hx = hx - nvarx * dterm - psiterms[:nvarx].sum()
hy = hy - nvary * dterm - psiterms[:nvary].sum()
hxy = hxy - nvarxy * dterm - psiterms[:nvarxy].sum()
# MI in bits
i = (hx + hy - hxy) / ln2
return i
def cmi_1d_gpu_ggg(x, y, z, biascorrect=True, demeaned=False):
"""Conditional MI between two Gaussian variables conditioned on a third.
I = cmi_ggg(x,y,z) returns the CMI between two (possibly multidimensional)
Gaussian variables, x and y, conditioned on a third, z, with bias
correction.
Parameters
----------
x, y, z : array_like
Gaussians arrays of shape (n_epochs,) or (n_dimensions, n_epochs).
biascorrect : bool | True
Specifies whether bias correction should be applied to the estimated MI
demeaned : bool | False
Specifies whether the input data already has zero mean (true if it has
been copula-normalized)
Returns
-------
i : float
Information shared by x and y conditioned by z (in bits)
"""
x, y, z = cp.atleast_2d(x), cp.atleast_2d(y), cp.atleast_2d(z)
if x.ndim > 2 or y.ndim > 2 or z.ndim > 2:
raise ValueError("x, y and z must be at most 2d")
ntrl = x.shape[1]
nvarx = x.shape[0]
nvary = y.shape[0]
nvarz = z.shape[0]
nvaryz = nvary + nvarz
nvarxy = nvarx + nvary
nvarxz = nvarx + nvarz
nvarxyz = nvarx + nvaryz
if y.shape[1] != ntrl or z.shape[1] != ntrl:
raise ValueError("number of trials do not match")
# joint variable
xyz = cp.vstack((x, y, z))
if not demeaned:
xyz = xyz - xyz.mean(axis=1)[:, cp.newaxis]
cxyz = cp.dot(xyz, xyz.T) / float(ntrl - 1)
# submatrices of joint covariance
cz = cxyz[nvarxy:, nvarxy:]
cyz = cxyz[nvarx:, nvarx:]
cxz = cp.zeros((nvarxz, nvarxz))
cxz[:nvarx, :nvarx] = cxyz[:nvarx, :nvarx]
cxz[:nvarx, nvarx:] = cxyz[:nvarx, nvarxy:]
cxz[nvarx:, :nvarx] = cxyz[nvarxy:, :nvarx]
cxz[nvarx:, nvarx:] = cxyz[nvarxy:, nvarxy:]
chcz = cp.linalg.cholesky(cz)
chcxz = cp.linalg.cholesky(cxz)
chcyz = cp.linalg.cholesky(cyz)
chcxyz = cp.linalg.cholesky(cxyz)
# entropies in nats
# normalizations cancel for cmi
hz = cp.sum(cp.log(cp.diagonal(chcz)))
hxz = cp.sum(cp.log(cp.diagonal(chcxz)))
hyz = cp.sum(cp.log(cp.diagonal(chcyz)))
hxyz = cp.sum(cp.log(cp.diagonal(chcxyz)))
ln2 = cp.log(2)
if biascorrect:
psiterms = psi(
(ntrl - cp.arange(1, nvarxyz + 1)).astype(cp.float) / 2.) / 2.
dterm = (ln2 - cp.log(ntrl - 1.)) / 2.
hz = hz - nvarz * dterm - psiterms[:nvarz].sum()
hxz = hxz - nvarxz * dterm - psiterms[:nvarxz].sum()
hyz = hyz - nvaryz * dterm - psiterms[:nvaryz].sum()
hxyz = hxyz - nvarxyz * dterm - psiterms[:nvarxyz].sum()
# MI in bits
i = (hxz + hyz - hxyz - hz) / ln2
return i
def get_error(y, t):
# 2値の交差エントロピー誤差を返す。
eps = 1e-7
return -np.sum(t * np.log(y + eps) +
(1 - t) * np.log(1 - y + eps)) / len(y)
def slogdet(a):
"""Returns sign and logarithm of the determinant of an array.
It calculates the natural logarithm of the determinant of a given value.
Args:
a (cupy.ndarray): The input matrix with dimension ``(..., N, N)``.
Returns:
tuple of :class:`~cupy.ndarray`:
It returns a tuple ``(sign, logdet)``. ``sign`` represents each
sign of the determinant as a real number ``0``, ``1`` or ``-1``.
'logdet' represents the natural logarithm of the absolute of the
determinant.
If the determinant is zero, ``sign`` will be ``0`` and ``logdet``
will be ``-inf``.
The shapes of both ``sign`` and ``logdet`` are equal to
``a.shape[:-2]``.
.. warning::
This function calls one or more cuSOLVER routine(s) which may yield
invalid results if input conditions are not met.
To detect these invalid results, you can set the `linalg`
configuration to a value that is not `ignore` in
:func:`cupyx.errstate` or :func:`cupyx.seterr`.
.. warning::
To produce the same results as :func:`numpy.linalg.slogdet` for
singular inputs, set the `linalg` configuration to `raise`.
