Source code for pymepix.processing.logic.centroid_calculator

# This file is part of Pymepix
#
# In all scientific work using Pymepix, please reference it as
#
# A. F. Al-Refaie, M. Johny, J. Correa, D. Pennicard, P. Svihra, A. Nomerotski, S. Trippel, and J. Küpper:
# "PymePix: a python library for SPIDR readout of Timepix3", J. Inst. 14, P10003 (2019)
# https://doi.org/10.1088/1748-0221/14/10/P10003
# https://arxiv.org/abs/1905.07999
#
# Pymepix 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 <https://www.gnu.org/licenses/>.
import multiprocessing as mp


import numpy as np
import scipy.ndimage as nd
from joblib import Parallel, delayed
from sklearn.cluster import DBSCAN

from pymepix.clustering.cluster_stream import ClusterStream
from pymepix.processing.logic.processing_step import ProcessingStep


[docs] def perform_clustering_dbscan(shot, x, y, tof, _tof_scale, epsilon, min_samples): """The clustering with DBSCAN, which is performed in this function is dependent on the order of the data in rare cases. Therefore, reordering in any means can lead to slightly changed results, which should not be an issue. Martin Ester, Hans-Peter Kriegel, Jiirg Sander, Xiaowei Xu: A Density Based Algorithm for Discovering Clusters [p. 229-230] (https://www.aaai.org/Papers/KDD/1996/KDD96-037.pdf) A more specific explaination can be found here: https://stats.stackexchange.com/questions/306829/why-is-dbscan-deterministic""" if x.size >= 0: X = np.column_stack((shot * epsilon * 1_000, x, y, tof * _tof_scale)) dist = DBSCAN( eps=epsilon, min_samples=min_samples, metric="euclidean", n_jobs=1, ).fit(X) return dist.labels_ + 1 return None
[docs] def calculate_centroids_properties(shot, x, y, tof, tot, labels, _cent_timewalk_lut): """ Calculates the properties of the centroids from labeled data points. ATTENTION! The order of the points can have an impact on the result due to errors in the floating point arithmetics. Very simple example: arr = np.random.random(100) arr.sum() - np.sort(arr).sum() This example shows that there is a very small difference between the two sums. The inaccuracy of floating point arithmetics can depend on the order of the values. Strongly simplified (3.2 + 3.4) + 2.7 and 3.2 + (3.4 + 2.7) can be unequal for floating point numbers. Therefore there is no guarantee for strictly equal results. Even after sorting. The error we observed can be about 10^-22 nano seconds. Currently this is issue exists only for the TOF-column as the other columns are integer-based values. """ label_index, cluster_size = np.unique(labels, return_counts=True) tot_max = np.array(nd.maximum_position(tot, labels=labels, index=label_index)).flatten() tot_sum = nd.sum(tot, labels=labels, index=label_index) tot_mean = nd.mean(tot, labels=labels, index=label_index) cluster_x = np.array(nd.sum(x * tot, labels=labels, index=label_index) / tot_sum).flatten() cluster_y = np.array(nd.sum(y * tot, labels=labels, index=label_index) / tot_sum).flatten() cluster_tof = np.array(nd.sum(tof * tot, labels=labels, index=label_index) / tot_sum).flatten() cluster_totMax = tot[tot_max] cluster_totAvg = tot_mean cluster_shot = shot[tot_max] if _cent_timewalk_lut is not None: # cluster_tof -= self._timewalk_lut[(cluster_tot / 25).astype(np.int) - 1] # cluster_tof *= 1e6 cluster_tof -= _cent_timewalk_lut[np.int_(cluster_totMax // 25) - 1] * 1e3 # cluster_tof *= 1e-6 return ( cluster_shot, cluster_x, cluster_y, cluster_tof, cluster_totAvg, cluster_totMax, cluster_size, )
