def run_GN_iter(self, num_cg_iter):
"""Runs a single GN iteration."""
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Create copy with graph detached
self.g = self.f0.vdetach()
self.g.requires_grad(True)
# Get df/dx^t @ f0
self.dfdxt_g = TensorList(
torch.autograd.grad(self.f0, self.x, self.g, create_graph=True))
# Get the right hand side
self.b = -self.dfdxt_g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
def __init__(self, training_samples: TensorList, y: TensorList,
filter_reg: torch.Tensor, sample_weights: TensorList,
response_activation):
self.training_samples = training_samples.variable()
self.y = y.variable()
self.filter_reg = filter_reg
self.sample_weights = sample_weights
self.response_activation = response_activation
def get_attribute(self, name: str, ignore_missing: bool = False):
if ignore_missing:
return TensorList([
getattr(f, name) for f in self.features
if self._return_feature(f) and hasattr(f, name)
])
else:
return TensorList([
getattr(f, name, None) for f in self.features
if self._return_feature(f)
])
def run(self, num_iter, dummy=None):
if num_iter == 0:
return
for i in range(num_iter):
self.x.requires_grad(True)
# Evaluate function at current estimate
loss = self.problem(self.x)
# Compute grad
grad = TensorList(torch.autograd.grad(loss, self.x))
# Update direction
if self.dir is None:
self.dir = grad
else:
self.dir = grad + self.momentum * self.dir
self.x.detach()
self.x -= self.step_legnth * self.dir
self.x.detach()
self.clear_temp()
def A(self, x):
dfdx_x = torch.autograd.grad(self.dfdxt_g,
self.g,
x,
retain_graph=True)
return TensorList(
torch.autograd.grad(self.f0, self.x, dfdx_x, retain_graph=True))
def evaluate_CG_iteration(self, delta_x):
if self.analyze_convergence:
x = (self.x + delta_x).detach()
x.requires_grad_(True)
# compute loss and gradient
loss = self.problem(x)
grad = TensorList(torch.autograd.grad(loss, x))
# store in the vectors
self.losses = torch.cat(
(self.losses, loss.detach().cpu().view(-1)))
self.gradient_mags = torch.cat(
(self.gradient_mags,
sum(grad.view(-1)
@ grad.view(-1)).cpu().sqrt().detach().view(-1)))
def __init__(self,
problem: L2Problem,
variable: TensorList,
cg_eps=0.0,
fletcher_reeves=True,
standard_alpha=True,
direction_forget_factor=0,
debug=False,
plotting=False,
fig_num=(10, 11)):
super().__init__(fletcher_reeves, standard_alpha,
direction_forget_factor, debug or plotting)
self.problem = problem
self.x = variable.variable()
self.plotting = plotting
self.fig_num = fig_num
self.cg_eps = cg_eps
self.f0 = None
self.g = None
self.dfdxt_g = None
self.residuals = torch.zeros(0)
self.losses = torch.zeros(0)
def get_fparams(self, name: str = None):
if name is None:
return [
f.fparams for f in self.features if self._return_feature(f)
]
return TensorList([
getattr(f.fparams, name) for f in self.features
if self._return_feature(f)
]).unroll()
def run_newton_iter(self, num_cg_iter):
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Gradient of loss
self.g = TensorList(
torch.autograd.grad(self.f0, self.x, create_graph=True))
# Get the right hand side
self.b = -self.g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
def extract_patches(self, im_patches):
"""Extract features.
args:
im: Image.