.. seealso:: :func:`numpy.linalg.slogdet`
"""
if a.ndim < 2:
msg = ('%d-dimensional array given. '
'Array must be at least two-dimensional' % a.ndim)
raise linalg.LinAlgError(msg)
_util._assert_nd_squareness(a)
dtype = numpy.promote_types(a.dtype.char, 'f')
real_dtype = numpy.dtype(dtype.char.lower())
if dtype not in (numpy.float32, numpy.float64,
numpy.complex64, numpy.complex128):
msg = ('dtype must be float32, float64, complex64, or complex128'
' (actual: {})'.format(a.dtype))
raise ValueError(msg)
a_shape = a.shape
shape = a_shape[:-2]
n = a_shape[-2]
if a.size == 0:
# empty batch (result is empty, too) or empty matrices det([[]]) == 1
sign = cupy.ones(shape, dtype)
logdet = cupy.zeros(shape, real_dtype)
return sign, logdet
lu, ipiv, dev_info = _decomposition._lu_factor(a, dtype)
# dev_info < 0 means illegal value (in dimensions, strides, and etc.) that
# should never happen even if the matrix contains nan or inf.
# TODO(kataoka): assert dev_info >= 0 if synchronization is allowed for
# debugging purposes.
diag = cupy.diagonal(lu, axis1=-2, axis2=-1)
logdet = cupy.log(cupy.abs(diag)).sum(axis=-1)
# ipiv is 1-origin
non_zero = cupy.count_nonzero(ipiv != cupy.arange(1, n + 1), axis=-1)
if dtype.kind == "f":
non_zero += cupy.count_nonzero(diag < 0, axis=-1)
# Note: sign == -1 ** (non_zero % 2)
sign = (non_zero % 2) * -2 + 1
if dtype.kind == "c":
sign = sign * cupy.prod(diag / cupy.abs(diag), axis=-1)
singular = dev_info > 0
return (
cupy.where(singular, dtype.type(0), sign.astype(dtype)).reshape(shape),
cupy.where(singular, real_dtype.type('-inf'), logdet).reshape(shape),
)
def log_softmax(x: xp.ndarray, axis=-1):
c = xp.max(x, axis=axis, keepdims=True) # [*, 1, *]
x2 = x - c # [*, ?, *]
logsumexp = xp.log(xp.exp(x2).sum(axis=axis, keepdims=True)) # [*, 1, *]
return x2 - logsumexp
def main():
nwaves = 256
pw = 256
dw = 512+16+3
A = lt.Propagation(nwaves, dw, pw)
B = to.Propagation(dw, pw, array_module=cp, asnumpy=cp.asnumpy)
shape = (nwaves, pw, pw)
nearplane = cp.random.rand(*shape) + 1j * cp.random.rand(*shape)
amplitude = plt.imread("/home/beams/DCHING/Pictures/images/Cryptomeria_japonica-0256.tif") / 255
phase = plt.imread("/home/beams/DCHING/Pictures/images/Erdhummel_Bombus_terrestris-0256.tif") / 255
nearplane[0] = cp.asarray(amplitude + 1j * phase)
nearplane = cp.ascontiguousarray(nearplane, dtype='complex64')
start = time.time()
farplaneB = B.fwd(nearplane)
stop = time.time()
print(farplaneB.shape, stop-start)
start = time.time()
farplaneA = A.fwd(nearplane)
stop = time.time()
print(farplaneA.shape, stop-start)
return
plt.figure()
plt.subplot(1, 3, 1)
plt.imshow(cp.log(cp.abs(farplaneA)).get()[0])
plt.colorbar()
plt.title('CUDA')
plt.subplot(1, 3, 2)
plt.imshow(cp.log(cp.abs(farplaneB)).get()[0])
plt.colorbar()
plt.title('CUPY')
plt.subplot(1, 3, 3)
plt.imshow(
cp.log(
cp.abs(farplaneA)
- cp.abs(farplaneB)
).get()[0]
)
plt.colorbar()
plt.title('DIFF')
plt.show()
# cp.testing.assert_array_equal(farplaneA, farplaneB)
nearplaneA = A.adj(farplaneB)
nearplaneB = B.adj(farplaneA)
plt.figure()
plt.subplot(1, 3, 1)
plt.imshow(nearplaneA.real.get()[0])
plt.colorbar()
plt.title('CUDA')
plt.subplot(1, 3, 2)
plt.imshow(nearplaneB.real.get()[0])
plt.colorbar()
plt.title('CUPY')
plt.subplot(1, 3, 3)
plt.imshow(
cp.log(
nearplaneB.real - nearplaneA.real
).get()[0]
)
plt.colorbar()
plt.title('DIFF')
plt.show()
cp.testing.assert_array_equal(nearplaneA, nearplaneB)
def choice(self, a, size=None, replace=True, p=None):
"""Returns an array of random values from a given 1-D array.
