def yangDistributionDifference(aNeg, bNeg, aPos, bPos, p=1):
"""
Eq. (7) from :
Yang, R., Jiang, Y., Mathews, S. et al.
Data Min Knowl Disc (2019) 33: 995.
https://doi.org/10.1007/s10618-019-00622-6
"""
sampleSize = 1000
negSample = xp.random.beta(aNeg, bNeg, sampleSize)
posSample = xp.random.beta(aPos, bPos, sampleSize)
negPDF_NEG, posPDF_NEG, pdfDiffPos_NEG, pdfDiffNeg_NEG, pdfMax_NEG = calcDifference(
negSample, aNeg, bNeg, aPos, bPos)
negPDF_POS, posPDF_POS, pdfDiffPos_POS, pdfDiffPOS_POS, pdfMax_POS = calcDifference(
posSample, aNeg, bNeg, aPos, bPos)
numerator1 = xp.mean(pdfDiffNeg_NEG / negPDF_NEG)
numerator2 = xp.mean(pdfDiffPos_POS / posPDF_POS)
sumVecs = xp.power(numerator1,
xp.ones_like(numerator1) * p) + xp.power(
numerator2,
xp.ones_like(numerator2) * p)
dPHat = xp.power(sumVecs, xp.ones_like(sumVecs) * (1 / p))
dTermNeg = (posPDF_NEG * 0.5) + (negPDF_NEG * 0.5)
dTermPos = (posPDF_POS * 0.5) + (negPDF_POS * 0.5)
denominator = (xp.sum(pdfMax_NEG / dTermNeg) +
xp.sum(pdfMax_POS / dTermPos)) / (2 * sampleSize)
return dPHat / denominator
def test_cross_correlate_masked_output_shape():
"""Masked normalized cross-correlation should return a shape
of N + M + 1 for each transform axis."""
shape1 = (15, 4, 5)
shape2 = (6, 12, 7)
expected_full_shape = tuple(np.array(shape1) + np.array(shape2) - 1)
expected_same_shape = shape1
arr1 = cp.zeros(shape1)
arr2 = cp.zeros(shape2)
# Trivial masks
m1 = cp.ones_like(arr1)
m2 = cp.ones_like(arr2)
full_xcorr = cross_correlate_masked(arr1,
arr2,
m1,
m2,
axes=(0, 1, 2),
mode="full")
assert full_xcorr.dtype.kind != "c" # grlee77: output should be real
assert full_xcorr.shape == expected_full_shape
same_xcorr = cross_correlate_masked(arr1,
arr2,
m1,
m2,
axes=(0, 1, 2),
mode="same")
assert same_xcorr.shape == expected_same_shape
def triple_phase_boundary(img):
phases = cp.unique(cp.asarray(img))
if len(phases) != 3:
return None
shape = img.shape
dim = len(shape)
ph_maps = []
img = cp.pad(cp.asarray(img), 1, 'constant', constant_values=-1)
if dim == 2:
x, y = shape
total_edges = (x - 1) * (y - 1)
for ph in phases:
ph_map = cp.zeros_like(img)
ph_map_temp = cp.zeros_like(img)
ph_map_temp[img == ph] = 1
for i in [0, 1]:
for j in [0, 1]:
ph_map += cp.roll(cp.roll(ph_map_temp, i, 0), j, 1)
ph_maps.append(ph_map)
tpb_map = cp.ones_like(img)
for ph_map in ph_maps:
tpb_map *= ph_map
tpb_map[tpb_map > 1] = 1
tpb_map = tpb_map[1:-1, 1:-1]
tpb = np.sum(tpb_map)
else:
tpb = 0
x, y, z = shape
total_edges = z * (x - 1) * (y - 1) + x * (y - 1) * (z - 1) + y * (
x - 1) * (z - 1)
print(total_edges)
for d in range(dim):
ph_maps = []
for ph in phases:
ph_map = cp.zeros_like(img)
ph_map_temp = cp.zeros_like(img)
ph_map_temp[img == ph] = 1
for i in [0, 1]:
for j in [0, 1]:
d1 = (d + 1) % 3
d2 = (d + 2) % 3
ph_map += cp.roll(cp.roll(ph_map_temp, i, d1), j, d2)
ph_maps.append(ph_map)
tpb_map = cp.ones_like(img)
for ph_map in ph_maps:
tpb_map *= ph_map
tpb_map[tpb_map > 1] = 1
tpb_map = tpb_map[1:-1, 1:-1, 1:-1]
tpb += np.sum(tpb_map)
return tpb / total_edges
def test_cross_correlate_masked_test_against_mismatched_dimensions():
"""Masked normalized cross-correlation should raise an error if array
dimensions along non-transformation axes are mismatched."""
