added composite bn experiments

paper_experiments/composite_bayesian_networks_experiments/CBN_GA/composite_BN_GA.py

@@ -0,0 +1,232 @@
+from datetime import timedelta
+import sys
+import os
+import pathlib
+current_path = pathlib.Path().resolve()
+parentdir = os.getcwd() 
+sys.path.insert(0, parentdir)
+sys.path.append(str(current_path) + '\\examples\\bn')
+import pandas as pd
+from sklearn import preprocessing
+import bamt.preprocessors as pp
+from golem.core.optimisers.optimization_parameters import GraphRequirements
+from golem.core.adapter import DirectAdapter
+from golem.core.optimisers.genetic.gp_optimizer import EvoGraphOptimizer
+from golem.core.optimisers.genetic.gp_params import GPAlgorithmParameters
+from golem.core.optimisers.genetic.operators.inheritance import GeneticSchemeTypesEnum
+from golem.core.optimisers.genetic.operators.selection import SelectionTypesEnum
+from golem.core.optimisers.genetic.operators.crossover import exchange_parents_one_crossover, exchange_parents_both_crossover, exchange_edges_crossover
+from golem.core.optimisers.objective.objective import Objective
+from golem.core.optimisers.optimizer import GraphGenerationParams
+from examples.composite_bn.composite_model import CompositeModel
+from examples.composite_bn.composite_node import CompositeNode
+from sklearn.model_selection import train_test_split
+from examples.composite_bn.composite_bn_genetic_operators import (
+ custom_crossover_all_model as composite_crossover_all_model, 
+ custom_mutation_add_structure as composite_mutation_add_structure, 
+ custom_mutation_delete_structure as composite_mutation_delete_structure, 
+ custom_mutation_reverse_structure as composite_mutation_reverse_structure,
+ custom_mutation_add_model as composite_mutation_add_model,
+)
+from functools import partial
+from examples.composite_bn.comparison import Comparison
+from examples.composite_bn.fitness_function import FitnessFunction
+from examples.composite_bn.rule import Rule
+from examples.composite_bn.likelihood import Likelihood
+from examples.composite_bn.write_txt import Write
+from examples.bn.bn_genetic_operators import (
+ custom_mutation_add_structure as classical_mutation_add_structure,
+ custom_mutation_delete_structure as classical_mutation_delete_structure,
+ custom_mutation_reverse_structure as classical_mutation_reverse_structure
+)
+
+
+
+def run_example(file):
+ if exist_true_str:
+ with open('examples/data/1000/txt/'+(file)+'.txt') as f:
+ lines = f.readlines()
+ true_net = []
+ for l in lines:
+ e0 = l.split()[0]
+ e1 = l.split()[1].split('\n')[0]
+ true_net.append((e0, e1)) 
+
+ fitness_function = FitnessFunction()
+ FF_classical = fitness_function.classical_K2
+ FF_composite = fitness_function.composite_metric
+
+ if bn_type == 'classical':
+ fitness_function_GA = FF_classical
+
+ mutations = [
+ classical_mutation_add_structure, 
+ classical_mutation_delete_structure, 
+ classical_mutation_reverse_structure
+ ]
+
+ crossovers = [
+ exchange_parents_one_crossover,
+ exchange_parents_both_crossover,
+ exchange_edges_crossover
+ ]
+ elif bn_type == 'composite':
+
+ composite_FF = FF_composite
+ complexity_FF = FF_classical
+
+
+ mutations = [
+ composite_mutation_add_structure, 
+ composite_mutation_delete_structure, 
+ composite_mutation_reverse_structure, 
+ composite_mutation_add_model 
+ ] 
+
+ crossovers = [
+ exchange_parents_one_crossover,
+ exchange_parents_both_crossover,
+ exchange_edges_crossover,
+ composite_crossover_all_model
+ ]
+ else:
+ print('There is no such type of BN: "{}". You can only use "classical" or "composite".'.format(bn_type)) 
+ return 
+
+
+ if file in ['abalone', 'adult', 'australian_statlog', 'liver_disorders']:
+ data = pd.read_csv('examples/data/1000/UCI/' + file + '.data') 
+ else:
+ data = pd.read_csv('examples/data/1000/csv/' + file + '.csv') 
+
+ if 'Unnamed: 0' in data.columns:
+ data.drop(['Unnamed: 0'], axis=1, inplace=True)
+ data.dropna(inplace=True)
+ data.reset_index(inplace=True, drop=True)
+ vertices = list(data.columns)
+
+ encoder = preprocessing.LabelEncoder()
+ discretizer = preprocessing.KBinsDiscretizer(n_bins=5, encode='ordinal', strategy='quantile')
+
+ p_for_composite = pp.Preprocessor([('encoder', encoder)])
+ discretized_data_composite, _ = p_for_composite.apply(data)
+
