def _get_case_data(mongo_client):
"""
Get a :class:`.CaseData` instance.
Parameters
----------
mongo_client : :class:`pymongo.MongoClient`
The mongo client the database should connect to.
"""
# The basic idea here is that the _counter.get_count method will
# return a different "fitness value" each time it is called.
# When the test runs fitness_calculator.get_fitness_value(), if
# caching is working, the same number as before will be returned.
# However, if caching is not working, a different number will be
# returned as the fitness value.
db = stk.ValueMongoDb(
mongo_client=mongo_client,
collection='test_caching',
database='_stk_pytest_database',
)
fitness_calculator = stk.PropertyVector(
property_functions=(_counter.get_count, ),
input_database=db,
output_database=db,
)
molecule = stk.BuildingBlock('BrCCBr')
fitness_value = fitness_calculator.get_fitness_value(molecule)
return CaseData(
fitness_calculator=fitness_calculator,
molecule=molecule,
fitness_value=fitness_value,
)
scores = []
for bb in mol.get_building_blocks():
rdkit_mol = bb.to_rdkit_mol()
rdkit_mol.UpdatePropertyCache()
rdkit.GetSymmSSSR(rdkit_mol)
rdkit_mol.GetRingInfo()
scores.append(scscorer(rdkit_mol)[1])
return sum(scores)
# Defines synthetic accesibility function to use.
synthetic_accesibility_func = scscore
cage_fitness_calculator = stk.PropertyVector(
pore_diameter,
largest_window,
window_std,
synthetic_accesibility_func,
)
fitness_calculator = stk.If(
condition=lambda mol: failed_optimizer.is_in_cache(mol),
true_calculator=stk.FitnessFunction(lambda mol: None),
false_calculator=cage_fitness_calculator,
)
# #####################################################################
# Fitness normalizer.
# #####################################################################
def valid_fitness(population, mol):
pw_mol = pywindow.Molecule.load_rdkit_mol(mol.to_rdkit_mol())
mol.pore_diameter = abs(pw_mol.calculate_pore_diameter() - 5)
return mol.pore_diameter
def window_std(mol):
pw_mol = pywindow.Molecule.load_rdkit_mol(mol.to_rdkit_mol())
windows = pw_mol.calculate_windows()
mol.window_std = None
if windows is not None and len(windows) > 3:
mol.window_std = np.std(windows)
return mol.window_std
fitness_calculator = stk.PropertyVector(
pore_diameter,
window_std,
)
def valid_fitness(population, mol):
return None not in population.get_fitness_values()[mol]
fitness_normalizer = stk.Sequence(
stk.Power([1, -1], filter=valid_fitness),
stk.DivideByMean(filter=valid_fitness),
stk.Multiply([1.0, 1.0], filter=valid_fitness),
stk.Sum(filter=valid_fitness),
stk.ReplaceFitness(
replacement_fn=lambda population:
min(
# #####################################################################
# Optimizer.
# #####################################################################
optimizer = stk.NullOptimizer(use_cache=True)
# #####################################################################
# Fitness calculator.
# #####################################################################
def num_atoms(mol):
return len(mol.atoms)
fitness_calculator = stk.PropertyVector(num_atoms)
# #####################################################################
# Fitness normalizer.
# #####################################################################
# The PropertyVector fitness calculator will set the fitness as
# [n_atoms] use the Sum() fitness normalizer to convert the fitness to
# just n_atoms^0.5. The sqrt is because we use the Power normalizer.
fitness_normalizer = stk.NormalizerSequence(stk.Power(0.5), stk.Sum())
# #####################################################################
# Exit condition.
# #####################################################################
terminator = stk.NumGenerations(25)
import numpy as np
import pytest
import stk
from ..case_data import CaseData
@pytest.fixture(
scope='session',
params=(lambda: CaseData(
fitness_calculator=stk.PropertyVector(property_functions=(
stk.Molecule.get_num_atoms,
stk.Molecule.get_num_bonds,
stk.Molecule.get_maximum_diameter,
), ),
molecule=stk.BuildingBlock('BrCCBr').with_position_matrix(
position_matrix=np.array([
[0, 0, 0],
[10, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
],
dtype=np.float64), ),
fitness_value=(8, 7, 10),
), ),
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--mongodb_uri',
help='The MongoDB URI for the database to connect to.',
default='mongodb://localhost:27017/')
args = parser.parse_args()
logging.basicConfig(level=logging.INFO)
# Use a random seed to get reproducible results.
random_seed = 4
generator = np.random.RandomState(random_seed)
logger.info('Making building blocks.')
