reflectorch 1.0.0__py3-none-any.whl

reflectorch/data_generation/priors/subprior_sampler.py

@@ -0,0 +1,377 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import lru_cache
+from typing import Tuple
+
+import torch
+from torch import Tensor
+
+from reflectorch.data_generation.utils import (
+    uniform_sampler,
+    logdist_sampler,
+    triangular_sampler,
+    get_slds_from_d_rhos,
+)
+
+from reflectorch.data_generation.priors.params import Params
+from reflectorch.data_generation.priors.no_constraints import (
+    BasicPriorSampler,
+    DEFAULT_ROUGHNESS_RANGE,
+    DEFAULT_THICKNESS_RANGE,
+    DEFAULT_SLD_RANGE,
+    DEFAULT_NUM_LAYERS,
+    DEFAULT_DEVICE,
+    DEFAULT_DTYPE,
+    DEFAULT_SCALED_RANGE,
+    DEFAULT_USE_DRHO,
+)
+
+
+class UniformSubPriorParams(Params):
+    """Parameters class for thicknesses, roughnesses and slds, together with their subprior bounds."""
+    __slots__ = ('thicknesses', 'roughnesses', 'slds', 'min_bounds', 'max_bounds')
+    PARAM_NAMES = __slots__
+
+    def __init__(self,
+                 thicknesses: Tensor,
+                 roughnesses: Tensor,
+                 slds: Tensor,
+                 min_bounds: Tensor,
+                 max_bounds: Tensor,
+                 ):
+        super().__init__(thicknesses, roughnesses, slds)
+        self.min_bounds = min_bounds
+        self.max_bounds = max_bounds
+
+    @staticmethod
+    def rearrange_context_from_params(
+            scaled_params: Tensor, context: Tensor, inference: bool = False, from_params: bool = False
+    ):
+        if inference:
+            if from_params:
+                num_params = scaled_params.shape[1] // 3
+                scaled_params = scaled_params[:, num_params:]
+            context = torch.cat([context, scaled_params], dim=-1)
+            return context
+
+        num_params = scaled_params.shape[1] // 3
+        assert num_params * 3 == scaled_params.shape[1]
+        scaled_params, bound_context = torch.split(scaled_params, [num_params, 2 * num_params], dim=-1)
+        context = torch.cat([context, bound_context], dim=-1)
+        return scaled_params, context
+
+    @staticmethod
+    def restore_params_from_context(scaled_params: Tensor, context: Tensor):
+        num_params = scaled_params.shape[-1]
+        scaled_bounds = context[:, -2 * num_params:]
+        scaled_params = torch.cat([scaled_params, scaled_bounds], dim=-1)
+        return scaled_params
+
+    @staticmethod
+    def input_context_split(t_params):
+        num_params = t_params.shape[1] // 3
+        return torch.split(t_params, [num_params, 2 * num_params])
+
+    def as_tensor(self, use_drho: bool = False, add_bounds: bool = True) -> Tensor:
+        t_list = [self.thicknesses, self.roughnesses]
+        if use_drho:
+            t_list.append(self.d_rhos)
+        else:
+            t_list.append(self.slds)
+        if add_bounds:
+            t_list += [self.min_bounds, self.max_bounds]
+        return torch.cat(t_list, -1)
+
+    @classmethod
+    def from_tensor(cls, params: Tensor):
+        layers_num = (params.shape[-1] - 6) // 9
+        num_params = 3 * layers_num + 2
+
+        thicknesses, roughnesses, slds, min_bounds, max_bounds = torch.split(
+            params,
+            [layers_num, layers_num + 1, layers_num + 1, num_params, num_params],
+            dim=-1
+        )
+
+        return cls(thicknesses, roughnesses, slds, min_bounds, max_bounds)
+
+    @property
+    def num_params(self) -> int:
+        return self.layers_num2size(self.max_layer_num)
+
+    @staticmethod
+    def size2layers_num(size: int) -> int:
+        return (size - 6) // 9
+
+    @staticmethod
+    def layers_num2size(layers_num: int) -> int:
+        return layers_num * 9 + 6
+
+    def scale_with_q(self, q_ratio: float):
+        super().scale_with_q(q_ratio)
+
+        layer_num = self.max_layer_num
+        scales = torch.tensor(
+            [1 / q_ratio] * (2 * layer_num + 1) + [q_ratio ** 2] * (layer_num + 1),
+            device=self.device, dtype=self.dtype
+        )
+
+        self.min_bounds *= scales
+        self.max_bounds *= scales
+
+
+class UniformSubPriorSampler(BasicPriorSampler):
+    """Prior sampler for thicknesses, roughnesses, slds and their subprior bounds
+
+    Args:
+        thickness_range (Tuple[float, float], optional): the range of the layer thicknesses. Defaults to DEFAULT_THICKNESS_RANGE.
