(self, name, params, data_loader)
| 51 | """EmQL Implementation.""" |
| 52 | |
| 53 | def __init__(self, name, params, data_loader): |
| 54 | |
| 55 | with tf.variable_scope('embeddings_mat', reuse=tf.AUTO_REUSE): |
| 56 | self.entity_embeddings_mat = tf.get_variable( |
| 57 | name='entity_embeddings_mat', |
| 58 | shape=[data_loader.num_entities, params['entity_emb_size']], |
| 59 | initializer=tf.random_normal_initializer(), |
| 60 | trainable=params['train_entity_emb']) |
| 61 | |
| 62 | self.relation_embeddings_mat = tf.get_variable( |
| 63 | name='relation_embeddings_mat', |
| 64 | shape=[data_loader.num_relations, params['relation_emb_size']], |
| 65 | initializer=tf.random_normal_initializer(), |
| 66 | trainable=params['train_relation_emb']) |
| 67 | |
| 68 | if name in ['metaqa2', 'metaqa3']: |
| 69 | self.word_embeddings_mat = tf.get_variable( |
| 70 | name='word_embeddings_mat', |
| 71 | shape=[data_loader.num_vocab, params['vocab_emb_size']], |
| 72 | initializer=tf.random_normal_initializer(), |
| 73 | trainable=True) |
| 74 | |
| 75 | ### Convert fact info to tf tensors. |
| 76 | self.all_entity_ids = tf.constant( |
| 77 | data_loader.all_entity_ids, dtype=tf.int32) |
| 78 | self.all_entity_sketches = tf.constant( |
| 79 | data_loader.all_entity_sketches, dtype=tf.int32) |
| 80 | self.all_relation_sketches = tf.constant( |
| 81 | data_loader.all_relation_sketches, dtype=tf.int32) |
| 82 | self.all_fact_subjids = tf.constant( |
| 83 | data_loader.all_fact_subjids, dtype=tf.int32) |
| 84 | self.all_fact_relids = tf.constant( |
| 85 | data_loader.all_fact_relids, dtype=tf.int32) |
| 86 | self.all_fact_objids = tf.constant( |
| 87 | data_loader.all_fact_objids, dtype=tf.int32) |
| 88 | |
| 89 | # Construct fact embeddings. |
| 90 | self.kb_embeddings_mat = self.load_kb_embeddings(name, params) |
| 91 | |
| 92 | # Construct other utilities |
| 93 | if name == 'webqsp': |
| 94 | self.all_entity_is_cvt = tf.constant( |
| 95 | data_loader.all_entity_is_cvt, dtype=tf.float32) |
| 96 | self.bert_module = hub.Module( |
| 97 | params['bert_handle'], |
| 98 | tags={'train'} if params['train_bert'] else {}, |
| 99 | trainable=params['train_bert']) |
| 100 | |
| 101 | def load_kb_embeddings(self, name, params): |
| 102 | """Construct or load KB embeddings. |
nothing calls this directly
no test coverage detected