# Load the building block databases.
fluoros = tuple(
get_building_blocks(
path=pathlib.Path(__file__).parent / 'fluoros.txt',
functional_group_factory=stk.FluoroFactory(),
))
bromos = tuple(
get_building_blocks(
path=pathlib.Path(__file__).parent / 'bromos.txt',
functional_group_factory=stk.BromoFactory(),
))
initial_population = tuple(get_initial_population(fluoros, bromos))
# Write the initial population.
for i, record in enumerate(initial_population):
write(record.get_molecule(), f'initial_{i}.mol')
client = pymongo.MongoClient(args.mongodb_uri)
db = stk.ConstructedMoleculeMongoDb(client)
fitness_db = stk.ValueMongoDb(client, 'fitness_values')
# Plot selections.
generation_selector = stk.Best(
num_batches=25,
duplicate_molecules=False,
)
stk.SelectionPlotter('generation_selection', generation_selector)
mutation_selector = stk.Roulette(
num_batches=5,
random_seed=generator.randint(0, 1000),
)
stk.SelectionPlotter('mutation_selection', mutation_selector)
crossover_selector = stk.Roulette(
num_batches=3,
batch_size=2,
random_seed=generator.randint(0, 1000),
)
stk.SelectionPlotter('crossover_selection', crossover_selector)
fitness_calculator = stk.PropertyVector(
property_functions=(
get_num_rotatable_bonds,
get_complexity,
get_num_bad_rings,
),
input_database=fitness_db,
output_database=fitness_db,
)
fitness_normalizer = stk.NormalizerSequence(
fitness_normalizers=(
# Prevent division by 0 error in DivideByMean, by ensuring
# a value of each property to be at least 1.
stk.Add((1, 1, 1)),
stk.DivideByMean(),
# Obviously, because all coefficients are equal, the
# Multiply normalizer does not need to be here. However,
# it's here to show that you can easily change the relative
# importance of each component of the fitness value, by
# changing the values of the coefficients.
stk.Multiply((1, 1, 1)),
stk.Sum(),
stk.Power(-1),
), )
ea = stk.EvolutionaryAlgorithm(
num_processes=1,
initial_population=initial_population,
fitness_calculator=fitness_calculator,
mutator=stk.RandomMutator(
mutators=(
stk.RandomBuildingBlock(
building_blocks=fluoros,
is_replaceable=is_fluoro,
random_seed=generator.randint(0, 1000),
),
stk.SimilarBuildingBlock(
building_blocks=fluoros,
is_replaceable=is_fluoro,
random_seed=generator.randint(0, 1000),
),
stk.RandomBuildingBlock(
building_blocks=bromos,
is_replaceable=is_bromo,
random_seed=generator.randint(0, 1000),
),
stk.SimilarBuildingBlock(
building_blocks=bromos,
is_replaceable=is_bromo,
random_seed=generator.randint(0, 1000),
),
),
random_seed=generator.randint(0, 1000),
),
crosser=stk.GeneticRecombination(get_gene=get_functional_group_type, ),
generation_selector=generation_selector,
mutation_selector=mutation_selector,
crossover_selector=crossover_selector,
fitness_normalizer=fitness_normalizer,
)
logger.info('Starting EA.')
generations = []
for generation in ea.get_generations(50):
for record in generation.get_molecule_records():
db.put(record.get_molecule())
generations.append(generation)
# Write the final population.
for i, record in enumerate(generation.get_molecule_records()):
write(record.get_molecule(), f'final_{i}.mol')
logger.info('Making fitness plot.')
# Normalize the fitness values across the entire EA before
# plotting the fitness values.
generations = tuple(
normalize_generations(
fitness_calculator=fitness_calculator,
fitness_normalizer=fitness_normalizer,
generations=generations,
))
fitness_progress = stk.ProgressPlotter(
generations=generations,
get_property=lambda record: record.get_fitness_value(),
y_label='Fitness Value',
)
fitness_progress.write('fitness_progress.png')
fitness_progress.get_plot_data().to_csv('fitness_progress.csv')
logger.info('Making rotatable bonds plot.')
rotatable_bonds_progress = stk.ProgressPlotter(
generations=generations,
get_property=lambda record: get_num_rotatable_bonds(record.
get_molecule()),
y_label='Number of Rotatable Bonds',
)
rotatable_bonds_progress.write('rotatable_bonds_progress.png')