.. seealso::
:func:`cupy.random.choice` for full document,
:meth:`numpy.random.choice`
"""
if a is None:
raise ValueError('a must be 1-dimensional or an integer')
if isinstance(a, cupy.ndarray) and a.ndim == 0:
raise NotImplementedError
if isinstance(a, six.integer_types):
a_size = a
if a_size <= 0:
raise ValueError('a must be greater than 0')
else:
a = cupy.array(a, copy=False)
if a.ndim != 1:
raise ValueError('a must be 1-dimensional or an integer')
else:
a_size = len(a)
if a_size == 0:
raise ValueError('a must be non-empty')
if p is not None:
p = cupy.array(p)
if p.ndim != 1:
raise ValueError('p must be 1-dimensional')
if len(p) != a_size:
raise ValueError('a and p must have same size')
if not (p >= 0).all():
raise ValueError('probabilities are not non-negative')
p_sum = cupy.sum(p).get()
if not numpy.allclose(p_sum, 1):
raise ValueError('probabilities do not sum to 1')
if size is None:
raise NotImplementedError
shape = size
size = numpy.prod(shape)
if not replace and p is None:
if a_size < size:
raise ValueError(
'Cannot take a larger sample than population when '
'\'replace=False\'')
if isinstance(a, six.integer_types):
indices = cupy.arange(a, dtype='l')
else:
indices = a.copy()
self.shuffle(indices)
return indices[:size].reshape(shape)
if not replace:
raise NotImplementedError
if p is not None:
p = cupy.broadcast_to(p, (size, a_size))
index = cupy.argmax(cupy.log(p) +
cupy.random.gumbel(size=(size, a_size)),
axis=1)
if not isinstance(shape, six.integer_types):
index = cupy.reshape(index, shape)
else:
index = cupy.random.randint(0, a_size, size=shape)
# Align the dtype with NumPy
index = index.astype(cupy.int64, copy=False)
if isinstance(a, six.integer_types):
return index
if index.ndim == 0:
return cupy.array(a[index], dtype=a.dtype)
return a[index]
def select_log_next_cupy(X, gains, current_values, idxs):
gains[:] = cupy.sum(cupy.log(current_values + X + 1), axis=1)[idxs]
def beam_search(model, X, params, return_alphas=False, eos_sym=0, null_sym=2, model_ensemble=False, n_models=0):
"""
Beam search method for Cond models.
(https://en.wikibooks.org/wiki/Artificial_Intelligence/Search/Heuristic_search/Beam_search)
The algorithm in a nutshell does the following:
1. k = beam_size
2. open_nodes = [[]] * k
3. while k > 0:
3.1. Given the inputs, get (log) probabilities for the outputs.
3.2. Expand each open node with all possible output.
3.3. Prune and keep the k best nodes.
3.4. If a sample has reached the <eos> symbol:
3.4.1. Mark it as final sample.
3.4.2. k -= 1
3.5. Build new inputs (state_below) and go to 1.
4. return final_samples, final_scores
:param model: Model to use
:param X: Model inputs
:param params: Search parameters
:param return_alphas: Whether we should return attention weights or not.
:param eos_sym: <eos> symbol
:param null_sym: <null> symbol
:param model_ensemble: Whether we are using several models in an ensemble
:param n_models; Number of models in the ensemble.
:return: UNSORTED list of [k_best_samples, k_best_scores] (k: beam size)
"""
k = params['beam_size']
samples = []
sample_scores = []
pad_on_batch = params['pad_on_batch']
dead_k = 0 # samples that reached eos
live_k = 1 # samples that did not yet reach eos
hyp_samples = [[]] * live_k
hyp_scores = cp.zeros(live_k, dtype='float32')
ret_alphas = return_alphas or params['pos_unk']
if ret_alphas:
sample_alphas = []
hyp_alphas = [[]] * live_k
if pad_on_batch:
maxlen = int(len(X[params['dataset_inputs'][0]][0]) * params['output_max_length_depending_on_x_factor']) if \
params['output_max_length_depending_on_x'] else params['maxlen']
minlen = int(
len(X[params['dataset_inputs'][0]][0]) / params['output_min_length_depending_on_x_factor'] + 1e-7) if \
params['output_min_length_depending_on_x'] else 0
else:
minlen = int(np.argmax(X[params['dataset_inputs'][0]][0] == eos_sym) /
params['output_min_length_depending_on_x_factor'] + 1e-7) if \
params['output_min_length_depending_on_x'] else 0
maxlen = int(np.argmax(X[params['dataset_inputs'][0]][0] == eos_sym) * params[
'output_max_length_depending_on_x_factor']) if \
params['output_max_length_depending_on_x'] else params['maxlen']
maxlen = min(params['state_below_maxlen'] - 1, maxlen)
# we must include an additional dimension if the input for each timestep are all the generated "words_so_far"
if params['words_so_far']:
if k > maxlen:
raise NotImplementedError("BEAM_SIZE can't be higher than MAX_OUTPUT_TEXT_LEN on the current implementation.")