+        roughness_range (Tuple[float, float], optional): the range of the interlayer roughnesses. Defaults to DEFAULT_ROUGHNESS_RANGE.
+        sld_range (Tuple[float, float], optional): the range of the layer SLDs. Defaults to DEFAULT_SLD_RANGE.
+        num_layers (int, optional): the number of layers. Defaults to DEFAULT_NUM_LAYERS.
+        use_drho (bool, optional): whether to use differences in SLD values between neighboring layers instead of the actual SLD values. Defaults to DEFAULT_USE_DRHO.
+        device (torch.device, optional): the Pytorch device. Defaults to DEFAULT_DEVICE.
+        dtype (torch.dtype, optional): the Pytorch data type. Defaults to DEFAULT_DTYPE.
+        scaled_range (Tuple[float, float], optional): the range for scaling the parameters. Defaults to DEFAULT_SCALED_RANGE.
+        scale_by_subpriors (bool, optional): if True the film parameters are scaled with respect to their subprior bounds. Defaults to False.
+        smaller_roughnesses (bool, optional): if True the sampled roughnesses are biased towards smaller values. Defaults to False.
+        logdist (bool, optional): if True the relative widths of the subprior intervals are sampled uniformly on a logarithmic scale instead of uniformly. Defaults to False.
+        relative_min_bound_width (float, optional): defines the interval [relative_min_bound_width, 1.0] from which the relative bound widths for each parameter are sampled. Defaults to 1e-2.
+    """
+    PARAM_CLS = UniformSubPriorParams
+
+    def __init__(self,
+                 thickness_range: Tuple[float, float] = DEFAULT_THICKNESS_RANGE,
+                 roughness_range: Tuple[float, float] = DEFAULT_ROUGHNESS_RANGE,
+                 sld_range: Tuple[float, float] = DEFAULT_SLD_RANGE,
+                 num_layers: int = DEFAULT_NUM_LAYERS,
+                 use_drho: bool = DEFAULT_USE_DRHO,
+                 device: torch.device = DEFAULT_DEVICE,
+                 dtype: torch.dtype = DEFAULT_DTYPE,
+                 scaled_range: Tuple[float, float] = DEFAULT_SCALED_RANGE,
+                 scale_by_subpriors: bool = False,
+                 smaller_roughnesses: bool = False,
+                 logdist: bool = False,
+                 relative_min_bound_width: float = 1e-2,
+                 ):
+        super().__init__(
+            thickness_range,
+            roughness_range,
+            sld_range,
+            num_layers,
+            use_drho,
+            device,
+            dtype,
+            scaled_range,
+        )
+
+        self.scale_by_subpriors = scale_by_subpriors
+        self.smaller_roughnesses = smaller_roughnesses
+        self.logdist = logdist
+        self.relative_min_bound_width = relative_min_bound_width
+
+    @property
+    def max_num_layers(self) -> int:
+        return self.num_layers
+
+    @property
+    def param_dim(self) -> int:
+        return self.max_num_layers * 3 + 2
+
+    @lru_cache()
+    def min_vector(self, layers_num, drho: bool = False):
+        min_vector = super().min_vector(layers_num, drho)
+        min_vector = torch.cat([min_vector, min_vector, min_vector], dim=0)
+        return min_vector
+
+    def scale_params(self, params: UniformSubPriorParams) -> Tensor:
+        scaled_params = super().scale_params(params)
+
+        if self.scale_by_subpriors:
+            params_t = params.as_tensor(use_drho=self.use_drho, add_bounds=False)
+            scaled_params[:, :self.param_dim] = self._scale(params_t, params.min_bounds, params.max_bounds)
+
+        return scaled_params
+
+    def restore_params(self, scaled_params: Tensor) -> Params:
+        if not self.scale_by_subpriors:
+            return super().restore_params(scaled_params)
+
+        scaled_params, scaled_min_bounds, scaled_max_bounds = torch.split(
+            scaled_params, [self.param_dim, self.param_dim, self.param_dim], dim=1
+        )
+
+        min_vector = super().min_vector(self.max_num_layers, self.use_drho)