shape1 = (23, 1, 1)
shape2 = (6, 2, 2)
arr1 = cp.zeros(shape1)
arr2 = cp.zeros(shape2)
# Trivial masks
m1 = cp.ones_like(arr1)
m2 = cp.ones_like(arr2)
with pytest.raises(ValueError):
cross_correlate_masked(arr1, arr2, m1, m2, axes=(1, 2))
def __init__(self, basis, lows, highs, resolutions, fine_all=False, linspace=False):
# Grids
self.x = Grid1D(low=lows[0], high=highs[0], res=resolutions[0],
basis=basis.basis_x, spectrum=True, fine=fine_all, linspace=linspace)
self.y = Grid1D(low=lows[1], high=highs[1], res=resolutions[1],
basis=basis.basis_y, spectrum=True, fine=fine_all, linspace=linspace)
# res
self.res_ghosts = [self.x.res_ghosts, self.y.res_ghosts]
self.orders = [self.x.order, self.y.order]
# spectral radius squared (for laplacian)
self.kr_sq = (outer2(self.x.d_wave_numbers, cp.ones_like(self.y.d_wave_numbers)) ** 2.0 +
outer2(cp.ones_like(self.x.d_wave_numbers), self.y.d_wave_numbers) ** 2.0)
# wave number vector
self.wave_numbers = cp.array([self.x.d_wave_numbers, self.y.d_wave_numbers])
def liqlatt(self, sigma_A, gamma_A):
'''Apply liquidization transform to reciprocal lattice'''
if self.abc[0] == 0:
msg = 'Provide rlatt_vox to apply liqlatt (recip. lattice voxel dimensions)'
raise AttributeError(msg)
s_sq = (2 * cp.pi * sigma_A * self.dgen.qrad)**2
n_max = 0
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
if n_max == 0:
#bzone = cp.zeros(self.abc)
#bzone[self.abc[0]//2, self.abc[1]//2, self.abc[2]//2] = 1
#return cp.tile(bzone, ncells)
print('Returning ones array')
return cp.ones_like(s_sq)
liq = cp.zeros_like(s_sq)
for n in range(1, n_max):
weight = cp.exp(-s_sq + n * cp.log(s_sq) -
float(special.loggamma(n + 1)))
factor = self.corrdisp_factor(gamma_A / n)
liq += weight * factor
sys.stderr.write('\rLiquidizing: %d/%d' % (n, n_max - 1))
sys.stderr.write('\n')
return liq
def stack_backward(gradient: Tensor, tensors: List[Tensor], axis: int = 0):
_check_tensors(*tensors)
engine = _get_engine(*tensors)
grad_arrays = engine.split(gradient.data, len(tensors), axis=axis)
for idx, tensor in enumerate(tensors):
_set_grad(tensor, data=grad_arrays[idx] * engine.ones_like(tensor.data))
def test_prior_in_posteriors(self):
for frame in self.data:
frame["prior"] = 1
like = HyperparameterLikelihood(posteriors=self.data,
hyper_prior=self.model)
self.assertTrue(
xp.array_equal(like.sampling_prior, xp.ones_like(like.data["a"])))
def curvature_to_height(image, h2, iterations=2000):
f = image[..., 0]
A = image[..., 3]
u = cup.ones_like(f) * 0.5
k = 1
t = np.empty_like(u, dtype=np.float32)
# periodic gauss seidel iteration
for ic in range(iterations):
if ic % 100 == 0:
print(ic)
# roll k, axis=0
t[:-k, :] = u[k:, :]
t[-k:, :] = u[:k, :]
# roll -k, axis=0
t[k:, :] += u[:-k, :]
t[:k, :] += u[-k:, :]
# roll k, axis=1
t[:, :-k] += u[:, k:]