state_below = np.asarray([[null_sym]] * live_k) if pad_on_batch else np.asarray([np.zeros((maxlen, maxlen))] * live_k)
else:
state_below = np.asarray([null_sym] * live_k) if pad_on_batch else np.asarray([np.zeros(params['state_below_maxlen']) + null_sym] * live_k)
prev_out = [None] * n_models if model_ensemble else None
for ii in range(maxlen):
# for every possible live sample calc prob for every possible label
if params['optimized_search']: # use optimized search model if available
if model_ensemble:
[probs, prev_out, alphas] = model.predict_cond_optimized(X, state_below, params, ii, prev_out)
else:
[probs, prev_out] = model.predict_cond_optimized(X, state_below, params, ii, prev_out)
if ret_alphas:
alphas = prev_out[-1][0] # Shape: (k, n_steps)
prev_out = prev_out[:-1]
else:
probs = model.predict_cond(X, state_below, params, ii)
log_probs = cp.log(probs)
if minlen > 0 and ii < minlen:
log_probs[:, eos_sym] = -cp.inf
# total score for every sample is sum of -log of word prb
cand_scores = hyp_scores[:, None] - log_probs
cand_flat = cand_scores.flatten()
# Find the best options by calling argsort of flatten array
ranks_flat = cp.argsort(cand_flat)[:(k - dead_k)]
# Decypher flatten indices
voc_size = log_probs.shape[1]
trans_indices = ranks_flat // voc_size # index of row
word_indices = ranks_flat % voc_size # index of col
costs = cand_flat[ranks_flat]
best_cost = costs[0]
if cupy:
trans_indices = cp.asnumpy(trans_indices)
word_indices = cp.asnumpy(word_indices)
if ret_alphas:
alphas = cp.asnumpy(alphas)
# Form a beam for the next iteration
new_hyp_samples = []
new_trans_indices = []
new_hyp_scores = cp.zeros(k - dead_k, dtype='float32')
if ret_alphas:
new_hyp_alphas = []
for idx, [ti, wi] in list(enumerate(zip(trans_indices, word_indices))):
if params['search_pruning']:
if costs[idx] < k * best_cost:
new_hyp_samples.append(hyp_samples[ti] + [wi])
new_trans_indices.append(ti)
new_hyp_scores[idx] = copy.copy(costs[idx])
if ret_alphas:
new_hyp_alphas.append(hyp_alphas[ti] + [alphas[ti]])
else:
dead_k += 1
else:
new_hyp_samples.append(hyp_samples[ti] + [wi])
new_trans_indices.append(ti)
new_hyp_scores[idx] = copy.copy(costs[idx])
if ret_alphas:
new_hyp_alphas.append(hyp_alphas[ti] + [alphas[ti]])
# check the finished samples
new_live_k = 0
hyp_samples = []
hyp_scores = []
hyp_alphas = []
indices_alive = []
for idx in range(len(new_hyp_samples)):
if new_hyp_samples[idx][-1] == eos_sym: # finished sample
samples.append(new_hyp_samples[idx])
sample_scores.append(new_hyp_scores[idx])
if ret_alphas:
sample_alphas.append(new_hyp_alphas[idx])
dead_k += 1
else:
indices_alive.append(new_trans_indices[idx])
new_live_k += 1
hyp_samples.append(new_hyp_samples[idx])
hyp_scores.append(new_hyp_scores[idx])
if ret_alphas:
hyp_alphas.append(new_hyp_alphas[idx])
hyp_scores = cp.array(np.asarray(hyp_scores, dtype='float32'), dtype='float32')
live_k = new_live_k
if new_live_k < 1:
break
if dead_k >= k:
break
state_below = np.asarray(hyp_samples, dtype='int64')
state_below = np.hstack((np.zeros((state_below.shape[0], 1), dtype='int64') + null_sym, state_below)) \
if pad_on_batch else \
np.hstack((np.zeros((state_below.shape[0], 1), dtype='int64') + null_sym,
state_below,
np.zeros((state_below.shape[0],
max(params['state_below_maxlen'] - state_below.shape[1] - 1, 0)), dtype='int64')))
# we must include an additional dimension if the input for each timestep are all the generated words so far
if params['words_so_far']:
state_below = np.expand_dims(state_below, axis=0)
if params['optimized_search'] and ii > 0:
# filter next search inputs w.r.t. remaining samples
if model_ensemble:
for n_model in range(n_models):
# filter next search inputs w.r.t. remaining samples
for idx_vars in range(len(prev_out[n_model])):
prev_out[n_model][idx_vars] = prev_out[n_model][idx_vars][indices_alive]
else:
for idx_vars in range(len(prev_out)):
prev_out[idx_vars] = prev_out[idx_vars][indices_alive]
# dump every remaining one
if live_k > 0:
for idx in range(live_k):
samples.append(hyp_samples[idx])
sample_scores.append(hyp_scores[idx])
if ret_alphas:
sample_alphas.append(hyp_alphas[idx])
if ret_alphas:
return samples, sample_scores, np.asarray(sample_alphas)
else:
return samples, sample_scores, None
def logit(x):
return cp.log(x / (1 - x))
def convolutional_barycenter_gpu(
Hv ,
reg : float,
alpha : np.ndarray,
stabThresh = 1e-30,
niter = 1500,
tol = 1e-9,
sharpening = False,
verbose = False):
"""Main function solving wasserstein barycenter problem using gpu
Parameters:
Hv {Set of distributions (cparray)} --
reg {regularization term "gamma" } -- float superior to 0, generally equals size of space/40
alpha {list } -- set of weights
Keyword Parameters:
