def assess_model(self, model: DistributionMatchingGenerator) -> DistributionLearningBenchmarkResult:
chemnet = self._load_chemnet()
start_time = time.time()
generated_molecules = sample_valid_molecules(model=model, number_molecules=self.number_samples)
end_time = time.time()
if len(generated_molecules) != self.number_samples:
logger.warning('The model could not generate enough valid molecules.')
mu_ref, cov_ref = self._calculate_distribution_statistics(chemnet, self.reference_molecules)
mu, cov = self._calculate_distribution_statistics(chemnet, generated_molecules)
FCD = fcd.calculate_frechet_distance(mu1=mu_ref, mu2=mu,
sigma1=cov_ref, sigma2=cov)
score = np.exp(-0.2 * FCD)
metadata = {
'number_reference_molecules': len(self.reference_molecules),
'number_generated_molecules': len(generated_molecules),
'FCD': FCD
}
return DistributionLearningBenchmarkResult(benchmark_name=self.name,
score=score,
sampling_time=end_time - start_time,
metadata=metadata)
def __call__(self, valid_molecule_bags: typing.List[multiset.BaseMultiset]):
if not len(valid_molecule_bags) >= self.fcd_scorer.sample_size:
print(f"less samples than ideal... @{len(valid_molecule_bags)}")
sample_size = len(valid_molecule_bags)
else:
sample_size = self.fcd_scorer.sample_size
# Sample the generated molecule multisets
samples_bags = random.sample(valid_molecule_bags, sample_size)
# And then for each pick one molecule randomly from the multiset:
samples = []
for s in samples_bags:
samples.append(random.choice(list(s.distinct_elements())))
chemnet = self.chemnet
print("FCD: calculating dist stats on training data...")
mu_ref, cov_ref = self.cached_ref_stats
print("FCD: calculating dist stats on new generated molecules...")
mu, cov = self.fcd_scorer._calculate_distribution_statistics(chemnet, samples)
print("FCD: ... computed stats!")
FCD = fcd.calculate_frechet_distance(mu1=mu_ref, mu2=mu,
sigma1=cov_ref, sigma2=cov)
# ^ See note 2 in class docstring.
return FCD