[docs] def calculate_centroids_dbscan(chunk, tot_threshold, _tof_scale, epsilon, min_samples, _cent_timewalk_lut): shot, x, y, tof, tot = chunk tot_filter = tot > tot_threshold # Filter out pixels shot = shot[tot_filter] x = x[tot_filter] y = y[tot_filter] tof = tof[tot_filter] tot = tot[tot_filter] labels = perform_clustering_dbscan(shot, x, y, tof, _tof_scale, epsilon, min_samples) label_filter = labels != 0 if labels is not None and labels[label_filter].size > 0: return calculate_centroids_properties( shot[label_filter], x[label_filter], y[label_filter], tof[label_filter], tot[label_filter], labels[label_filter], _cent_timewalk_lut, ) return None
[docs] class CentroidCalculator(ProcessingStep): """ Class responsible for calculating centroids in timepix data. This includes the calculation of the clusters first and the centroids. The data processed is not the direct raw data but the data that has been processed by the PacketProcessor before (x, y, tof, tot). Methods ------- process(data): Process data and return the result. To use this class only this method should be used! Use the other methods only for testing or if you are sure about what you are doing """ def __init__( self, cent_timewalk_lut=None, number_of_processes=1, clustering_args={}, dbscan_clustering=True, *args, **kwargs, ): """ Constructor for the CentroidCalculator. Parameters ---------- tot_threshold : int Threshold applied before calculating the clusters. A higher threshold can lead to more clusters found in some cases. epsilon : float see DBSCAN min_samples : int see DBSCAN triggers_processed : int every triggers_processed trigger is used for the calculation. Increasing the value can speed up online processing if the data rate is too high to process all triggers directly. chunk_size_limit : int Maximum size of the chunks to increase the performance of DBSCAN. Higher and Lower values might increase the runtime. cent_timewalk_lut Data for correction of the time-walk parameter_wrapper_classe : ProcessingParameter Class used to wrap the processing parameters to make them changable while processing is running (useful for online optimization) """ super().__init__(name="CentroidCalculator", *args, **kwargs) self._epsilon = self.parameter_wrapper_class(clustering_args.pop("epsilon", 2.0)) self._min_samples = self.parameter_wrapper_class(clustering_args.pop("min_samples", 3)) self._tot_threshold = self.parameter_wrapper_class(clustering_args.pop("tot_threshold", 0)) self._triggers_processed = self.parameter_wrapper_class(clustering_args.pop("triggers_processed", 1)) self._cs_sensor_size = self.parameter_wrapper_class(clustering_args.pop("cs_sensor_size", 256)) self._cs_min_cluster_size = self.parameter_wrapper_class(clustering_args.pop("cs_min_cluster_size", 3)) self._cs_max_dist_tof = self.parameter_wrapper_class(clustering_args.pop("cs_max_dist_tof", 5e-8)) self._cs_tot_offset = self.parameter_wrapper_class(clustering_args.pop("cs_tot_offset", 0.5)) self._chunk_size_limit = clustering_args.pop("chunk_size_limit", 6_500) self._tof_scale = 1.7e7 self._cent_timewalk_lut = cent_timewalk_lut self._dbscan_clustering = self.parameter_wrapper_class(int(dbscan_clustering)) self.number_of_processes = number_of_processes @property def epsilon(self): return self._epsilon.value @epsilon.setter def epsilon(self, epsilon): self._epsilon.value = epsilon @property def min_samples(self): return self._min_samples.value @min_samples.setter def min_samples(self, min_samples): self._min_samples.value = min_samples @property def tot_threshold(self): """Determines which time over threshold values to filter before centroiding This is useful in reducing the computational time in centroiding and can filter out noise.""" return self._tot_threshold.value @tot_threshold.setter def tot_threshold(self, tot_threshold): self._tot_threshold.value = tot_threshold @property def