t[:, -k:] += u[:, :k]
# roll -k, axis=1
t[:, k:] += u[:, :-k]
t[:, :k] += u[:, -k:]
t -= h2 * f
t *= 0.25
u = t * A
u = -u
u -= cup.min(u)
u /= cup.max(u)
return cup.dstack([u, u, u, image[..., 3]])
def forward(self, bottom, top): self.label = cp.asarray(copy.deepcopy(bottom[1].data),cp.uint8) prob = cp.asarray(copy.deepcopy(bottom[0].data),cp.float64) prob = cp.subtract(prob,cp.max(prob,axis=1)[:,cp.newaxis,...]) prob = cp.exp(prob) self.softmax = cp.divide(prob,cp.sum(prob,axis=1)[:,cp.newaxis,...]) ## mask self.weight_mask = cp.ones_like(self.label, cp.float64) for weight_id in self.weight_dic: self.weight_mask[self.label == weight_id] = self.weight_dic[weight_id] if self.has_ignore_label: self.weight_mask[self.label == self.ignore_label] = 0 # self.weight_mask[self.label == 0] = 0.3 # self.weight_mask[self.label == 1] = 0.25 # self.weight_mask[self.label == 2] = 5 # self.weight_mask[self.label == 4] = 2 self.label[self.label == 3] = 2 compute_count = self.weight_mask[self.weight_mask != 0].size ## nomalize mask self.weight_mask = cp.divide(self.weight_mask, cp.divide(cp.sum(self.weight_mask), compute_count)) ## compute loss prob_compute_matrix = copy.deepcopy(self.softmax[self.index_0,self.label,self.index_2,self.index_3]) prob_compute_matrix[prob_compute_matrix < (1e-10)] = 1e-10 loss = - cp.divide(cp.sum(cp.multiply(cp.log(prob_compute_matrix),self.weight_mask)),compute_count) loss = cp.asnumpy(loss) top[0].data[...] = loss
def forward(self, x):
self.mask = (cp.absolute(x) > 1)
out = cp.ones_like(x, dtype=np.float32)
out[x < 0.5] = 0
out[x < -0.5] = -1
return out
def _generate_init_mask(I, T):
"""
Generate mask estimation based on SML.
Parameters
----------
I : list of np.ndarray
List of original raw images.
T : float
Blur level criteria.
"""
# temporary buffer for SML
S = cp.array(I[0])
nelem = S.size
n_threads = 1024
block_sz = (n_threads, )
grid_sz = (int(ceil(nelem / n_threads)), )
M, V = cp.ones_like(S, dtype=cp.int32), sml(S, T)
for i, iI in enumerate(I[1:], 2):
S = cp.array(iI)
S = sml(S, T)
keep_max_kernel(grid_sz, block_sz, (M, V, S, nelem, i))
return cp.asnumpy(M)
def all_pair_dist_cuda(X1, X2, feat, metric='cosine'):
if metric=='cosine':
norm1 = cp.einsum('ij, ij->i', X1, X1)
norm1 = cp.sqrt(norm1, norm1).reshape(-1, 1)
norm2 = cp.einsum('ij, ij->i', X2, X2)
norm2 = cp.sqrt(norm2, norm2).reshape(-1, 1)
return cp.dot(X2/norm2, (X1/norm1).T)
else:
n1 = len(X1)
n2 = len(X2)
nf = len(feat)
feat = cp.array(feat)
X1 = cp.array(X1.reshape(1, n1, -1))
X2 = cp.array(X2.reshape(1, n2, -1))
mat1 = cp.repeat(X1, n2, axis=0)
mat2 = cp.repeat(X2.reshape(n2, 1, nf), n1, axis=1)
isbow = cp.repeat(cp.repeat((feat < 2100).reshape(1, 1, nf), n1, axis=1), n2, axis=0)
count_mat = nf - cp.sum((mat1 == 0) & (mat2 == 0) & isbow, axis=2)
zeros = (count_mat != 0)
