def _get_indexed_samples(self, indices):
graph_lists = []
targets = []
for index in indices:
sentence = graph.copy_graph(
self._questions_data[index]) # type: Dict
question_tokens = self.get_question_tokens(index)
del sentence['tokens']
graph_list = [(sentence, 1.0 * self._p.get("mult.f1.by", 1.0))]
negative_pool_size = self._p.get("max.negative.samples",
30) - len(graph_list)
negative_pool = self._get_negative_instances_for_sentence(
negative_pool_size, graph_list)
negative_pool = [(g, 0.0) for g in negative_pool]
one_negative = negative_pool[-1]
instance = graph_list + negative_pool[:-1]
np.random.shuffle(instance)
instance = [one_negative] + instance
target = [g[1] for g in instance] + [0.0] * (
self._p.get("max.negative.samples", 30) - len(instance))
instance = [el[0] for el in instance]
graph_lists.append((question_tokens, instance))
targets.append(target)
return graph_lists, np.asarray(targets)
def _get_negative_instances_for_sentence(self, pool_size, graph_list):
sentence = graph_list[0][0]
negative_pool = []
for i in range(pool_size):
neg_graph = graph.copy_graph(sentence)
for edge in neg_graph['edgeSet']:
edge['kbID'] = self._idx2property[np.random.randint(
len(self._idx2property))] + "v"
if "label" in edge:
del edge["label"]
negative_pool.append(neg_graph)
return negative_pool