+        max_vector = super().max_vector(self.max_num_layers, self.use_drho)
+
+        min_bounds = self._restore(scaled_min_bounds, min_vector, max_vector)
+        max_bounds = self._restore(scaled_max_bounds, min_vector, max_vector)
+
+        param_t = self._restore(scaled_params, min_bounds, max_bounds)
+        param_t = torch.cat([param_t, min_bounds, max_bounds], dim=-1)
+
+        params = UniformSubPriorParams.from_tensor(param_t)
+
+        if self.use_drho:
+            params.slds = get_slds_from_d_rhos(params.slds)
+        return params
+
+    @lru_cache()
+    def max_vector(self, layers_num, drho: bool = False):
+        max_vector = super().max_vector(layers_num, drho)
+        max_vector = torch.cat([max_vector, max_vector, max_vector], dim=0)
+        return max_vector
+
+    @lru_cache()
+    def delta_vector(self, layers_num, drho: bool = False):
+        delta_vector = self.max_vector(layers_num, drho) - self.min_vector(layers_num, drho)
+        delta_vector[delta_vector == 0.] = 1.
+        return delta_vector
+
+    def get_indices_within_bounds(self, params: UniformSubPriorParams) -> Tensor:
+        t_params = torch.cat([
+            params.thicknesses,
+            params.roughnesses,
+            params.slds
+        ], dim=-1)
+
+        indices = (
+                torch.all(t_params >= params.min_bounds, dim=-1) &
+                torch.all(t_params <= params.max_bounds, dim=-1)
+        )
+
+        return indices
+
+    def clamp_params(self, params: UniformSubPriorParams) -> UniformSubPriorParams:
+        params = UniformSubPriorParams.from_tensor(
+            torch.cat([
+                torch.clamp(
+                    params.as_tensor(add_bounds=False),
+                    params.min_bounds, params.max_bounds
+                ),
+                params.min_bounds, params.max_bounds
+            ], dim=1)
+        )
+        return params
+
+    def get_indices_within_domain(self, params: UniformSubPriorParams) -> Tensor:
+        return self.get_indices_within_bounds(params)
+
+    def sample(self, batch_size: int) -> UniformSubPriorParams:
+        min_bounds, max_bounds = self.sample_bounds(batch_size)
+
+        params = torch.rand(
+            *min_bounds.shape,
+            device=self.device,
+            dtype=self.dtype
+        ) * (max_bounds - min_bounds) + min_bounds
+
+        thicknesses, roughnesses, slds = torch.split(
+            params, [self.max_num_layers, self.max_num_layers + 1, self.max_num_layers + 1], dim=-1
+        )
+
+        params = UniformSubPriorParams(thicknesses, roughnesses, slds, min_bounds, max_bounds)
+
+        return params
+
+    def sample_bounds(self, batch_size: int):
+        min_vector, max_vector = (
+            super().min_vector(self.num_layers)[None],
+            super().max_vector(self.num_layers)[None]
+        )
+
+        delta_vector = max_vector - min_vector
+
+        if self.logdist:
+            widths_sampler_func = logdist_sampler
+        else:
+            widths_sampler_func = uniform_sampler
+
+        prior_widths = widths_sampler_func(
+            self.relative_min_bound_width, 1.,
+            batch_size, delta_vector.shape[1],
+            device=self.device, dtype=self.dtype
+        ) * delta_vector
+
+        prior_centers = uniform_sampler(
+            min_vector + prior_widths / 2, max_vector - prior_widths / 2,
+            *prior_widths.shape,
+            device=self.device, dtype=self.dtype
+        )
+
+        if self.smaller_roughnesses:
+            idx_min, idx_max = self.num_layers, self.num_layers * 2 + 1
+            prior_centers[:, idx_min:idx_max] = triangular_sampler(
+                min_vector[:, idx_min:idx_max] + prior_widths[:, idx_min:idx_max] / 2,
+                max_vector[:, idx_min:idx_max] - prior_widths[:, idx_min:idx_max] / 2,
+                batch_size, self.num_layers + 1,