dist_mat = cp.ones_like(count_mat)
dist_mat[zeros] = cp.sum(np.cbs(mat1 - mat2), axis=2)[zeros] / count_mat[zeros]
return cp.asnumpy(dist_mat)
def test_transform_resnet50(self):
""" """
cp.random.seed(0)
cp.cuda.Device(0).use()
with chainer.using_config('dtype', chainer.mixed16):
net = ResNet50(arch='he') # stride_first is True
net.to_gpu()
x = chainer.Variable(
cp.random.normal(size=(1, 3, 224, 224)).astype('float16'))
y1 = net(x)
net_ = AdaLossScaled(net,
init_scale=16.,
transforms=[
AdaLossTransformLinear(),
AdaLossTransformBottleneck(),
AdaLossTransformConv2DBNActiv(),
],
verbose=False,
cfg=CFG)
net_.to_gpu()
y2 = net_(x)
self.assertTrue(cp.allclose(y1.array, y2.array))
y2.grad = cp.ones_like(y2.array, dtype='float16')
y2.backward()
self.assertTrue('loss_scale' in x.grad_var.__dict__)
self.assertEqual(x.grad_var.__dict__['loss_scale'],
CFG['accum_upper_bound'])
def lrelu(x, alpha=0.01, derivative=False):
res = x
if derivative:
dx = np.ones_like(res)
dx[res < 0] = alpha
return dx
else:
return np.maximum(x, x * alpha, x)
def test_background_one_region_center(self):
x = cp.zeros((3, 3, 3), int)
x[1, 1, 1] = 1
lb = cp.ones_like(x) * BG
lb[1, 1, 1] = 1
assert_array_equal(label(x, connectivity=1, background=0), lb)
def prob(self, dataset, axis=None):
self._parameter_conversion(dataset) # Do parameter conversion regardless
p = 1
for i in [1,2]:
p *= bilby.core.prior.Uniform(name='magn', minimum=0, maximum=1).prob(dataset["a_{}".format(i)])
p *= bilby.core.prior.Sine(name='tilt').prob(dataset["tilt_{}".format(i)])
p *= xp.ones_like(dataset["a_{}".format(i)])*1./(2*np.pi) # Independent of how one defines the domain
return xp.array(p)
def scaled_dot_product_attention(queries, keys, values, scale=1., mask=None):
x1 = F.matmul(queries, keys, transb=True) * xp.array(scale,
dtype=keys.dtype)
x2 = F.where(mask,
xp.ones_like(x1.array) *
-xp.inf, x1) if mask is not None else x1
x3 = F.softmax(x2, axis=-1)
x4 = F.matmul(x3, values)
return x4
def generate_q_u_matrix(x_coordinate: cp.array,
y_coordinate: cp.array) -> tuple:
flatten_flag = x_coordinate.ndim > 1
if flatten_flag:
x_coordinate = x_coordinate.flatten()
y_coordinate = y_coordinate.flatten()
t, u = cp.modf(y_coordinate)
u = u.astype(int)
uy = cp.vstack([
cp.minimum(cp.maximum(u - 1, 0), h - 1),
cp.minimum(cp.maximum(u, 0), h - 1),
cp.minimum(cp.maximum(u + 1, 0), h - 1),
cp.minimum(cp.maximum(u + 2, 0), h - 1),
]).astype(int)
Qy = cp.dot(
coeff,
cp.vstack([
cp.ones_like(t, dtype=cp.float32), t,
cp.power(t, 2),
cp.power(t, 3)
]))
t, u = cp.modf(x_coordinate)
u = u.astype(int)
ux = cp.vstack([
cp.minimum(cp.maximum(u - 1, 0), w - 1),
cp.minimum(cp.maximum(u, 0), w - 1),
cp.minimum(cp.maximum(u + 1, 0), w - 1),
cp.minimum(cp.maximum(u + 2, 0), w - 1),
])
Qx = cp.dot(
coeff,
cp.vstack([
