def calc_ll(self,theta):
"""
Computes the ll per measurments.
"""
cpa_space = self.cpa_space
src = self.src
dst = self.dst
params_flow_int = self.params_flow_int
transformed = self.transformed
sigma_lm = self.sigma_lm
ll = self.ll
if src.shape[1]!=cpa_space.dim_domain:
raise ValueError(src.shape)
cpa_space.theta2Avees(theta)
cpa_space.update_pat()
cpa_space.calc_T_fwd(pts = src, out=transformed,
**params_flow_int)
if transformed.shape != src.shape:
raise ValueError(transformed.shape , src.shape)
if 1:
calc_err_per_sample(transformed.gpu,dst.gpu,self.err.gpu)
# print np.allclose(transformed.gpu.get()-dst.gpu.get(),
# self.err.gpu.get())
calc_ll_per_sample(self.err.gpu,sigma_lm ,ll.gpu)
# print np.allclose((-0.5/(sigma_lm**2)*
# (transformed.gpu.get()-dst.gpu.get())**2).sum(axis=1),
# ll.gpu.get())
# idx = (dst.cpu<0.5).nonzero()[0][-1]
# idx = 499
#
# weight_it(ll.gpu,np.int32(idx))
# ipshell('hi')
# 1/0
else:
raise Exception("BAD IDEA")
ll_by_der = self.ll_by_der
der_dt = 0.002
calc_err_by_der_per_sample(transformed.gpu,dst.gpu,self.err_by_der.gpu,
der_dt)
calc_ll_by_der_per_sample(self.err_by_der.gpu,.1 ,ll_by_der.gpu)
ll = ll_by_der
def calc_ll(self,theta):
"""
Returns the ll per measurments.
"""
cpa_space = self.cpa_space
src = self.src
dst = self.dst
params_flow_int = self.params_flow_int
transformed = self.transformed
sigma_lm = self.sigma_lm
ll = self.ll
if src.shape[1]!=cpa_space.dim_domain:
raise ValueError(src.shape)
# Avees = cpa_space.theta2Avees(theta)
# pat = self._pat
# pat.update(Avees=Avees)
# cpa_space.calc_T(pat,pts = src, mysign=1,out=transformed,
# **params_flow_int)
cpa_space.theta2Avees(theta)
cpa_space.update_pat()
cpa_space.calc_T_fwd(pts = src,out=transformed,
**params_flow_int)
if transformed.shape != src.shape:
raise ValueError(transformed.shape , src.shape)
calc_err_per_sample(transformed.gpu,dst.gpu,self.err.gpu)
# print np.allclose(transformed.gpu.get()-dst.gpu.get(),
# self.err.gpu.get())
calc_ll_per_sample(self.err.gpu,sigma_lm ,ll.gpu)
# print np.allclose((-0.5/(sigma_lm**2)*
# (transformed.gpu.get()-dst.gpu.get())**2).sum(axis=1),
# ll.gpu.get())
return ll
def calc_ll(self, theta):
"""
Returns the ll per measurments.
"""
cpa_space = self.cpa_space
src = self.src
dst = self.dst
params_flow_int = self.params_flow_int
transformed = self.transformed
sigma_lm = self.sigma_lm
ll = self.ll
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape)
# Avees = cpa_space.theta2Avees(theta)
# pat = self._pat
# pat.update(Avees=Avees)
# cpa_space.calc_T(pat,pts = src, mysign=1,out=transformed,
# **params_flow_int)
cpa_space.theta2Avees(theta)
cpa_space.update_pat()
cpa_space.calc_T_fwd(pts=src, out=transformed, **params_flow_int)
if transformed.shape != src.shape:
raise ValueError(transformed.shape, src.shape)
calc_err_per_sample(transformed.gpu, dst.gpu, self.err.gpu)
# print np.allclose(transformed.gpu.get()-dst.gpu.get(),
# self.err.gpu.get())
calc_ll_per_sample(self.err.gpu, sigma_lm, ll.gpu)
# print np.allclose((-0.5/(sigma_lm**2)*
# (transformed.gpu.get()-dst.gpu.get())**2).sum(axis=1),
# ll.gpu.get())
return ll
def calc_ll(self, theta):
"""
Computes the ll per measurments.