triggers_processed(self): """Setting for the number of packets skipped during processing. Every packet_skip packet is processed. This means for a value of 1 every packet is processed. For 2 only every 2nd packet is processed.""" return self._triggers_processed.value @triggers_processed.setter def triggers_processed(self, triggers_processed): self._triggers_processed.value = triggers_processed @property def cs_sensor_size(self): """Setting for the number of packets skipped during processing. Every packet_skip packet is processed. This means for a value of 1 every packet is processed. For 2 only every 2nd packet is processed.""" return self._cs_sensor_size.value @cs_sensor_size.setter def cs_sensor_size(self, cs_sensor_size): self._cs_sensor_size.value = cs_sensor_size @property def cs_min_cluster_size(self): """Setting the minimal cluster size in Cluster Streaming algorithm""" return self._cs_min_cluster_size.value @cs_min_cluster_size.setter def cs_min_cluster_size(self, cs_min_cluster_size): self._cs_min_cluster_size.value = cs_min_cluster_size @property def cs_max_dist_tof(self): """Setting the maximal ToF distance between the voxels belonging to the cluster in Cluster Streaming algorithm""" return self._cs_max_dist_tof.value @cs_max_dist_tof.setter def cs_max_dist_tof(self, cs_max_dist_tof): self._cs_max_dist_tof.value = cs_max_dist_tof @property def cs_tot_offset(self): """Setting the ToT ratio factor of the voxel to the ToT of previous voxel in Cluster Streaming algorithm. Zero factor means ToT of prev. voxel should be larger. 0.5 factor means ToT of prev voxel could be high than the half of the considered voxel""" return self._cs_tot_offset.value @cs_tot_offset.setter def cs_tot_offset(self, cs_tot_offset): self._cs_tot_offset.value = cs_tot_offset @property def dbscan_clustering(self): return bool(self._dbscan_clustering.value) @dbscan_clustering.setter def dbscan_clustering(self, dbscan_clustering): self._dbscan_clustering.value = int(dbscan_clustering)
[docs] def process(self, data): if data is None: return None if self.dbscan_clustering: shot, x, y, tof, tot = self.__skip_triggers(*data) chunks = self.__divide_into_chunks(shot, x, y, tof, tot) centroids_in_chunks = self.perform_centroiding_dbscan(chunks) else: chunks = self.cluster_stream_preprocess(*data) centroids_in_chunks = self.perform_centroiding_cluster_stream(chunks) return self.centroid_chunks_to_centroids(centroids_in_chunks)
def __skip_triggers(self, shot, x, y, tof, tot): unique_shots = np.unique(shot) selected_shots = unique_shots[:: self.triggers_processed] mask = np.isin(shot, selected_shots) return shot[mask], x[mask], y[mask], tof[mask], tot[mask]
[docs] def cluster_stream_preprocess(self, shot, x, y, tof, tot): order = shot.argsort() shot, x, y, tof, tot = shot[order], x[order], y[order], tof[order], tot[order] _, unique_trig_nr_indices, unique_trig_nr_counts = np.unique(shot, return_index=True, return_counts=True) shot, x, y, tof, tot = [np.split(arr, unique_trig_nr_indices[1:]) for arr in [shot, x, y, tof, tot]] for i, sh in enumerate(shot): sorting_indexes = np.lexsort((-tot[i], tof[i])) x[i] = x[i][sorting_indexes] y[i] = y[i][sorting_indexes] tof[i] = tof[i][sorting_indexes] tot[i] = tot[i][sorting_indexes] chunks = [] for i in range(len(shot)): chunks.append(np.vstack((shot[i], x[i], y[i], tof[i], tot[i]))) return chunks