cp.ones_like(t, dtype=cp.float32), t,
cp.power(t, 2),
cp.power(t, 3)
]))
if flatten_flag:
Qx = Qx.reshape(4, frame_n, int(w * mag)).transpose(1, 0, 2).copy()
Qy = Qy.reshape(4, frame_n, int(h * mag)).transpose(1, 0, 2).copy()
ux = ux.reshape(4, frame_n, int(w * mag)).transpose(1, 0, 2).copy()
uy = uy.reshape(4, frame_n, int(h * mag)).transpose(1, 0, 2).copy()
return Qx, Qy, ux, uy
def __init__(self,basis_number,extended_basis_number,t_start = 0,t_end=1,mempool=None):
self.basis_number = basis_number
self.extended_basis_number = extended_basis_number
self.basis_number_2D = (2*basis_number-1)*basis_number
self.basis_number_2D_ravel = (2*basis_number*basis_number-2*basis_number+1)
self.basis_number_2D_sym = (2*basis_number-1)*(2*basis_number-1)
self.extended_basis_number_2D = (2*extended_basis_number-1)*extended_basis_number
self.extended_basis_number_2D_sym = (2*extended_basis_number-1)*(2*extended_basis_number-1)
self.t_end = t_end
self.t_start = t_start
self.verbose = True
if mempool is None:
mempool = cp.get_default_memory_pool()
# pinned_mempool = cp.get_default_pinned_memory_pool()
# self.ix = cp.zeros((2*self.basis_number-1,2*self.basis_number-1),dtype=cp.int32)
# self.iy = cp.zeros((2*self.basis_number-1,2*self.basis_number-1),dtype=cp.int32)
temp = cp.arange(-(self.basis_number-1),self.basis_number,dtype=cp.int32)
# for i in range(2*self.basis_number-1):
# self.ix[i,:] = temp
# self.iy[:,i] = temp
self.ix,self.iy = cp.meshgrid(temp,temp)
if self.verbose:
print("Used bytes so far, after creating ix and iy {}".format(mempool.used_bytes()))
self.Ddiag = -(2*util.PI)**2*(self.ix.ravel(ORDER)**2+self.iy.ravel(ORDER)**2)
self.OnePerDdiag = 1/self.Ddiag
self.Dmatrix = cpx.scipy.sparse.diags(self.Ddiag,dtype=cp.float32)
if self.verbose:
print("Used bytes so far, after creating Dmatrix {}".format(mempool.used_bytes()))
# self.Imatrix = cp.eye((2*self.basis_number-1)**2,dtype=cp.int8)
self.Imatrix = cpx.scipy.sparse.identity((2*self.basis_number-1)**2,dtype=cp.float32)
self.Imatrix_dense = cp.eye((2*self.basis_number-1)**2,dtype=cp.float32)
if self.verbose:
print("Used bytes so far, after creating Imatrix {}".format(mempool.used_bytes()))
#K matrix --> 1D fourier index to 2D fourier index
ones_temp = cp.ones_like(temp)
self.Kmatrix = cp.vstack((cp.kron(temp,ones_temp),cp.kron(ones_temp,temp)))
if self.verbose:
print("Used bytes so far, after creating Kmatrix {}".format(mempool.used_bytes()))
#implement fft plan here
self._plan_fft2 = None
# self._fft2_axes = (-2, -1)
self._plan_ifft2 = None
x = np.concatenate((np.arange(self.basis_number)+1,np.zeros(self.basis_number-1)))
toep = sla.toeplitz(x)
self._Umask = cp.asarray(np.kron(toep,toep),dtype=cp.int16)
def test_background_one_region_center(self):
x = cp.zeros((3, 3, 3), int)
x[1, 1, 1] = 1
