Belle II Software development
DeepFlavorTagger.py
1#!/usr/bin/env python3
2
3
10
11import json
12import os
13import basf2_mva
14from basf2 import B2ERROR, B2FATAL
15import basf2
16from variables import variables as vm
17import modularAnalysis as ma
18
19
20def get_variables(particle_list, ranked_variable, variables=None, particleNumber=1):
21 """ creates variable name pattern requested by the basf2 variable getVariableByRank()
22 :param particle_list:
23 :param ranked_variable:
24 :param variables:
25 :param particleNumber:
26 :return:
27 """
28 var_list = []
29 for var in variables:
30 for i_num in range(1, particleNumber + 1):
31 var_list.append('getVariableByRank(' + particle_list + ', ' + ranked_variable + ', ' + var + ', ' +
32 str(i_num) + ')')
33 return var_list
34
35
36def construct_default_variable_names(particle_lists=None, ranked_variable='p', variables=None, particleNumber=5):
37 """ construct default variables (that are sorted by charge and ranked by momentum)
38 :param particle_lists:
39 :param ranked_variable:
40 :param variables:
41 :param particleNumber:
42 :return:
43 """
44 from ROOT import Belle2 # noqa
45 if particle_lists is None:
46 particle_lists = ['pi+:pos_charged', 'pi+:neg_charged']
47
48 variable_names = []
49 for p_list in particle_lists:
50 variable_names += get_variables(p_list, ranked_variable, variables, particleNumber)
51
52 # make root compatible
53 root_compatible_list = []
54 for var in variable_names:
55 root_compatible_list.append(Belle2.MakeROOTCompatible.makeROOTCompatible(var))
56
57 return root_compatible_list
58
59
60def DeepFlavorTagger(particle_lists, mode='expert', working_dir='', uniqueIdentifier='standard', variable_list=None,
61 target='qrCombined', overwrite=False,
62 transform_to_probability=False, signal_fraction=-1.0, classifier_args=None,
63 train_valid_fraction=.92, mva_steering_file='analysis/scripts/dft/tensorflow_dnn_interface.py',
64 maskName='all',
65 path=None):
66 """
67 Interfacing for the DeepFlavorTagger. This function can be used for training (``teacher``), preparation of
68 training datasets (``sampler``) and inference (``expert``).
69
70 This function requires reconstructed B meson signal particle list and where an RestOfEvent is built.
71
72 :param particle_lists: string or list[string], particle list(s) of the reconstructed signal B meson
73 :param mode: string, valid modes are ``expert`` (default), ``teacher``, ``sampler``
74 :param working_dir: string, working directory for the method
75 :param uniqueIdentifier: string, database identifier for the method
76 :param variable_list: list[string], name of the basf2 variables used for discrimination
77 :param target: string, target variable
78 :param overwrite: bool, overwrite already (locally!) existing training
79 :param transform_to_probability: bool, enable a purity transformation to compensate potential over-training,
80 can only be set during training
81 :param signal_fraction: float, (experimental) signal fraction override,
82 transform to output to a probability if an uneven signal/background fraction is used in the training data,
83 can only be set during training
84 :param classifier_args: dictionary, customized arguments for the mlp
85 possible attributes of the dictionary are:
86 lr_dec_rate: learning rate decay rate
87 lr_init: learning rate initial value
88 mom_init: momentum initial value
89 min_epochs: minimal number of epochs
90 max_epochs: maximal number of epochs
91 stop_epochs: epochs to stop without improvements on the validation set for early stopping
92 batch_size: batch size
93 seed: random seed for tensorflow
94 layers: [[layer name, activation function, input_width, output_width, init_bias, init_weights],..]
95 wd_coeffs: weight decay coefficients, length of layers
96 cuda_visible_devices: selection of cuda devices
97 tensorboard_dir: addition directory for logging the training process
98 :param train_valid_fraction: float, train-valid fraction (.92). If transform to probability is
99 enabled, train valid fraction will be split into a test set (.5)
100 :param maskName: get ROE particles from a specified ROE mask
101 :param path: basf2 path obj
102 :return: None
103 """
104
105 if isinstance(particle_lists, str):
106 particle_lists = [particle_lists]
107
108 if mode not in ['expert', 'teacher', 'sampler']:
109 B2FATAL(f'Invalid mode {mode}')
110
111 if variable_list is None and mode in ['sampler', 'teacher']:
112 variable_list = [
113 'useCMSFrame(p)',
114 'useCMSFrame(cosTheta)',
115 'useCMSFrame(phi)',
116 'kaonID',
117 'electronID',
118 'muonID',
119 'protonID',
120 'nCDCHits',
121 'nPXDHits',
122 'nSVDHits',
123 'dz',
124 'dr',
125 'chiProb']
126
127 if variable_list is not None and mode == 'expert':