+                device=self.device, dtype=self.dtype
+            )
+
+        min_bounds, max_bounds = prior_centers - prior_widths / 2, prior_centers + prior_widths / 2
+
+        return min_bounds, max_bounds
+
+    def scale_bounds(self, bounds: Tensor) -> Tensor:
+        layers_num = bounds.shape[-1] // 2
+
+        return self._scale(
+            bounds,
+            self.min_vector(layers_num, drho=self.use_drho).to(bounds),
+            self.max_vector(layers_num, drho=self.use_drho).to(bounds),
+        )
+
+
+class NarrowSldUniformSubPriorSampler(UniformSubPriorSampler):
+    """Prior sampler for thicknesses, roughnesses, slds and their subprior bounds. The subprior bound widths for SLDs are restricted to be lower than a specified value. """
+    def __init__(self,
+                 thickness_range: Tuple[float, float] = DEFAULT_THICKNESS_RANGE,
+                 roughness_range: Tuple[float, float] = DEFAULT_ROUGHNESS_RANGE,
+                 sld_range: Tuple[float, float] = DEFAULT_SLD_RANGE,
+                 num_layers: int = DEFAULT_NUM_LAYERS,
+                 use_drho: bool = DEFAULT_USE_DRHO,
+                 device: torch.device = DEFAULT_DEVICE,
+                 dtype: torch.dtype = DEFAULT_DTYPE,
+                 scaled_range: Tuple[float, float] = DEFAULT_SCALED_RANGE,
+                 scale_by_subpriors: bool = False,
+                 max_sld_prior_width: float = 10.,
+                 ):
+        super().__init__(
+            thickness_range,
+            roughness_range,
+            sld_range,
+            num_layers,
+            use_drho,
+            device,
+            dtype,
+            scaled_range,
+            scale_by_subpriors,
+        )
+
+        self.max_sld_prior_width = max_sld_prior_width
+
+    def sample_bounds(self, batch_size: int):
+        min_vector, max_vector = (
+            BasicPriorSampler.min_vector(self, self.num_layers),
+            BasicPriorSampler.max_vector(self, self.num_layers),
+        )
+
+        delta_vector = max_vector - min_vector
+        delta_vector[-self.num_layers:] = self.max_sld_prior_width
+
+        prior_widths = uniform_sampler(
+            delta_vector * self.relative_min_bound_width, delta_vector,
+            batch_size, min_vector.shape[0],
+            device=self.device, dtype=self.dtype
+        )
+
+        prior_centers = uniform_sampler(
+            min_vector + prior_widths / 2, max_vector - prior_widths / 2,
+            *prior_widths.shape,
+            device=self.device, dtype=self.dtype
+        )
+
+        return prior_centers - prior_widths / 2, prior_centers + prior_widths / 2

reflectorch/data_generation/priors/utils.py

@@ -0,0 +1,124 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Tuple
+
+import torch
+from torch import Tensor
+
+from reflectorch.data_generation.utils import (
+    get_d_rhos,
+    uniform_sampler,
+)
+
+
+def get_max_allowed_roughness(thicknesses: Tensor, mask: Tensor = None, coef: float = 0.5):
+    """gets the maximum allowed interlayer roughnesses such that they do not exceed a fraction of the thickness of either layers meeting at that interface"""
+    batch_size, layers_num = thicknesses.shape
+    max_roughness = torch.ones(
+        batch_size, layers_num + 1, device=thicknesses.device, dtype=thicknesses.dtype
+    ) * float('inf')
+
+    boundary = thicknesses * coef
+    if mask is not None:
+        boundary[get_thickness_mask_from_sld_mask(mask)] = float('inf')
+
+    max_roughness[:, :-1] = boundary
+    max_roughness[:, 1:] = torch.minimum(max_roughness[:, 1:], boundary)
+    return max_roughness
+
+
+def get_allowed_contrast_indices(slds: Tensor, min_contrast: float, mask: Tensor = None) -> Tensor:
+    d_rhos = get_d_rhos(slds)
+    indices = d_rhos.abs() >= min_contrast
+    if mask is not None:
+        indices = indices | mask
+    indices = torch.all(indices, -1)
+    return indices