lb = cp.ones_like(x) * BG
lb[1, 1, 1] = 1
with expected_warnings(["use 'connectivity'"]):
assert_array_equal(label(x, neighbors=4, background=0), lb)
assert_array_equal(label(x, connectivity=1, background=0), lb)
def forward(self, bottom, top): self.label = cp.asarray(copy.deepcopy(bottom[1].data),cp.uint8) prob = cp.asarray(copy.deepcopy(bottom[0].data),cp.float64) prob = cp.subtract(prob,cp.max(prob,axis=1)[:,cp.newaxis,...]) prob = cp.exp(prob) self.softmax = cp.divide(prob,cp.sum(prob,axis=1)[:,cp.newaxis,...]) ## mask self.weight_mask = cp.ones_like(self.label, cp.float64) for weight_id in self.weight_dic: self.weight_mask[self.label == weight_id] = self.weight_dic[weight_id] if self.has_ignore_label: self.weight_mask[self.label == self.ignore_label] = 0 # num_total = 15422668800 # empty_num = 3679002314 # road_num = 10565335603 # ped_num = 99066996 # car_num = 995347874 #self.label[self.label == 3] = 2 # w_empty = float((num_total-empty_num)/num_total) # w_road = float((num_total-road_num)/num_total) # w_ped = float((num_total-ped_num)/num_total) # w_car = float((num_total-car_num)/num_total) # print(w_empty) # print(w_road) # print(w_ped) # print(w_car) # empty:0.3 # road:0.25 self.weight_mask[self.label == 0] = 0.3 self.weight_mask[self.label == 1] = 0.25 self.weight_mask[self.label == 3] = 0.5 # self.weight_mask[self.label == 2] = w_ped # self.weight_mask[self.label == 4] = w_car compute_count = self.weight_mask[self.weight_mask != 0].size ## nomalize mask self.weight_mask = cp.divide(self.weight_mask, cp.divide(cp.sum(self.weight_mask), compute_count)) ## compute loss prob_compute_matrix = copy.deepcopy(self.softmax[self.index_0,self.label,self.index_2,self.index_3]) prob_compute_matrix[prob_compute_matrix < (1e-10)] = 1e-10 loss = - cp.divide(cp.sum(cp.multiply(cp.log(prob_compute_matrix),self.weight_mask)),compute_count) loss = cp.asnumpy(loss) top[0].data[...] = loss
def fft_zdist(Q, S, epsilon, alignment=10000, Kahan=0):
"""
Rolling mean- and amplitude-adjusted Euclidean Distance using FFT to run in loglinear time
Equation exploiting cross-correlation (Fourier) theorem:
d[k] = sum_i (f(Q[i]) - f(S[i+k]))**2
= sum_i (f(Q[i])**2 - 2*f(Q[i])*f(S[i+k]) + f(S[i+k])**2)
= sum_i (f(Q[i])**2 - 2*f(Q[i])*(S[i+k]-mu[k]) + (S[i+k]-mu[k])**2)
= sum_i (f(Q[i])**2 - 2*f(Q[i])*S[i+k] + 2*f(Q[i])*mu[k] + (S[i+k]-mu[k])**2)
= sum_i (f(Q[i])**2 - 2*f(Q[i])*S[i+k] + (S[i+k]-mu[k])**2)
since sum_i f(Q[i]) = 0 by definition
= sum_i (f(Q[i])**2 - 2*f(Q[i])*S[i+k] + S[i+k]**2 - 2*S[i+k]*mu[k] + mu[k]**2)
= sum_i (f(Q[i])**2 - 2*f(Q[i])*S[i+k] + S[i+k]**2 - 2*|Q|*mu[k]*mu[k] + mu[k]**2)
= sum_i f(Q[i])**2 - 2*correlation(k) + Y[k] - 2*X[k]**2/|Q| + X[k]**2/|Q|
= sum_i f(Q[i])**2 - 2*correlation(k) + Y[k] - X[k]**2/|Q|
= sum_i f(Q[i])**2 - 2*correlation(k) + |Q|*variance[k]