+ data_train_composite , data_val_composite = train_test_split(discretized_data_composite, test_size=0.2, shuffle = True, random_state=random_seed[number-1])
+ data_train_composite , data_test_composite = train_test_split(data_train_composite, test_size=0.2, shuffle = True, random_state=random_seed[number-1])
+
+ rules = Rule().bn_rules()
+
+ initial = [CompositeModel(nodes=[CompositeNode(nodes_from=None,
+ content={'name': vertex,
+ 'type': p_for_composite.nodes_types[vertex],
+ 'parent_model': None}) 
+ for vertex in vertices])] 
+
+
+ objective = Objective(
+ {'fitness function': partial(composite_FF, data_train = data_train_composite, data_test = data_test_composite)},
+ is_multi_objective=False,
+ )
+
+ requirements = GraphRequirements(
+ max_arity=100,
+ max_depth=100, 
+ num_of_generations=n_generation,
+ timeout=timedelta(minutes=time_m),
+ early_stopping_iterations = early_stopping_iterations,
+ n_jobs=-1
+ )
+
+ optimiser_parameters = GPAlgorithmParameters(
+ multi_objective=objective.is_multi_objective,
+ pop_size=pop_size,
+ max_pop_size = pop_size,
+ crossover_prob=crossover_probability, 
+ mutation_prob=mutation_probability,
+ genetic_scheme_type = GeneticSchemeTypesEnum.steady_state,
+ selection_types = [SelectionTypesEnum.tournament],
+ mutation_types = mutations,
+ crossover_types = crossovers,
+ )
+
+ graph_generation_params = GraphGenerationParams(
+ adapter=DirectAdapter(base_graph_class=CompositeModel, base_node_class=CompositeNode),
+ rules_for_constraint=rules,
+ )
+
+ optimiser = EvoGraphOptimizer(
+ graph_generation_params=graph_generation_params,
+ graph_optimizer_params=optimiser_parameters,
+ requirements=requirements,
+ initial_graphs=initial,
+ objective=objective)
+
+ optimized_graphs = optimiser.optimise(objective)
+
+ vars_of_interest = {}
+ comparison = Comparison()
+ LL = Likelihood() 
+
+ for optimized_graph in optimized_graphs:
+ optimized_structure = [(str(edge[0]), str(edge[1])) for edge in optimized_graph.get_edges()]
+ score = composite_FF(optimized_graph, data_train = data_train_composite, data_test = data_val_composite)
+ spent_time = optimiser.timer.minutes_from_start
+
+ if bn_type == 'composite':
+ p = pp.Preprocessor([('encoder', encoder), ('discretizer', discretizer)]) 
+ discretized_data, _ = p.apply(data)
+ data_train_test , data_val = train_test_split(discretized_data, test_size=0.2, shuffle = True, random_state=random_seed[number-1])
+ data_train , data_test = train_test_split(data_train_test, test_size=0.2, shuffle = True, random_state=random_seed[number-1])
+
+ likelihood_classical = LL.likelihood_function(optimized_graph, data_train=data_train, data_val=data_val)
+ likelihood_composite = LL.likelihood_function_composite(optimized_graph, data_train_composite, data_val_composite)
+
+ if exist_true_str:
+ f1 = comparison.F1(optimized_structure, true_net)
+ SHD = comparison.precision_recall(optimized_structure, true_net)['SHD']
+ if bn_type == 'composite': 
+ models = {node:node.content['parent_model'] for node in optimized_graph.nodes}
+
+ vars_of_interest['Structure'] = optimized_structure
+ vars_of_interest['Score'] = -score
+ vars_of_interest['Likelihood classical'] = likelihood_classical
+ vars_of_interest['Likelihood composite'] = likelihood_composite
+
+ vars_of_interest['Spent time'] = spent_time
+ if exist_true_str:
+ vars_of_interest['f1'] = f1
+ vars_of_interest['SHD'] = SHD
+ if bn_type == 'composite': 
+ vars_of_interest['Models'] = models
+
+ vars_of_interest['Generation number'] = optimiser.current_generation_num
+ vars_of_interest['Population number'] = optimiser.graph_optimizer_params.pop_size
+
+ write = Write()
+ write.write_txt(vars_of_interest, path = os.path.join(parentdir, 'examples'), file_name = 'composite_BN_' + file + '_run_' + str(number) + '.txt')
+
+
+
+if __name__ == '__main__':
+ files = ['asia', 'cancer', 'earthquake', 'sachs', 'survey', 'healthcare', 'sangiovese', 'alarm', 'barley', 'child', 'insurance', 'mildew', 'water', 'ecoli70', 'magic-niab', 'mehra-complete', 'hailfinder']
+ exist_true_str = True
+ pop_size = 20 
+ n_generation = 1000
+ crossover_probability = 0.8
+ mutation_probability = 0.9
+ early_stopping_iterations = 10 
+ time_m = 15
+ random_seed = [87, 60, 37, 99, 42, 92, 48, 91, 86, 33]
+
+ n = 10
+ for file in files:
+
+ for bn_type in ['composite']: 
+ number = 1
+ while number <= n:
+ run_example(file) 
+ number += 1 
+
+
+
+