def backward(self, obj):
theta = -1 * self.theta
if theta == 0:
return obj
else:
if not obj.is_cuda:
obj = obj.cuda()
theta *= np.pi / 180.0
alpha = 1.0 * np.tan(theta / 2.0)
beta = np.sin(-1.0 * theta)
shear_phase_1 = op.exp(
op.multiply_complex(op._j, self.coord_phase_1 * alpha))
shear_phase_2 = op.exp(
op.multiply_complex(op._j, self.coord_phase_2 * beta))
self.dim[self.axis] = self.slice_per_tile
self.obj_rotate = op.r2c(
torch.zeros([self.dim[0], self.dim[1], self.dim[2]],
dtype=self.dtype,
device=self.device))
for idx_start in range(0, obj.shape[self.axis],
self.slice_per_tile):
idx_end = np.min(
[obj.shape[self.axis], idx_start + self.slice_per_tile])
idx_slice = slice(idx_start, idx_end)
self.dim[self.axis] = int(idx_end - idx_start)
if self.axis == 0:
self.range_crop_y = slice(0, self.dim[self.axis])
obj[idx_slice, :, :] = self._rotate_3d(
obj[idx_slice, :, :], alpha, beta, shear_phase_1,
shear_phase_2)
elif self.axis == 1:
self.range_crop_x = slice(0, self.dim[self.axis])
obj[:, idx_slice, :] = self._rotate_3d(
obj[:, idx_slice, :], alpha, beta, shear_phase_1,
shear_phase_2)
elif self.axis == 2:
self.range_crop_z = slice(0, self.dim[self.axis])
obj[:, :,
idx_slice] = self._rotate_3d(obj[:, :,
idx_slice], alpha,
beta, shear_phase_1,
shear_phase_2)
self.obj_rotate[:] = 0.0
self.dim[self.axis] = obj.shape[self.axis]
self.obj_rotate = None
if not obj.is_cuda:
obj = obj.cpu()
return obj
def generate_angular_spectrum_kernel(shape, pixel_size, wavelength, \
numerical_aperture=None, flag_band_limited=True, \
dtype=torch.float32, device=torch.device('cuda')):
"""
Function that generates angular spectrum propagation kernel WITHOUT the distance
The angular spectrum has the following form:
p = exp(distance * kernel)
kernel = 1j * 2 * pi * sqrt((ri/wavelength)**2-x**2-y**2)
and this function generates the kernel only!
"""
assert len(shape) == 2, "pupil should be two dimensional!"
ky_lin, kx_lin = util.generate_grid_2d(shape,
pixel_size,
flag_fourier=True,
dtype=dtype,
device=device)
if flag_band_limited:
assert numerical_aperture is not None, "need to provide numerical aperture of the system!"
pupil_crop = op.r2c(
generate_hard_pupil(shape, pixel_size, numerical_aperture,
wavelength))
else:
pupil_crop = 1.0
prop_kernel = 2.0 * np.pi * pupil_crop * \
op.exponentiate(op.r2c((1./wavelength)**2 - kx_lin**2 - ky_lin**2), 0.5)
return op.multiply_complex(op._j, prop_kernel)
def backward(ctx, grad_output):
kernel, field = ctx.saved_tensors
defocus_list = ctx.defocus_list
grad_defocus_list = defocus_list.clone()
grad_output = torch.fft(grad_output.permute(2, 0, 1, 3), signal_ndim=2)
for defocus_idx in range(len(defocus_list)):
kernel_temp = op.exp(abs(defocus_list[defocus_idx]) * kernel)
kernel_temp = kernel_temp if defocus_list[
defocus_idx] < 0. else op.conj(kernel_temp)
grad_output[defocus_idx, ...] = op.multiply_complex(
grad_output[defocus_idx, ...], kernel_temp)
# adaptive_step_size = op.r2c(op.abs(field) / (1e-8 + (torch.max(op.abs(field)) * op.abs(field)**2)))
# grad_defocus_list[defocus_idx] = op.multiply_complex(op.multiply_complex(op.multiply_complex(grad_output[defocus_idx,...],adaptive_step_size), op.conj(field)), op.conj(kernel)).sum((0,1))[0]
grad_defocus_list[defocus_idx] = op.multiply_complex(
op.multiply_complex(grad_output[defocus_idx, ...],
op.conj(field)), op.conj(kernel)).sum(
(0, 1))[0]
grad_output = torch.ifft(grad_output,
signal_ndim=2).permute(1, 2, 0, 3)
temp_f = grad_output.sum(2)
return grad_output.sum(2), None, grad_defocus_list
def compute_prox(self, x): if self.parameter_list is not None: self.set_parameter(self.parameter_list[self.itr_count]) x_device = x.device x = x.to(device=self.device) if self.pure_real: x[...,0] = self._compute_prox_real(op.real(x), self.realProjector) x[...,1] = 0.0 elif self.pure_imag: x[...,0] = 0.0 x[...,1] = op.multiply_complex(op._j, op.r2c(self._compute_prox_real(op.imag(x), self.imagProjector))) elif self.pure_amplitude: x[...,0] = self._compute_prox_real(op.abs(x), self.realProjector) x[...,1] = 0.0 elif self.pure_phase: x = op.exp(op.multiply_complex(op._j, op.r2c(self._compute_prox_real(op.angle(x), self.realProjector)))) else: x[...,0] = self._compute_prox_real(op.real(x), self.realProjector) self.set_parameter(self.parameter / 1.0, self.maxitr) x[...,1] = self._compute_prox_real(op.imag(x), self.imagProjector) self.set_parameter(self.parameter * 1.0, self.maxitr) self.itr_count += 1 return x.to(x_device)
def _shift_stack_inplace(self, stack, shift_list):
for img_idx in range(stack.shape[2]):
y_shift = shift_list[0, img_idx]
x_shift = shift_list[1, img_idx]
kernel = op.exp(
op.multiply_complex(
op._j,
op.r2c(2 * np.pi *
(self.kx_lin * x_shift + self.ky_lin * y_shift))))
stack[..., img_idx] = op.convolve_kernel(op.r2c(stack[...,
img_idx]),
kernel,
n_dim=2)[..., 0]
return stack
def forward(ctx, field, kernel, defocus_list=[0.0]):
field = torch.fft(field, signal_ndim=2)
ctx.save_for_backward(kernel, field)
# ctx.save_for_backward(field)
ctx.defocus_list = defocus_list
field = field.unsqueeze(2).repeat(1, 1, len(defocus_list),
1).permute(2, 0, 1, 3)
for defocus_idx in range(len(defocus_list)):
kernel_temp = op.exp(abs(defocus_list[defocus_idx]) * kernel)
kernel_temp = kernel_temp if defocus_list[
defocus_idx] > 0. else op.conj(kernel_temp)
field[defocus_idx,
...] = op.multiply_complex(field[defocus_idx, ...],
kernel_temp)
field = torch.ifft(field, signal_ndim=2).permute(1, 2, 0, 3)
return field
def compute_prox(self, x): x = op.exp(op.multiply_complex(op._j, op.r2c(op.angle(x)))) return x