"""
# Compute features
# print (im_patches.shape)
feature_map = TensorList(
[f.get_feature(im_patches) for f in self.features]).unroll()
return feature_map
def __init__(self, training_samples: TensorList, y: TensorList, filter_reg: torch.Tensor, projection_reg, params, sample_weights: TensorList,
projection_activation, response_activation):
self.training_samples = training_samples
self.y = y.variable()
self.filter_reg = filter_reg
self.sample_weights = sample_weights
self.params = params
self.projection_reg = projection_reg
self.projection_activation = projection_activation
self.response_activation = response_activation
self.diag_M = self.filter_reg.concat(projection_reg)
def run(self, num_cg_iter):
"""Run the oprimizer with the provided number of iterations."""
if num_cg_iter == 0:
return
lossvec = None
if self.debug:
lossvec = torch.zeros(2)
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Create copy with graph detached
self.g = self.f0.vdetach()
if self.debug:
lossvec[0] = self.problem.ip_output(self.g, self.g)
self.g.requires_grad(True)
# Get df/dx^t @ f0
self.dfdxt_g = TensorList(
torch.autograd.grad(self.f0, self.x, self.g, create_graph=True))
# Get the right hand side
self.b = -self.dfdxt_g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
self.x.vdetach_()
self.clear_temp()
def extract(self, im, pos, scales, image_sz):
if isinstance(scales, (int, float)):
scales = [scales]
# Get image patches
im_patches = torch.cat(
[sample_patch(im, pos, s * image_sz, image_sz) for s in scales])
# Compute features
feature_map = torch.cat(TensorList(
[f.get_feature(im_patches) for f in self.features]).unroll(),
dim=1)
return feature_map
def extract(self, im, pos, scales, image_sz):
"""Extract features.
args:
im: Image.
pos: Center position for extraction.
scales: Image scales to extract features from.
image_sz: Size to resize the image samples to before extraction.
"""
if isinstance(scales, (int, float)):
scales = [scales]
# Get image patches
im_patches = torch.cat(
[sample_patch(im, pos, s * image_sz, image_sz) for s in scales])
# Compute features
feature_map = TensorList(
[f.get_feature(im_patches) for f in self.features]).unroll()
return feature_map
def extract_transformed(self, im, pos, scale, image_sz, transforms):
"""Extract features from a set of transformed image samples.
args:
im: Image.
pos: Center position for extraction.
scale: Image scale to extract features from.
image_sz: Size to resize the image samples to before extraction.
transforms: A set of image transforms to apply.
"""
# Get image patche
im_patch = sample_patch(im, pos, scale * image_sz, image_sz).data
# Apply transforms
im_patches = torch.cat([T(im_patch).data for T in transforms])
# Compute features
feature_map = TensorList(
[f.get_feature(im_patches) for f in self.features]).unroll()
return feature_map
def A(self, x):
return TensorList(
torch.autograd.grad(self.g, self.x, x,
retain_graph=True)) + self.hessian_reg * x
def stride(self):
return torch.Tensor(
TensorList([
f.stride() for f in self.features if self._return_feature(f)
]).unroll())
class NewtonCG(ConjugateGradientBase):
"""Newton with Conjugate Gradient. Handels general minimization problems."""