stabThresh {float} -- Stabilization threshold to prevent division by 0 (default: {1e-30})
niter {int } -- Maximum number of loop iteration (default: {1500})
tolerance {float} -- convergence tolerance at which point iterations stop (default: {1e-9})
sharpening {bool } -- Whether or not entropic sharpening is used (default: {False})
verbose {bool } -- verbose option
Returns:
cparray -- solution of weighted wassertein barycenter problem
"""
import cupy as cp
from cupyx.scipy.ndimage import gaussian_filter as cupyx_gaussian_filter
def K_cupyx(x):
return cupyx_gaussian_filter(x,sigma=reg)
def to_find_root(barycenter, H0, beta):
return entropy(barycenter**beta) - H0
alpha = cp.array(alpha)
alpha = alpha/alpha.sum()
Hv = cp.array(Hv)
for i in range(len(Hv)):
Hv[i] = Hv[i]/Hv[i].sum()
v = cp.ones(Hv.shape)
Kw = cp.ones(Hv.shape)
entropy_max = max_entropy (Hv )
barycenter = cp .zeros(Hv[0].shape)
cumtime_agg = 0
rolling_delta = []
cumtime = []
iterations = []
change = 1
for j in range(niter):
print("For every iteration.. ")
t0 = time.time()
barycenterOld = barycenter
barycenter = cp.zeros_like(Hv[0, :, :])
print("Hv shape is", Hv.shape)
for i in range(Hv.shape[0]):
#* for each of two distributions(which are identical)
#* distribution *i* becomes Kernel of (dist1 over the Kernel of v(i))
Kw[i, :, :] = K_cupyx(Hv[i, :, :] / cp.maximum(stabThresh,K_cupyx(v[i, :, :])) )
#* barycenter is barycenter plus weighted log of v(i)*Kw(i)
barycenter += alpha[i] * cp.log(cp.maximum(stabThresh, v[i, :, :]*Kw[i, :, :]))
barycenter = cp.exp(barycenter)
change = cp.sum(cp.abs(barycenter-barycenterOld))
if sharpening :
if (entropy(barycenter)) > (entropy_max):
beta = newton(lambda beta : to_find_root(barycenter,entropy_max,beta), 1, tol=1e-6)
if beta < 0 :
beta = 1
else :
beta = 1
barycenter = barycenter**beta
for i in range(Hv.shape[0]):
# assign to v(i) barycenter normalized by Kw(i)'s largest
v[i, :, :] = barycenter / cp.maximum(stabThresh, Kw[i, : ,: ])
elapsed = np.around(time.time() - t0, 4)
delta = np.around(change,10)
cumtime_agg += elapsed
iterations.append(j)
cumtime.append(cumtime_agg)
rolling_delta.append(float( delta ))
print(f"Refinement iter {j} | delta: {delta} | elapsed : {elapsed}")
if change < tol :
print(f"Exited. Change {change} under tolerance.")
log = {
"iterations" : iterations,
"cumtime" : cumtime,
"rolling_delta" : rolling_delta,
"exited_on" : j,
"exited_under_tolerance": True
}
# print(f"Exited with 0 on iter {j}")
return [ cp.asnumpy(barycenter),log ]
break
log = {
"iterations" : iterations,
"cumtime" : cumtime,
"rolling_delta" : rolling_delta,
"exited_on" : j,
"exited_under_tolerance": False
}
# print(f"Exited with 0 on iter {j}")
return [ cp.asnumpy(barycenter),log ]
def learnAndSolve8b(ctx):
"""This is the main optimization. Takes the longest time and uses the GPU heavily."""
Nbatch = ctx.intermediate.Nbatch
params = ctx.params
probe = ctx.probe
ir = ctx.intermediate
proc = ir.proc
iorig = ir.iorig
# TODO: move_to_config
NrankPC = 6 # this one is the rank of the PCs, used to detect spikes with threshold crossings
Nrank = 3 # this one is the rank of the templates
wTEMP, wPCA = extractTemplatesfromSnippets(proc=proc,
probe=probe,
params=params,
Nbatch=Nbatch,
nPCs=NrankPC)
# move these to the GPU
wPCA = cp.asarray(wPCA[:, :Nrank], dtype=np.float32, order='F')
wTEMP = cp.asarray(wTEMP, dtype=np.float32, order='F')
wPCAd = cp.asarray(wPCA, dtype=np.float64,
order='F') # convert to double for extra precision
nt0 = params.nt0
nt0min = params.nt0min
nBatches = Nbatch
NT = params.NT
Nfilt = params.Nfilt
Nchan = probe.Nchan
# two variables for the same thing? number of nearest channels to each primary channel
# TODO: unclear - let's fix this
NchanNear = min(probe.Nchan, 32)
Nnearest = min(probe.Nchan, 32)
# decay of gaussian spatial mask centered on a channel
sigmaMask = params.sigmaMask
batchstart = list(range(0, NT * nBatches + 1, NT))
# find the closest NchanNear channels, and the masks for those channels
iC, mask, C2C = getClosestChannels(probe, sigmaMask, NchanNear)
# sorting order for the batches
isortbatches = iorig
nhalf = int(ceil(nBatches / 2)) - 1 # halfway point
# this batch order schedule goes through half of the data forward and backward during the model
# fitting and then goes through the data symmetrically-out from the center during the final
# pass
ischedule = np.concatenate(
(np.arange(nhalf, nBatches), np.arange(nBatches - 1, nhalf - 1, -1)))
i1 = np.arange(nhalf - 1, -1, -1)
i2 = np.arange(nhalf, nBatches)
irounds = np.concatenate((ischedule, i1, i2))
niter = irounds.size
if irounds[niter - nBatches - 1] != nhalf:
# this check is in here in case I do somehting weird when I try different schedules
raise ValueError('Mismatch between number of batches')