"""
cpa_space = self.cpa_space
src = self.src
signal = self.signal
params_flow_int = self.params_flow_int
transformed = self.transformed
sigma_signal = self.sigma_signal
ll = self.ll
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape)
cpa_space.theta2Avees(theta)
cpa_space.update_pat()
cpa_space.calc_T_inv(pts=src, out=transformed, **params_flow_int)
if transformed.shape != src.shape:
raise ValueError(transformed.shape, src.shape)
if 0:
# I think this is irrelevant here.
# (it was relevant for the 1d case)
scale_pts = src.shape[0] # Assuming we normalized src
# to [0,1],
# we need to scale it
# t0 [0,1,..,num_of_samples+1]
transformed.gpu *= scale_pts
if self.interp_type_for_ll == 'gpu_linear':
resampler(pts_gpu=transformed.gpu,
img_gpu=signal.src.gpu,
img_wrapped_gpu=signal.transformed.gpu)
else:
remap_fwd_opencv(pts_inv=transformed,
img=signal.src,
img_wrapped_fwd=signal.transformed,
interp_method=self.interp_type_for_ll)
# if signal.dst.shape[1]>1:
# raise NotImplementedError("Only 1D signals for now")
# plt.figure(17);
# signal.transformed.gpu2cpu()
# if not signal.transformed.cpu.any():
# raise ValueError('wtf')
# plt.subplot(221)
# plt.imshow(signal.src.cpu.reshape(256,-1))
# plt.subplot(222)
# plt.imshow(signal.transformed.cpu.reshape(256,-1))
#
# ipshell('hi')
# 2/0
calc_signal_err_per_sample(signal.transformed.gpu, signal.dst.gpu,
self.signal.err.gpu)
# print np.allclose(signal.transformed.gpu.get()-
# signal.dst.gpu.get(),
# self.signal.err.gpu.get())
# if self.dim_signal != 1:
# raise NotImplementedError
calc_ll_per_sample(ll=ll.gpu,
err=self.signal.err.gpu,
sigma=sigma_signal)
if 0:
res = -0.5 / (sigma_signal**2) * (
(signal.transformed.gpu - signal.dst.gpu)**2).get()
np.allclose(res.ravel(), ll.gpu.get())
ipshell('stop')
1 / 0
# 1/0
# ipshell('stop');
# if any(theta):
# 1/0
def calc_ll(self, theta):
"""
Computes the ll per measurments.
"""
cpa_space = self.cpa_space
src = self.src
signal = self.signal
params_flow_int = self.params_flow_int
transformed = self.transformed
sigma_signal = self.sigma_signal
ll = self.ll
if src.shape[1] != cpa_space.dim_domain:
raise ValueError(src.shape)
cpa_space.theta2Avees(theta)
cpa_space.update_pat()
cpa_space.calc_T_inv(pts=src, out=transformed, **params_flow_int)
if transformed.shape != src.shape:
raise ValueError(transformed.shape, src.shape)
if 0:
# I think this is irrelevant here.
# (it was relevant for the 1d case)
scale_pts = src.shape[0] # Assuming we normalized src
# to [0,1],
# we need to scale it
# t0 [0,1,..,num_of_samples+1]
transformed.gpu *= scale_pts
if self.interp_type_for_ll == "gpu_linear":
resampler(pts_gpu=transformed.gpu, img_gpu=signal.src.gpu, img_wrapped_gpu=signal.transformed.gpu)
else:
remap_fwd_opencv(
pts_inv=transformed,
img=signal.src,
img_wrapped_fwd=signal.transformed,
interp_method=self.interp_type_for_ll,
)
# if signal.dst.shape[1]>1:
# raise NotImplementedError("Only 1D signals for now")
# plt.figure(17);
# signal.transformed.gpu2cpu()
# if not signal.transformed.cpu.any():
# raise ValueError('wtf')
# plt.subplot(221)
# plt.imshow(signal.src.cpu.reshape(256,-1))
# plt.subplot(222)
# plt.imshow(signal.transformed.cpu.reshape(256,-1))
#
# ipshell('hi')
# 2/0
calc_signal_err_per_sample(signal.transformed.gpu, signal.dst.gpu, self.signal.err.gpu)
# print np.allclose(signal.transformed.gpu.get()-
# signal.dst.gpu.get(),
# self.signal.err.gpu.get())
# if self.dim_signal != 1:
# raise NotImplementedError
calc_ll_per_sample(ll=ll.gpu, err=self.signal.err.gpu, sigma=sigma_signal)
if 0:
res = -0.5 / (sigma_signal ** 2) * ((signal.transformed.gpu - signal.dst.gpu) ** 2).get()
np.allclose(res.ravel(), ll.gpu.get())
ipshell("stop")
1 / 0
