Fix division by 0 issue when all scores are zero

[AlanKerstjens]

In graph_ga/goal_directed_generation.py the mating pool is selected through weighted random sampling, with the probability of a molecule being selected given by its score divided by the sum of all scores (sum_scores).

guacamol_baselines/graph_ga/goal_directed_generation.py

Lines 25 to 42 in 44d24c5

	def make_mating_pool(population_mol: List[Mol], population_scores, offspring_size: int):
	"""
	Given a population of RDKit Mol and their scores, sample a list of the same size
	with replacement using the population_scores as weights
	Args:
	population_mol: list of RDKit Mol
	population_scores: list of un-normalised scores given by ScoringFunction
	offspring_size: number of molecules to return
	Returns: a list of RDKit Mol (probably not unique)
	"""
	# scores -> probs
	sum_scores = sum(population_scores)
	population_probs = [p / sum_scores for p in population_scores]
	mating_pool = np.random.choice(population_mol, p=population_probs, size=offspring_size, replace=True)
	return mating_pool

If sum_scores == 0.0 we get a division by 0. This actually occurs during the Valsartan SMARTS benchmark when using random starting molecules due to its binary nature.