Arguments:
-------
Q: cupy.core.core.ndarray
the input query of length m to be aligned
S: cupy.core.core.ndarray
the input stream of length n>=m to be scanned
epsilon: float
non-negative number for regularizing zero stdev
Kahan: int
non-negative number of Kahan summation adjustment rounds
Returns
-------
cupy.core.core.ndarray
the computed distance array of length n-m+1
"""
assert(epsilon > 0)
m, Q = len(Q), znorm(Q, epsilon)
n = (len(S)+alignment-1)//alignment*alignment
iS = cp.zeros(n, dtype=S.dtype)
iS[:len(S)] = S
delta = n-len(S)
X, Y = cumsum(iS, Kahan), cumsum(iS**2, Kahan)
X = X[+m:]-X[:-m]
Y = Y[+m:]-Y[:-m]
Z = cp.sqrt(cp.maximum(Y/m-cp.square(X/m), 0))
E = cp.zeros(n, dtype=Q.dtype)
E[:m] = Q
R = cp.fft.irfft(cp.fft.rfft(E).conj()*cp.fft.rfft(iS), n=n)
F = cp.where(Z > 0 , 2*(m-R[:-m+1]/Z), m*cp.ones_like(Z))
return F[:len(S)-m+1]
def backward(self, dout=1.0):
dout *= np.ones_like(self.y)
dout[self.mask_border == False] /= self.t[self.mask_border == False]
error = (self.y - self.t) / self.t
mask = (error == 0.0)
error[mask] += 1e-7
dout *= error / np.abs(error)
dout[self.mask_border == True] = (
self.y[self.mask_border == True] - self.t[self.mask_border == True]
) / self.t[self.mask_border == True]**2
dout /= float(self.y.size)
return dout
def __call__(self,
data,
iters=5,
width=3.0,
weights=None,
mask=None,
keepdims=False):
data = cp.asarray(data, dtype=self.dtype)
filt = cp.ones_like(data)
if mask is not None:
mask = cp.asarray(mask, dtype=self.dtype)
elementwise_not(cp.broadcast_to(mask, data.shape), filt)
if weights is not None:
weights = cp.asarray(weights, dtype=self.dtype)
try:
filt *= weights
except ValueError:
if isinstance(self.axis, int):
ndim = data.ndim
axis = self.axis % ndim
if weights.size == data.shape[axis]:
filt *= weights.reshape(-1 if i == axis else 1
for i in range(ndim))
else:
raise ValueError(
'length of weights must be same as'
' the length of data along specified axis.')
else:
raise ValueError(
'If weights and data are not broadcastable, '
'axis must be specified as int.')
checkfinite(data, filt, filt)
iterator = count() if (iters is None) else range(iters)
csum = check_sum(filt, axis=self.axis)
for _ in iterator:
self.updatefilt(data, filt, width)
tsum = check_sum(filt, axis=self.axis)
if all_equal(csum, tsum):
break
else:
csum = tsum
if self.rtnmask:
result = elementwise_not(filt, filt)
else:
result = self.reduce(data, filt, keepdims=keepdims)
return result
def forward(self, x):
# テンソル対応
self.original_x_shape = x.shape
x = x.reshape(x.shape[0], -1)
self.x = x
bW = cp.ones_like(self.W)
bW[self.W<0] = -1
self.bW = (bW).T
out = cp.dot(self.x, bW) #+ self.b
return out
def test_cross_correlate_masked_side_effects():
"""Masked normalized cross-correlation should not modify the inputs."""