128 B2ERROR('DFT: Variables from identifier file are used. Input variables will be ignored.')
129
130 if classifier_args is None:
131 classifier_args = {}
132 else:
133 assert isinstance(classifier_args, dict)
134
135 classifier_args['transform_to_prob'] = transform_to_probability
136
137 output_file_name = os.path.join(working_dir, uniqueIdentifier + '_training_data.root')
138
139 # create roe specific paths
140 roe_path = basf2.create_path()
141 dead_end_path = basf2.create_path()
142
143 # define dft specific lists to enable multiple calls, if someone really wants to do that
144 extension = particle_lists[0].replace(':', '_to_')
145 roe_particle_list_cut = ''
146 roe_particle_list = 'pi+:dft' + '_' + extension
147
148 tree_name = 'dft_variables'
149
150 # filter rest of events only for specific particle list
151 ma.signalSideParticleListsFilter(particle_lists, 'hasRestOfEventTracks > 0', roe_path, dead_end_path)
152
153 # TODO: particles with empty rest of events seems not to show up in efficiency statistics anymore
154
155 # create final state particle lists
156 ma.fillParticleList(roe_particle_list, roe_particle_list_cut, path=roe_path)
157
158 dft_particle_lists = ['pi+:pos_charged', 'pi+:neg_charged']
159
160 pos_cut = 'charge > 0 and isInRestOfEvent == 1 and passesROEMask(' + maskName + ') > 0.5 and p < infinity'
161 neg_cut = 'charge < 0 and isInRestOfEvent == 1 and passesROEMask(' + maskName + ') > 0.5 and p < infinity'
162
163 ma.cutAndCopyList(dft_particle_lists[0], roe_particle_list, pos_cut, writeOut=True, path=roe_path)
164 ma.cutAndCopyList(dft_particle_lists[1], roe_particle_list, neg_cut, writeOut=True, path=roe_path)
165
166 # sort pattern for tagging specific variables
167 rank_variable = 'p'
168 # rank_variable = 'useCMSFrame(p)'
169
170 # create tagging specific variables
171 if mode != 'expert':
172 features = get_variables(dft_particle_lists[0], rank_variable, variable_list, particleNumber=5)
173 features += get_variables(dft_particle_lists[1], rank_variable, variable_list, particleNumber=5)
174
175 for particles in dft_particle_lists:
176 ma.rankByHighest(particles, rank_variable, path=roe_path)
177
178 if mode == 'sampler':
179 if os.path.isfile(output_file_name) and not overwrite:
180 B2FATAL(f'Outputfile {output_file_name} already exists. Aborting writeout.')
181
182 # and add target
183 all_variables = features + [target]
184
185 # write to ntuples
186 ma.variablesToNtuple('', all_variables, tree_name, output_file_name, roe_path)
187
188 # write the command line output for the extern teacher to a file
189 extern_command = f'basf2_mva_teacher --datafile {output_file_name} --treename {tree_name}' + \
190 f' --identifier {uniqueIdentifier} ' + \
191 '--variables "{}" '.format('" "'.join(features)) + \
192 f'--target_variable {target}' + \
193 f' --method Python --training_fraction {train_valid_fraction}' + \
194 f" --config '{json.dumps(classifier_args)}' --framework tensorflow" + \
195 f' --steering_file {mva_steering_file}'
196
197 with open(os.path.join(working_dir, uniqueIdentifier + '_teacher_command'), 'w') as f:
198 f.write(extern_command)
199
200 elif mode == 'teacher':
201 if not os.path.isfile(output_file_name):
202 B2FATAL('There is no training data file available. Run flavor tagger in sampler mode first.')
203 general_options = basf2_mva.GeneralOptions()
204 general_options.m_datafiles = basf2_mva.vector(output_file_name)
205
206 general_options.m_treename = tree_name
207 general_options.m_target_variable = target
208 general_options.m_variables = basf2_mva.vector(*features)
209
210 general_options.m_identifier = uniqueIdentifier
211
212 specific_options = basf2_mva.PythonOptions()
213 specific_options.m_framework = 'tensorflow'
214 specific_options.m_steering_file = mva_steering_file
215 specific_options.m_training_fraction = train_valid_fraction
216
217 specific_options.m_config = json.dumps(classifier_args)
218
219 basf2_mva.teacher(general_options, specific_options)
220
221 elif mode == 'expert':
222
223 flavorTaggerInfoBuilder = basf2.register_module('FlavorTaggerInfoBuilder')
224 path.add_module(flavorTaggerInfoBuilder)
225
226 expert_module = basf2.register_module('MVAExpert')
227 expert_module.param('listNames', particle_lists)
228 expert_module.param('identifier', uniqueIdentifier)
229
230 expert_module.param('extraInfoName', 'dnn_output')
231 expert_module.param('signalFraction', signal_fraction)
232
233 roe_path.add_module(expert_module)
234
235 flavorTaggerInfoFiller = basf2.register_module('FlavorTaggerInfoFiller')
236 flavorTaggerInfoFiller.param('DNNmlp', True)
237 roe_path.add_module(flavorTaggerInfoFiller)
238
239 # Create standard alias for the output of the flavor tagger
240 vm.addAlias('DNN_qrCombined', 'qrOutput(DNN)')
241
242 path.for_each('RestOfEvent', 'RestOfEvents', roe_path)
static std::string makeROOTCompatible(std::string str)
Remove special characters that ROOT dislikes in branch names, e.g.