smoothing_operator.py

# NIFTy
# Copyright (C) 2017  Theo Steininger
#
# Author: Theo Steininger
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

import numpy as np

import nifty.nifty_utilities as utilities
from nifty.operators.endomorphic_operator import EndomorphicOperator
from nifty.operators.fft_operator import FFTOperator
from nifty.operators.smoothing_operator import smooth_util as su
from d2o import STRATEGIES


class SmoothingOperator(EndomorphicOperator):
    # ---Overwritten properties and methods---
    def __init__(self, domain=(), sigma=0, log_distances=False):

        self._domain = self._parse_domain(domain)

        if len(self.domain) != 1:
            raise ValueError(
                'ERROR: SmoothOperator accepts only exactly one '
                'space as input domain.'
            )

        self.sigma = sigma
        self.log_distances = log_distances

        self._direct_smoothing_width = 3.

    def _inverse_times(self, x, spaces):
        return self._smoothing_helper(x, spaces, inverse=True)

    def _times(self, x, spaces):
        return self._smoothing_helper(x, spaces, inverse=False)

    # ---Mandatory properties and methods---
    @property
    def domain(self):
        return self._domain

    @property
    def self_adjoint(self):
        return True

    @property
    def unitary(self):
        return False

    # ---Added properties and methods---

    @property
    def sigma(self):
        return self._sigma

    @sigma.setter
    def sigma(self, sigma):
        self._sigma = np.float(sigma)

    @property
    def log_distances(self):
        return self._log_distances

    @log_distances.setter
    def log_distances(self, log_distances):
        self._log_distances = bool(log_distances)

    def _smoothing_helper(self, x, spaces, inverse):
        if self.sigma == 0:
            return x.copy()

        # the domain of the smoothing operator contains exactly one space.
        # Hence, if spaces is None, but we passed LinearOperator's
        # _check_input_compatibility, we know that x is also solely defined
        # on that space
        if spaces is None:
            spaces = (0,)
        else:
            spaces = utilities.cast_axis_to_tuple(spaces, len(x.domain))

        try:
            result = self._fft_smoothing(x, spaces, inverse)
        except ValueError:
            result = self._direct_smoothing(x, spaces, inverse)
        return result

    def _fft_smoothing(self, x, spaces, inverse):
        Transformator = FFTOperator(x.domain[spaces[0]])

        # transform to the (global-)default codomain and perform all remaining
        # steps therein
        transformed_x = Transformator(x, spaces=spaces)
        codomain = transformed_x.domain[spaces[0]]
        coaxes = transformed_x.domain_axes[spaces[0]]

        # create the kernel using the knowledge of codomain about itself
        axes_local_distribution_strategy = \
            transformed_x.val.get_axes_local_distribution_strategy(axes=coaxes)

        kernel = codomain.get_distance_array(
            distribution_strategy=axes_local_distribution_strategy)

        if self.log_distances:
            kernel.apply_scalar_function(np.log, inplace=True)

        kernel.apply_scalar_function(
            codomain.get_fft_smoothing_kernel_function(self.sigma),
            inplace=True)

        # now, apply the kernel to transformed_x
        # this is done node-locally utilizing numpys reshaping in order to
        # apply the kernel to the correct axes
        local_transformed_x = transformed_x.val.get_local_data(copy=False)
        local_kernel = kernel.get_local_data(copy=False)

        reshaper = [transformed_x.shape[i] if i in coaxes else 1
                    for i in xrange(len(transformed_x.shape))]
        local_kernel = np.reshape(local_kernel, reshaper)

        # apply the kernel
        if inverse:
            local_transformed_x /= local_kernel
        else:
            local_transformed_x *= local_kernel

        transformed_x.val.set_local_data(local_transformed_x, copy=False)

        smoothed_x = Transformator.inverse_times(transformed_x, spaces=spaces)

        result = x.copy_empty()
        result.set_val(smoothed_x, copy=False)

        return result

    def _direct_smoothing(self, x, spaces, inverse):
        # infer affected axes
        # we rely on the knowledge, that `spaces` is a tuple with length 1.
        affected_axes = x.domain_axes[spaces[0]]

        if len(affected_axes) > 1:
            raise ValueError("By this implementation only one-dimensional "
                             "spaces can be smoothed directly.")