def __init__(self,
problem: MinimizationProblem,
variable: TensorList,
init_hessian_reg=0.0,
hessian_reg_factor=1.0,
cg_eps=0.0,
fletcher_reeves=True,
standard_alpha=True,
direction_forget_factor=0,
debug=False,
analyze=False,
plotting=False,
fig_num=(10, 11, 12)):
super().__init__(fletcher_reeves, standard_alpha,
direction_forget_factor, debug or analyze or plotting)
self.problem = problem
self.x = variable
self.analyze_convergence = analyze
self.plotting = plotting
self.fig_num = fig_num
self.hessian_reg = init_hessian_reg
self.hessian_reg_factor = hessian_reg_factor
self.cg_eps = cg_eps
self.f0 = None
self.g = None
self.residuals = torch.zeros(0)
self.losses = torch.zeros(0)
self.gradient_mags = torch.zeros(0)
def clear_temp(self):
self.f0 = None
self.g = None
def run(self, num_cg_iter, num_newton_iter=None):
if isinstance(num_cg_iter, int):
if num_cg_iter == 0:
return
if num_newton_iter is None:
num_newton_iter = 1
num_cg_iter = [num_cg_iter] * num_newton_iter
num_newton_iter = len(num_cg_iter)
if num_newton_iter == 0:
return
if self.analyze_convergence:
self.evaluate_CG_iteration(0)
for cg_iter in num_cg_iter:
self.run_newton_iter(cg_iter)
self.hessian_reg *= self.hessian_reg_factor
self.x.vdetach_()
self.clear_temp()
return self.losses, self.residuals
def run_newton_iter(self, num_cg_iter):
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Gradient of loss
self.g = TensorList(
torch.autograd.grad(self.f0, self.x, create_graph=True))
# Get the right hand side
self.b = -self.g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
def A(self, x):
return TensorList(
torch.autograd.grad(self.g, self.x, x,
retain_graph=True)) + self.hessian_reg * x
def ip(self, a, b):
# Implements the inner product
return self.problem.ip_input(a, b)
def M1(self, x):
return self.problem.M1(x)
def M2(self, x):
return self.problem.M2(x)
def evaluate_CG_iteration(self, delta_x):
if self.analyze_convergence:
x = (self.x + delta_x).detach()
x.requires_grad_(True)
# compute loss and gradient
loss = self.problem(x)
grad = TensorList(torch.autograd.grad(loss, x))
# store in the vectors
self.losses = torch.cat(
(self.losses, loss.detach().cpu().view(-1)))
self.gradient_mags = torch.cat(
(self.gradient_mags,
sum(grad.view(-1)
@ grad.view(-1)).cpu().sqrt().detach().view(-1)))
def size(self, input_sz):
return TensorList([
f.size(input_sz) for f in self.features if self._return_feature(f)
]).unroll()
def dim(self):
return TensorList([
f.dim() for f in self.features if self._return_feature(f)
]).unroll()
class GaussNewtonCG(ConjugateGradientBase):
"""Gauss-Newton with Conjugate Gradient optimizer."""
def __init__(self,
problem: L2Problem,
variable: TensorList,
cg_eps=0.0,
fletcher_reeves=True,
standard_alpha=True,
direction_forget_factor=0,
debug=False,
analyze=False,
plotting=False,
fig_num=(10, 11, 12)):
super().__init__(fletcher_reeves, standard_alpha,
direction_forget_factor, debug or analyze or plotting)
self.problem = problem
self.x = variable
self.analyze_convergence = analyze
self.plotting = plotting
self.fig_num = fig_num
self.cg_eps = cg_eps
self.f0 = None
self.g = None
self.dfdxt_g = None
self.residuals = torch.zeros(0)
self.losses = torch.zeros(0)
self.gradient_mags = torch.zeros(0)
def clear_temp(self):
self.f0 = None
self.g = None
self.dfdxt_g = None
def run_GN(self, *args, **kwargs):
return self.run(*args, **kwargs)
def run(self, num_cg_iter, num_gn_iter=None):
"""Run the optimizer.
args:
num_cg_iter: Number of CG iterations per GN iter. If list, then each entry specifies number of CG iterations
and number of GN iterations is given by the length of the list.
num_gn_iter: Number of GN iterations. Shall only be given if num_cg_iter is an integer.
"""
if isinstance(num_cg_iter, int):
if num_gn_iter is None:
raise ValueError(
'Must specify number of GN iter if CG iter is constant')
num_cg_iter = [num_cg_iter] * num_gn_iter
num_gn_iter = len(num_cg_iter)
if num_gn_iter == 0:
return
if self.analyze_convergence:
self.evaluate_CG_iteration(0)
# Outer loop for running the GN iterations.
for cg_iter in num_cg_iter:
self.run_GN_iter(cg_iter)
self.x.vdetach_()
self.clear_temp()
return self.losses, self.residuals
def run_GN_iter(self, num_cg_iter):
"""Runs a single GN iteration."""