# these two flags are used to keep track of what stage of model fitting we're at
# flag_final = 0
flag_resort = 1
# this is the absolute temporal offset in seconds corresponding to the start of the
# spike sorted time segment
t0 = 0 # ceil(params.trange(1) * ops.fs)
nInnerIter = 60 # this is for SVD for the power iteration
# schedule of learning rates for the model fitting part
# starts small and goes high, it corresponds approximately to the number of spikes
# from the past that were averaged to give rise to the current template
pmi = cp.exp(
-1. /
cp.linspace(params.momentum[0], params.momentum[1], niter - nBatches))
Nsum = min(
Nchan,
7) # how many channels to extend out the waveform in mexgetspikes
# lots of parameters passed into the CUDA scripts
Params = np.array([
NT, Nfilt, params.Th[0], nInnerIter, nt0, Nnearest, Nrank, params.lam,
pmi[0], Nchan, NchanNear, params.nt0min, 1, Nsum, NrankPC, params.Th[0]
],
dtype=np.float64)
# W0 has to be ordered like this
W0 = cp.transpose(
cp.atleast_3d(cp.asarray(wPCA, dtype=np.float64, order='F')),
[0, 2, 1])
# initialize the list of channels each template lives on
iList = cp.zeros((Nnearest, Nfilt), dtype=np.int32, order='F')
# initialize average number of spikes per batch for each template
nsp = cp.zeros((0, 1), dtype=np.float64, order='F')
# this flag starts 0, is set to 1 later
Params[12] = 0
# kernels for subsample alignment
Ka, Kb = getKernels(params)
p1 = .95 # decay of nsp estimate in each batch
ntot = 0
# this keeps track of dropped templates for debugging purposes
ndrop = np.zeros(2, dtype=np.float32, order='F')
# this is the minimum firing rate that all templates must maintain, or be dropped
m0 = params.minFR * params.NT / params.fs
# allocate variables when switching to extraction phase
# this holds spike times, clusters and other info per spike
st3 = [] # cp.zeros((int(1e7), 5), dtype=np.float32, order='F')
# these ones store features per spike
# Nnearest is the number of nearest templates to store features for
fW = LargeArrayWriter(ctx.path('fW', ext='.dat'),
dtype=np.float32,
shape=(Nnearest, -1))
# NchanNear is the number of nearest channels to take PC features from
fWpc = LargeArrayWriter(ctx.path('fWpc', ext='.dat'),
dtype=np.float32,
shape=(NchanNear, Nrank, -1))
for ibatch in tqdm(range(niter), desc="Optimizing templates"):
# korder is the index of the batch at this point in the schedule
korder = int(irounds[ibatch])
# k is the index of the batch in absolute terms
k = int(isortbatches[korder])
logger.debug("Batch %d/%d, %d templates.", ibatch, niter, Nfilt)
if ibatch > niter - nBatches - 1 and korder == nhalf:
# this is required to revert back to the template states in the middle of the
# batches
W, dWU = ir.W, ir.dWU
logger.debug('Reverted back to middle timepoint.')
if ibatch < niter - nBatches:
# obtained pm for this batch
Params[8] = float(pmi[ibatch])
pm = pmi[ibatch] * ones((Nfilt, ), dtype=np.float64, order='F')
# loading a single batch (same as everywhere)
offset = Nchan * batchstart[k]
dat = proc.flat[offset:offset + NT * Nchan].reshape((-1, Nchan),
order='F')
dataRAW = cp.asarray(dat, dtype=np.float32) / params.scaleproc
if ibatch == 0:
# only on the first batch, we first get a new set of spikes from the residuals,
# which in this case is the unmodified data because we start with no templates
# CUDA function to get spatiotemporal clips from spike detections
dWU, cmap = mexGetSpikes2(Params, dataRAW, wTEMP, iC)
dWU = cp.asarray(dWU, dtype=np.float64, order='F')
# project these into the wPCA waveforms
dWU = cp.reshape(cp.dot(
wPCAd,
cp.dot(wPCAd.T, dWU.reshape((dWU.shape[0], -1), order='F'))),
dWU.shape,
order='F')
# initialize the low-rank decomposition with standard waves
W = W0[:, cp.ones(dWU.shape[2], dtype=np.int32), :]
Nfilt = W.shape[1] # update the number of filters/templates
# initialize the number of spikes for new templates with the minimum allowed value,
# so it doesn't get thrown back out right away
nsp = _extend(nsp, 0, Nfilt, m0)
Params[1] = Nfilt # update in the CUDA parameters
if flag_resort:
# this is a flag to resort the order of the templates according to best peak
# channel
# this is important in order to have cohesive memory requests from the GPU RAM
# max channel (either positive or negative peak)
iW = cp.argmax(cp.abs(dWU[nt0min - 1, :, :]), axis=0)
# iW = int32(squeeze(iW))
isort = cp.argsort(iW) # sort by max abs channel
iW = iW[isort]
W = W[:,
isort, :] # user ordering to resort all the other template variables
dWU = dWU[:, :, isort]
nsp = nsp[isort]
# decompose dWU by svd of time and space (via covariance matrix of 61 by 61 samples)
# this uses a "warm start" by remembering the W from the previous iteration
W, U, mu = mexSVDsmall2(Params, dWU, W, iC, iW, Ka, Kb)