+
+
+def params_within_bounds(params_t: Tensor, min_t: Tensor, max_t: Tensor, mask: Tensor = None) -> Tensor:
+    indices = (params_t >= min_t[None]) & (params_t <= max_t[None])
+    if mask is not None:
+        indices = indices | mask
+    indices = torch.all(indices, -1)
+    return indices
+
+
+def get_allowed_roughness_indices(thicknesses: Tensor, roughnesses: Tensor, mask: Tensor = None) -> Tensor:
+    max_roughness = get_max_allowed_roughness(thicknesses, mask)
+    indices = roughnesses <= max_roughness
+    if mask is not None:
+        indices = indices | mask
+    indices = torch.all(indices, -1)
+    return indices
+
+
+def get_thickness_mask_from_sld_mask(mask: Tensor):
+    return mask[:, :-1]
+
+
+def generate_roughnesses(thicknesses: Tensor, roughness_range: Tuple[float, float], mask: Tensor = None):
+    batch_size, layers_num = thicknesses.shape
+    max_roughness = get_max_allowed_roughness(thicknesses, mask)
+    max_roughness = torch.clamp_(max_roughness, max=roughness_range[1])
+
+    roughnesses = uniform_sampler(
+        roughness_range[0], max_roughness, batch_size, layers_num + 1,
+        device=thicknesses.device, dtype=thicknesses.dtype
+    )
+
+    if mask is not None:
+        roughnesses[mask] = 0.
+
+    return roughnesses
+
+
+def generate_thicknesses(
+        thickness_range: Tuple[float, float],
+        batch_size: int,
+        layers_num: int,
+        device: torch.device,
+        dtype: torch.dtype,
+        mask: Tensor = None
+):
+    thicknesses = uniform_sampler(
+        *thickness_range, batch_size, layers_num, device=device, dtype=dtype
+    )
+    if mask is not None:
+        thicknesses[get_thickness_mask_from_sld_mask(mask)] = 0.
+    return thicknesses
+
+
+def generate_slds_with_min_contrast(
+        sld_range: Tuple[float, float],
+        batch_size: int,
+        layers_num: int,
+        min_contrast: float,
+        device: torch.device,
+        dtype: torch.dtype,
+        mask: Tensor = None,
+        *,
+        _depth: int = 0
+):
+    # rejection sampling
+    slds = uniform_sampler(
+        *sld_range, batch_size, layers_num + 1, device=device, dtype=dtype
+    )
+
+    if mask is not None:
+        slds[mask] = 0.
+
+    rejected_indices = ~get_allowed_contrast_indices(slds, min_contrast, mask)
+    rejected_num = rejected_indices.sum(0).item()
+
+    if rejected_num:
+        if mask is not None:
+            mask = mask[rejected_indices]
+        slds[rejected_indices] = generate_slds_with_min_contrast(
+            sld_range, rejected_num, layers_num, min_contrast, device, dtype, mask, _depth=_depth+1
+        )
+    return slds

reflectorch/data_generation/process_data.py

@@ -0,0 +1,47 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any
+
+__all__ = [
+    "ProcessData",
+    "ProcessPipeline",
+]
+
+
+class ProcessData(object):
+    def __add__(self, other):
+        if isinstance(other, ProcessData):
+            return ProcessPipeline(self, other)
+
+    def apply(self, args: Any, context: dict = None):
+        return args
+
+    def __call__(self, args: Any, context: dict = None):
+        return self.apply(args, context)
+
+    def __repr__(self):
+        return f'{self.__class__.__name__}()'
+
+
+class ProcessPipeline(ProcessData):
+    def __init__(self, *processes):
+        self._processes = list(processes)
+
+    def apply(self, args: Any, context: dict = None):
+        for process in self._processes:
+            args = process(args, context)
+        return args
+
+    def __add__(self, other):
+        if isinstance(other, ProcessPipeline):
+            return ProcessPipeline(*self._processes, *other._processes)
+        elif isinstance(other, ProcessData):
+            return ProcessPipeline(*self._processes, other)
+
+    def __repr__(self):
+        processes = ", ".join(repr(p) for p in self._processes)
+        return f'ProcessPipeline({processes})'