def __divide_into_chunks(self, shot, x, y, tof, tot): """Reordering the voxels can have an impact on the clusterings result. See CentroidCalculator.perform_clustering string doc for further information!""" order = shot.argsort() shot, x, y, tof, tot = shot[order], x[order], y[order], tof[order], tot[order] split_indices = self.__calc_trig_chunks_split_indices(shot) if len(split_indices) > 0: shot, x, y, tof, tot = [np.split(arr, split_indices) for arr in [shot, x, y, tof, tot]] chunks = [] for i in range(len(shot)): chunks.append((shot[i], x[i], y[i], tof[i], tot[i])) return chunks else: return [(shot, x, y, tof, tot)] def __calc_trig_chunks_split_indices(self, shot): _, unique_trig_nr_indices, unique_trig_nr_counts = np.unique(shot, return_index=True, return_counts=True) trigger_chunks = [] trigger_chunk_voxel_counter = 0 for index, unique_trig_nr_index in enumerate(unique_trig_nr_indices): if trigger_chunk_voxel_counter < self._chunk_size_limit: trigger_chunk_voxel_counter += unique_trig_nr_counts[index] else: trigger_chunks.append(unique_trig_nr_index) trigger_chunk_voxel_counter = unique_trig_nr_counts[index] return trigger_chunks
[docs] def centroid_chunks_to_centroids(self, chunks): # range(7) because the centroids have 7 dimensions: shot, x, y, tof, tot avg, tot max, cluster size """centroids = [[] for i in range(7)] for chunk in list(chunks): if chunk != None: for index, coordinate in enumerate(chunk): centroids[index].append(coordinate)""" joined_chunks = list(filter(None, chunks)) if joined_chunks: return np.concatenate(joined_chunks, axis=1) else: return None
[docs] def perform_centroiding_dbscan(self, chunks): # with Pool(self.number_of_processes) as p: # return p.map(self.calculate_centroids_dbscan, chunks) return Parallel(n_jobs=self.number_of_processes)( delayed(calculate_centroids_dbscan)( c, self.tot_threshold, self._tof_scale, self.epsilon, self.min_samples, self._cent_timewalk_lut ) for c in chunks )
# return map(self.calculate_centroids_dbscan, chunks)
[docs] def perform_centroiding_cluster_stream(self, chunks): self.cstream = ClusterStream( self.cs_sensor_size, self.cs_max_dist_tof, self.cs_min_cluster_size, self.cs_tot_offset ) # with Pool(self.number_of_processes) as p: # return p.map(self.calculate_centroids_cluster_stream, chunks) return map(self.calculate_centroids_cluster_stream, chunks)
[docs] def calculate_centroids_dbscan(self, chunk): shot, x, y, tof, tot = chunk tot_filter = tot > self.tot_threshold # Filter out pixels shot = shot[tot_filter] x = x[tot_filter] y = y[tot_filter] tof = tof[tot_filter] tot = tot[tot_filter] labels = self.perform_clustering_dbscan(shot, x, y, tof) label_filter = labels != 0 if labels is not None and labels[label_filter].size > 0: return self.calculate_centroids_properties( shot[label_filter], x[label_filter], y[label_filter], tof[label_filter], tot[label_filter], labels[label_filter], ) return None
[docs] def calculate_centroids_cluster_stream(self, chunk): labels = self.cstream.perform(np.transpose(chunk[1:, :])) label_filter = labels != 0 shot, x, y, tof, tot = chunk if labels is not None and labels[label_filter].size > 0: return self.calculate_centroids_properties( shot[label_filter], x[label_filter], y[label_filter], tof[label_filter], tot[label_filter], labels[label_filter], ) return None
[docs] class CentroidCalculatorPooled(CentroidCalculator): """ Parallelized implementation of CentroidCalculator using mp.Pool for parallelization. """ def __init__(self, number_of_processes=None, *args, **kwargs): super().__init__(*args, **kwargs) self._number_of_processes = number_of_processes
[docs] def perform_centroiding(self, chunks): return self._pool.map(self.calculate_centroids, chunks)
[docs] def pre_process(self): self._pool = mp.Pool(self._number_of_processes) return super().pre_process()
[docs] def post_process(self): self._pool.close() self._pool.join() return super().post_process()
def __getstate__(self): self_dict = self.__dict__.copy() del self_dict["_pool"] return self_dict