# UtU is the gram matrix of the spatial components of the low-rank SVDs
# it tells us which pairs of templates are likely to "interfere" with each other
# such as when we subtract off a template
# this needs to change (but I don't know why!)
UtU, maskU = getMeUtU(iW, iC, mask, Nnearest, Nchan)
# main CUDA function in the whole codebase. does the iterative template matching
# based on the current templates, gets features for these templates if requested
# (featW, featPC),
# gets scores for the template fits to each spike (vexp), outputs the average of
# waveforms assigned to each cluster (dWU0),
# and probably a few more things I forget about
st0, id0, x0, featW, dWU0, drez, nsp0, featPC, vexp = mexMPnu8(
Params, dataRAW, U, W, mu, iC, iW, UtU, iList, wPCA)
logger.debug("%d spikes.", x0.size)
# Sometimes nsp can get transposed (think this has to do with it being
# a single element in one iteration, to which elements are added
# nsp, nsp0, and pm must all be row vectors (Nfilt x 1), so force nsp
# to be a row vector.
# nsp = cp.atleast_2d(nsp)
# nsprow, nspcol = nsp.shape
# if nsprow < nspcol:
# nsp = nsp.T
nsp = nsp.squeeze()
# updates the templates as a running average weighted by recency
# since some clusters have different number of spikes, we need to apply the
# exp(pm) factor several times, and fexp is the resulting update factor
# for each template
fexp = np.exp(nsp0 * cp.log(pm[:Nfilt]))
fexp = cp.reshape(fexp, (1, 1, -1), order='F')
dWU = dWU * fexp + (1 - fexp) * (
dWU0 / cp.reshape(cp.maximum(1, nsp0), (1, 1, -1), order='F'))
# nsp just gets updated according to the fixed factor p1
nsp = nsp * p1 + (1 - p1) * nsp0
if ibatch == niter - nBatches - 1:
# if we reached this point, we need to disable secondary template updates
# like dropping, and adding new templates. We need to memorize the state of the
# templates at this timepoint, and set the processing mode to "extraction and
# tracking"
flag_resort = 0 # no need to resort templates by channel any more
# flag_final = 1 # this is the "final" pass
# final clean up, triage templates one last time
W, U, dWU, mu, nsp, ndrop = triageTemplates2(
params, iW, C2C, W, U, dWU, mu, nsp, ndrop)
# final number of templates
Nfilt = W.shape[1]
Params[1] = Nfilt
# final covariance matrix between all templates
WtW, iList = getMeWtW(W, U, Nnearest)
# iW is the final channel assigned to each template
iW = cp.argmax(cp.abs(dWU[nt0min - 1, :, :]), axis=0)
# extract ALL features on the last pass
Params[
12] = 2 # this is a flag to output features (PC and template features)
# different threshold on last pass?
Params[2] = params.Th[
-1] # usually the threshold is much lower on the last pass
# memorize the state of the templates
logger.debug("Memorized middle timepoint.")
ir.W, ir.dWU, ir.U, ir.mu = W, dWU, U, mu
ir.Wraw = cp.zeros((U.shape[0], W.shape[0], U.shape[1]),
dtype=np.float64,
order='F')
for n in range(U.shape[1]):
# temporarily use U rather Urot until I have a chance to test it
ir.Wraw[:, :, n] = mu[n] * cp.dot(U[:, n, :], W[:, n, :].T)
if ibatch < niter - nBatches - 1:
# during the main "learning" phase of fitting a model
if ibatch % 5 == 0:
# this drops templates based on spike rates and/or similarities to
# other templates
W, U, dWU, mu, nsp, ndrop = triageTemplates2(
params, iW, C2C, W, U, dWU, mu, nsp, ndrop)
Nfilt = W.shape[1] # update the number of filters
Params[1] = Nfilt
# this adds new templates if they are detected in the residual
dWU0, cmap = mexGetSpikes2(Params, drez, wTEMP, iC)
if dWU0.shape[2] > 0:
# new templates need to be integrated into the same format as all templates
# apply PCA for smoothing purposes
dWU0 = cp.reshape(cp.dot(
wPCAd,
cp.dot(
wPCAd.T,
dWU0.reshape(
(dWU0.shape[0], dWU0.shape[1] * dWU0.shape[2]),
order='F'))),
dWU0.shape,
order='F')
dWU = cp.concatenate((dWU, dWU0), axis=2)
m = dWU0.shape[2]
# initialize temporal components of waveforms
W = _extend(W,
Nfilt,
Nfilt + m,
W0[:, cp.ones(m, dtype=np.int32), :],
axis=1)
# initialize the number of spikes with the minimum allowed
nsp = _extend(nsp, Nfilt, Nfilt + m,
params.minFR * NT / params.fs)
# initialize the amplitude of this spike with a lowish number
mu = _extend(mu, Nfilt, Nfilt + m, 10)
# if the number of filters exceed the maximum allowed, clip it
Nfilt = min(params.Nfilt, W.shape[1])
Params[1] = Nfilt
W = W[:, :
Nfilt, :] # remove any new filters over the maximum allowed
dWU = dWU[:, :, :
Nfilt] # remove any new filters over the maximum allowed
nsp = nsp[:
Nfilt] # remove any new filters over the maximum allowed
mu = mu[:
Nfilt] # remove any new filters over the maximum allowed
if ibatch > niter - nBatches - 1:
# during the final extraction pass, this keeps track of all spikes and features
# we memorize the spatio-temporal decomposition of the waveforms at this batch
# this is currently only used in the GUI to provide an accurate reconstruction
# of the raw data at this time
ir.WA[..., k] = cp.asnumpy(W)
ir.UA[..., k] = cp.asnumpy(U)
ir.muA[..., k] = cp.asnumpy(mu)
# we carefully assign the correct absolute times to spikes found in this batch
ioffset = params.ntbuff - 1
if k == 0:
ioffset = 0 # the first batch is special (no pre-buffer)
toff = nt0min + t0 - ioffset + (NT - params.ntbuff) * k
st = toff + st0
st30 = np.c_[
cp.asnumpy(st), # spike times
cp.asnumpy(id0), # spike clusters (0-indexing)
cp.asnumpy(x0), # template amplitudes
cp.asnumpy(vexp), # residual variance of this spike
korder *
np.ones(st.size), # batch from which this spike was found
]