shape1 = (2, 2, 2)
shape2 = (2, 2, 2)
arr1 = cp.zeros(shape1)
arr2 = cp.zeros(shape2)
# Trivial masks
m1 = cp.ones_like(arr1)
m2 = cp.ones_like(arr2)
# for arr in (arr1, arr2, m1, m2):
# arr.setflags(write=False)
arr1c, arr2c, m1c, m2c = [a.copy() for a in (arr1, arr2, m1, m2)]
cross_correlate_masked(arr1, arr2, m1, m2)
cp.testing.assert_array_equal(arr1, arr1c)
cp.testing.assert_array_equal(arr2, arr2c)
cp.testing.assert_array_equal(m1, m1c)
cp.testing.assert_array_equal(m2, m2c)
def test_atomic_and(self, dtype):
@jit.rawkernel()
def f(x, out):
tid = jit.blockDim.x * jit.blockIdx.x + jit.threadIdx.x
if tid < x.size:
jit.atomic_and(out, tid, x[tid])
x = cupy.arange(1024, dtype=dtype)
out = cupy.ones_like(x)
f((32, ), (32, ), (x, out))
expected = cupy.zeros_like(out)
expected[1::2] = 1
self._check(out, expected)
def test_vertical_mask_line(grad_func):
"""Vertical edge filters mask pixels surrounding input mask."""
_, hgrad = cp.mgrid[:1:11j, :1:11j] # horizontal gradient with spacing 0.1
hgrad[:, 5] = 1 # bad vertical line
mask = cp.ones_like(hgrad)
mask[:, 5] = 0 # mask bad line
expected = cp.zeros_like(hgrad)
expected[1:-1, 1:-1] = 0.2 # constant gradient for most of image,
expected[1:-1, 4:7] = 0 # but line and neighbors masked
result = grad_func(hgrad, mask)
assert_allclose(result, expected)
def precompute_cc_factors(ad, bd, radius, mode="constant"):
# factors = cp.zeros((5,) + ad.shape, dtype=ad.dtype)
factors = [None] * 5
sum_h = cp.ones((2 * radius + 1, ), dtype=ad.dtype)
h_tuple = (sum_h, ) * ad.ndim
kwargs = dict(mode=mode)
sum_a = convolve_separable(ad, h_tuple, **kwargs)
sum_b = convolve_separable(bd, h_tuple, **kwargs)
sum_ab = convolve_separable(ad * bd, h_tuple, **kwargs)
sum_aa = convolve_separable(ad * ad, h_tuple, **kwargs)
sum_bb = convolve_separable(bd * bd, h_tuple, **kwargs)
if mode != "constant":
cnt = (2 * radius + 1)**ad.ndim
else:
cnt = convolve_separable(cp.ones_like(ad), (sum_h, ) * ad.ndim,
**kwargs).astype(cp.int32)
if True:
factors[0] = cp.empty_like(ad)
factors[1] = cp.empty_like(ad)
factors[2] = cp.empty_like(ad)
factors[3] = cp.empty_like(ad)
factors[4] = cp.empty_like(ad)
_cc_precompute(
ad,
bd,
sum_a,
sum_b,
sum_ab,
sum_aa,
sum_bb,
cnt,
factors[0],
factors[1],
factors[2],
factors[3],
factors[4],
)
else:
a_mean = sum_a / cnt
b_mean = sum_b / cnt
factors[0] = ad - a_mean
factors[1] = bd - b_mean
factors[2] = sum_ab - b_mean * sum_a - a_mean * sum_b + sum_a * b_mean
factors[3] = sum_aa - (a_mean + a_mean) * sum_a + sum_a * a_mean
factors[4] = sum_bb - (b_mean + b_mean) * sum_b + sum_b * b_mean
return factors
def ones_like(array, stream=None):
"""Creates a one-filled cupy.ndarray object like the given array.
Args:
array (cupy.ndarray or numpy.ndarray): Base array.
stream (cupy.cuda.Stream): CUDA stream.
Returns:
cupy.ndarray: One-filled array.
"""
warnings.warn("chainer.cuda.ones_like is deprecated. Use cupy.ones_like instead.", DeprecationWarning)
check_cuda_available()
assert stream is None
if isinstance(array, cupy.ndarray):
return cupy.ones_like(array)
return cupy.ones(array.shape, dtype=array.dtype)