        affected_axis = affected_axes[0]

        distance_array = x.domain[spaces[0]].get_distance_array(
            distribution_strategy='not')
        distance_array = distance_array.get_local_data(copy=False)

        if self.log_distances:
            np.log(distance_array, out=distance_array)

        # collect the local data + ghost cells
        local_data_Q = False

        if x.distribution_strategy == 'not':
            local_data_Q = True
        elif x.distribution_strategy in STRATEGIES['slicing']:
            # infer the local start/end based on the slicing information of
            # x's d2o. Only gets non-trivial for axis==0.
            if 0 != affected_axis:
                local_data_Q = True
            else:
                start_index = x.val.distributor.local_start
                start_distance = distance_array[start_index]
                augmented_start_distance = \
                    (start_distance - self._direct_smoothing_width*self.sigma)
                augmented_start_index = \
                    np.searchsorted(distance_array, augmented_start_distance)
                true_start = start_index - augmented_start_index
                end_index = x.val.distributor.local_end
                end_distance = distance_array[end_index-1]
                augmented_end_distance = \
                    (end_distance + self._direct_smoothing_width*self.sigma)
                augmented_end_index = \
                    np.searchsorted(distance_array, augmented_end_distance)
                true_end = true_start + x.val.distributor.local_length
                augmented_slice = slice(augmented_start_index,
                                        augmented_end_index)

                augmented_data = x.val.get_data(augmented_slice,
                                                local_keys=True,
                                                copy=False)
                augmented_data = augmented_data.get_local_data(copy=False)

                augmented_distance_array = distance_array[augmented_slice]

        else:
            raise ValueError("Direct smoothing not implemented for given"
                             "distribution strategy.")

        if local_data_Q:
            # if the needed data resides on the nodes already, the necessary
            # are the same; no matter what the distribution strategy was.
            augmented_data = x.val.get_local_data(copy=False)
            augmented_distance_array = distance_array
            true_start = 0
            true_end = x.shape[affected_axis]

        # perform the convolution along the affected axes
        # currently only one axis is supported
        data_axis = affected_axes[0]
        local_result = self._direct_smoothing_single_axis(
                                                    augmented_data,
                                                    data_axis,
                                                    augmented_distance_array,
                                                    true_start,
                                                    true_end,
                                                    inverse)
        result = x.copy_empty()
        result.val.set_local_data(local_result, copy=False)
        return result

    def _direct_smoothing_single_axis(self, data, data_axis, distances,
                                      true_start, true_end, inverse):
        if inverse:
            true_sigma = 1. / self.sigma
        else:
            true_sigma = self.sigma

        if data.dtype is np.dtype('float32'):
            distances = distances.astype(np.float32, copy=False)
            smoothed_data = su.apply_along_axis_f(
                                  data_axis, data,
                                  startindex=true_start,
                                  endindex=true_end,
                                  distances=distances,
                                  smooth_length=true_sigma,
                                  smoothing_width=self._direct_smoothing_width)
        elif data.dtype is np.dtype('float64'):
            distances = distances.astype(np.float64, copy=False)
            smoothed_data = su.apply_along_axis(
                                  data_axis, data,
                                  startindex=true_start,
                                  endindex=true_end,
                                  distances=distances,
                                  smooth_length=true_sigma,
                                  smoothing_width=self._direct_smoothing_width)

        elif np.issubdtype(data.dtype, np.complexfloating):
            real = self._direct_smoothing_single_axis(data.real,
                                                      data_axis,
                                                      distances,
                                                      true_start,
                                                      true_end, inverse)
            imag = self._direct_smoothing_single_axis(data.imag,
                                                      data_axis,
                                                      distances,
                                                      true_start,
                                                      true_end, inverse)

            return real + 1j*imag

        else:
            raise TypeError("Dtype %s not supported" % str(data.dtype))

        return smoothed_data