# Check the number of spikes.
assert st30.shape[0] == featW.shape[1] == featPC.shape[2]
st3.append(st30)
fW.append(featW)
fWpc.append(featPC)
ntot = ntot + x0.size # keeps track of total number of spikes so far
if ibatch == niter - nBatches - 1:
# these next three store the low-d template decompositions
ir.WA = np.zeros((nt0, Nfilt, Nrank, nBatches),
dtype=np.float32,
order='F')
ir.UA = np.zeros((Nchan, Nfilt, Nrank, nBatches),
dtype=np.float32,
order='F')
ir.muA = np.zeros((Nfilt, nBatches), dtype=np.float32, order='F')
if ibatch % 100 == 0:
# this is some of the relevant diagnostic information to be printed during training
logger.info(('%d / %d batches, %d units, nspks: %2.4f, mu: %2.4f, '
'nst0: %d, merges: %2.4f, %2.4f'), ibatch, niter,
Nfilt, nsp.sum(), median(mu), st0.size, *ndrop)
free_gpu_memory()
# Close the large array writers and save the JSON metadata files to disk.
fW.close()
fWpc.close()
# just display the total number of spikes
logger.info("Found %d spikes.", ntot)
# Save results to the ctx.intermediate object.
ir.st3 = np.concatenate(st3, axis=0)
# the similarity score between templates is simply the correlation,
# taken as the max over several consecutive time delays
ir.simScore = cp.asnumpy(cp.max(WtW, axis=2))
# NOTE: these are now already saved by LargeArrayWriter
# fWa = np.concatenate(fW, axis=-1)
# fWpca = np.concatenate(fWpc, axis=-1)
# the template features are stored in cProj, like in Kilosort1
# ir.cProj = fWa.T
# the neihboring templates idnices are stored in iNeigh
ir.iNeigh = cp.asnumpy(iList)
# permute the PC projections in the right order
# ir.cProjPC = np.transpose(fWpca, (2, 1, 0))
# iNeighPC keeps the indices of the channels corresponding to the PC features
ir.iNeighPC = cp.asnumpy(iC[:, iW])
# Number of spikes.
assert ir.st3.shape[0] == fW.shape[-1] == fWpc.shape[-1]
# this whole next block is just done to compress the compressed templates
# we separately svd the time components of each template, and the spatial components
# this also requires a careful decompression function, available somewhere in the GUI code
nKeep = min(Nchan * 3, 20) # how many PCs to keep
W_a = np.zeros((nt0 * Nrank, nKeep, Nfilt), dtype=np.float32)
W_b = np.zeros((nBatches, nKeep, Nfilt), dtype=np.float32)
U_a = np.zeros((Nchan * Nrank, nKeep, Nfilt), dtype=np.float32)
U_b = np.zeros((nBatches, nKeep, Nfilt), dtype=np.float32)
for j in tqdm(range(Nfilt), desc='Compressing templates'):
# do this for every template separately
WA = np.reshape(ir.WA[:, j, ...], (-1, nBatches), order='F')
# svd on the GPU was faster for this, but the Python randomized CPU version
# might be faster still
# WA = gpuArray(WA)
A, B, C = svdecon_cpu(WA)
# W_a times W_b results in a reconstruction of the time components
W_a[:, :, j] = np.dot(A[:, :nKeep], B[:nKeep, :nKeep])
W_b[:, :, j] = C[:, :nKeep]
UA = np.reshape(ir.UA[:, j, ...], (-1, nBatches), order='F')
# UA = gpuArray(UA)
A, B, C = svdecon_cpu(UA)
# U_a times U_b results in a reconstruction of the time components
U_a[:, :, j] = np.dot(A[:, :nKeep], B[:nKeep, :nKeep])
U_b[:, :, j] = C[:, :nKeep]
logger.info('Finished compressing time-varying templates.')
return Bunch(
wPCA=wPCA[:, :Nrank],
wTEMP=wTEMP,
st3=ir.st3,
simScore=ir.simScore,
# cProj=ir.cProj,
# cProjPC=ir.cProjPC,
iNeigh=ir.iNeigh,
iNeighPC=ir.iNeighPC,
WA=ir.WA,
UA=ir.UA,
W=ir.W,
U=ir.U,
dWU=ir.dWU,
mu=ir.mu,
W_a=W_a,
W_b=W_b,
U_a=U_a,
U_b=U_b,
)
def cross_entropy(label, prob):
loss = -np.sum(label * np.log(prob))
return loss