Do retrieval.
(features, retriever_beam_size, mode, params)
| 86 | |
| 87 | |
| 88 | def retrieve(features, retriever_beam_size, mode, params): |
| 89 | """Do retrieval.""" |
| 90 | tokenizer, vocab_lookup_table = bert_utils.get_tf_tokenizer( |
| 91 | params["retriever_module_path"]) |
| 92 | |
| 93 | question_token_ids = tokenizer.tokenize( |
| 94 | tf.expand_dims(features["question"], 0)) |
| 95 | question_token_ids = tf.cast( |
| 96 | question_token_ids.merge_dims(1, 2).to_tensor(), tf.int32) |
| 97 | cls_token_id = vocab_lookup_table.lookup(tf.constant("[CLS]")) |
| 98 | sep_token_id = vocab_lookup_table.lookup(tf.constant("[SEP]")) |
| 99 | question_token_ids = tf.concat( |
| 100 | [[[tf.cast(cls_token_id, tf.int32)]], question_token_ids, |
| 101 | [[tf.cast(sep_token_id, tf.int32)]]], -1) |
| 102 | |
| 103 | retriever_module = hub.Module( |
| 104 | params["retriever_module_path"], |
| 105 | tags={"train"} if mode == tf_estimator.ModeKeys.TRAIN else {}, |
| 106 | trainable=True) |
| 107 | |
| 108 | # [1, projection_size] |
| 109 | question_emb = retriever_module( |
| 110 | inputs=dict( |
| 111 | input_ids=question_token_ids, |
| 112 | input_mask=tf.ones_like(question_token_ids), |
| 113 | segment_ids=tf.zeros_like(question_token_ids)), |
| 114 | signature="projected") |
| 115 | |
| 116 | block_emb, searcher = scann_utils.load_scann_searcher( |
| 117 | var_name="block_emb", |
| 118 | checkpoint_path=os.path.join(params["retriever_module_path"], "encoded", |
| 119 | "encoded.ckpt"), |
| 120 | num_neighbors=retriever_beam_size) |
| 121 | |
| 122 | # [1, retriever_beam_size] |
| 123 | retrieved_block_ids, _ = searcher.search_batched(question_emb) |
| 124 | |
| 125 | # [1, retriever_beam_size, projection_size] |
| 126 | retrieved_block_emb = tf.gather(block_emb, retrieved_block_ids) |
| 127 | |
| 128 | # [retriever_beam_size] |
| 129 | retrieved_block_ids = tf.squeeze(retrieved_block_ids) |
| 130 | |
| 131 | # [retriever_beam_size, projection_size] |
| 132 | retrieved_block_emb = tf.squeeze(retrieved_block_emb) |
| 133 | |
| 134 | # [1, retriever_beam_size] |
| 135 | retrieved_logits = tf.matmul( |
| 136 | question_emb, retrieved_block_emb, transpose_b=True) |
| 137 | |
| 138 | # [retriever_beam_size] |
| 139 | retrieved_logits = tf.squeeze(retrieved_logits, 0) |
| 140 | |
| 141 | blocks_dataset = tf.data.TFRecordDataset( |
| 142 | params["block_records_path"], buffer_size=512 * 1024 * 1024) |
| 143 | blocks_dataset = blocks_dataset.batch( |
| 144 | params["num_block_records"], drop_remainder=True) |
| 145 | blocks = tf.get_local_variable( |