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Create copy with graph detached
self.g = self.f0.vdetach()
self.g.requires_grad(True)
# Get df/dx^t @ f0
self.dfdxt_g = TensorList(
torch.autograd.grad(self.f0, self.x, self.g, create_graph=True))
# Get the right hand side
self.b = -self.dfdxt_g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
def A(self, x):
dfdx_x = torch.autograd.grad(self.dfdxt_g,
self.g,
x,
retain_graph=True)
return TensorList(
torch.autograd.grad(self.f0, self.x, dfdx_x, retain_graph=True))
def ip(self, a, b):
return self.problem.ip_input(a, b)
def M1(self, x):
return self.problem.M1(x)
def M2(self, x):
return self.problem.M2(x)
def evaluate_CG_iteration(self, delta_x):
if self.analyze_convergence:
x = (self.x + delta_x).detach()
x.requires_grad_(True)
# compute loss and gradient
f = self.problem(x)
loss = self.problem.ip_output(f, f)
grad = TensorList(torch.autograd.grad(loss, x))
# store in the vectors
self.losses = torch.cat(
(self.losses, loss.detach().cpu().view(-1)))
self.gradient_mags = torch.cat(
(self.gradient_mags,
sum(grad.view(-1)
@ grad.view(-1)).cpu().sqrt().detach().view(-1)))
class ConjugateGradient(ConjugateGradientBase):
"""Conjugate Gradient optimizer, performing single linearization of the residuals in the start."""
def __init__(self,
problem: L2Problem,
variable: TensorList,
cg_eps=0.0,
fletcher_reeves=True,
standard_alpha=True,
direction_forget_factor=0,
debug=False,
plotting=False,
fig_num=(10, 11)):
super().__init__(fletcher_reeves, standard_alpha,
direction_forget_factor, debug or plotting)
self.problem = problem
self.x = variable.variable()
self.plotting = plotting
self.fig_num = fig_num
self.cg_eps = cg_eps
self.f0 = None
self.g = None
self.dfdxt_g = None
self.residuals = torch.zeros(0)
self.losses = torch.zeros(0)
def clear_temp(self):
self.f0 = None
self.g = None
self.dfdxt_g = None
def run(self, num_cg_iter):
"""Run the oprimizer with the provided number of iterations."""
if num_cg_iter == 0:
return
lossvec = None
if self.debug:
lossvec = torch.zeros(2)
self.x.requires_grad(True)
# Evaluate function at current estimate
self.f0 = self.problem(self.x)
# Create copy with graph detached
self.g = self.f0.vdetach()
if self.debug:
lossvec[0] = self.problem.ip_output(self.g, self.g)
self.g.requires_grad(True)
# Get df/dx^t @ f0
self.dfdxt_g = TensorList(
torch.autograd.grad(self.f0, self.x, self.g, create_graph=True))
# Get the right hand side
self.b = -self.dfdxt_g.vdetach()
# Run CG
delta_x, res = self.run_CG(num_cg_iter, eps=self.cg_eps)
self.x.vdetach_()
self.x.plus_(delta_x)
self.x.vdetach_()
self.clear_temp()
def A(self, x):
dfdx_x = torch.autograd.grad(self.dfdxt_g,
self.g,
x,
retain_graph=True)
return TensorList(
torch.autograd.grad(self.f0, self.x, dfdx_x, retain_graph=True))
def ip(self, a, b):
return self.problem.ip_input(a, b)
def M1(self, x):
return self.problem.M1(x)
def M2(self, x):
return self.problem.M2(x)
def size(self, im_sz):
if self.output_size is None:
return TensorList([im_sz / s for s in self.stride()])
if isinstance(im_sz, torch.Tensor):
return TensorList([im_sz / s if sz is None else torch.Tensor([sz[0], sz[1]]) for sz, s in zip(self.output_size, self.stride())])