retrievers.py

# coding=utf-8
# Copyright 2021 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Tensorflow 2 Retriever using a (poss. fine-tuned) BERT query encoder.
"""
# import logging
# import os
# from typing import Any, Dict, Union  # pylint: disable=unused-import
#
# from absl import flags
# import dataclasses
# import numpy as np
# import tensorflow as tf
# import tensorflow_hub as hub
# import tf_utils
# import utils
# # import h5py
# # import scann_utils
# # import bert_utils
#
#
# FLAGS = flags.FLAGS
# LOGGER = logging.getLogger(__name__)
#
#
# @dataclasses.dataclass
# class ScannConfig:
#
#   num_neighbors: int
#   training_sample_size: int
#   num_leaves: int
#   num_leaves_to_search: int
#   reordering_num_neighbors: int

#
# class BERTScaNNRetriever:
#   """Class used for BERT based retrievers such as REALM and DPR.
#
#   Parameters:
#     self.query_encoder: Model instance to encode the queries.
#     self.tokenizer: Tokenizer for the query model.
#     self.vocab_lookup_table: Vocabulary index for the query and key embedders.
#     self.scann_config: Configuration dataclass for the ScaNN builder.
#     self.block_emb: Contains the dense vectors over which MIPS is done.
#     self.scann_searcher: ScaNN MIPS index instance.
#     self.cls_token_id: Id of the CLS token for the query and the embedder
#     self.sep_token_id: Id of the SEP token for the query and the embedder
#       modules.
#     self.blocks: Object from which the raw text is obtained with indices.
#   """
#
#   def __init__(self, retriever_module_path: str,
#                block_records_path: str, num_block_records: int,
#                mode: tf.estimator.ModeKeys, scann_config: ScannConfig):
#     """Constructor for BERTScaNNRetriever.
#
#     Arguments:
#       retriever_module_path: Path of the BERT tf-hub checkpoint.
#       block_records_path: Path of the textual form of the retrieval dataset in
#                           the TFRecord format.
#       num_block_records: Number of samples in the retrieval dataset.
#       mode: tf.estimator.ModeKeys for the model, currently only eval is
#             supported.
#
#       scann_config: Configuration dataclass used to initialize the ScaNN MIPS
#                     searcher object.
#
#     """
#
#     # Two and a half min. on CPU
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever.__init__",
#                             "hub load query enc"):
#       self.query_encoder = hub.load(retriever_module_path, tags={"train"} if
#                                   mode == tf.estimator.ModeKeys.TRAIN else {})
#
#     # Instantaneous
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever.__init__",
#                             "build own tok info"):
#       # Building our own tokenization info saves us 5 min where we would load
#       # the BERT model again in bert_utils.get_tf_tokenizer
#       # Getting the vocab path from the tf2 hub object (from tf.load) seems
#       # broken
#       vocab_file = os.path.join(retriever_module_path, "assets", "vocab.txt")
#       utils.check_exists(vocab_file)
#       do_lower_case = self.query_encoder.signatures["tokenization_info"
#                                                     ]()["do_lower_case"]
#       tokenization_info = dict(vocab_file=vocab_file,
#                                do_lower_case=do_lower_case)
#
#     # Instantaneous (for something that happens once) if tokenization_info
#     # is passed (our addition) a few minutes otherwise, on CPU
#     # (not passing tokenization_info makes it have to load BERT).
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever.__init__",
#                             "get_tf_tokenizer"):
#
#       self.tokenizer, self.vocab_lookup_table = bert_utils.get_tf_tokenizer(
#           retriever_module_path, tokenization_info)
#
#     # 9 min on CPU if not in dev mode. Longuest part of the setup phase.
#     # We are using a lot of default values in the load_scann_searcher call
#     # that it would probably be helpful to finetune
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever.__init__",
#                             "load_scann_searcher"):
#       checkpoint_path = os.path.join(retriever_module_path, "encoded",
#                                      "encoded.ckpt")
#       self.scann_config = scann_config
#       self.block_emb, self.scann_searcher = scann_utils.load_scann_searcher(
#           var_name="block_emb", checkpoint_path=checkpoint_path,
#           **vars(scann_config))
#
#     # Instantaneous for something that happens once
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever",
#     "CLS and SEP tokens"):
#       self.cls_token_id = tf.cast(self.vocab_lookup_table.lookup(
#           tf.constant("[CLS]")), tf.int32)
#       self.sep_token_id = tf.cast(self.vocab_lookup_table.lookup(
#           tf.constant("[SEP]")), tf.int32)
#
#       # 3 min on CPU whwn nor in dev mode
#     with utils.log_duration(LOGGER, "BERTScaNNRetriever",
#                             "Load the textual dataset"):
#       # Extract the appropriate text
#       # The buffer_size is taken from the original ORQA code.
#       blocks_dataset = tf.data.TFRecordDataset(block_records_path,
#                                                # Value taken from the REALM
#                                                # code.
#                                                buffer_size=512 * 1024 * 1024)
#       # Get a single batch with all elements (?)
#       blocks_dataset = blocks_dataset.batch(num_block_records,
#                                             drop_remainder=True)
#       # Create a thing that gets single elements over the dataset
#       self.blocks = tf.data.experimental.get_single_element(blocks_dataset)
#
#   @tf.function
#   def retrieve(self, query_text: str) -> Dict[str, Any]:
#     """Retrieves over the retrieval dataset, from a batch of text queries.
#
#     First generates the query vector from the text, then queries the
#     approximate maximum inner-product search engine.
#     Args:
#       query_text: Batch of text queries. In string form.
#
#     Returns:
#       Returns the text of the approximate nearest neighbors, as well as their
#       inner product similarity with their query's vector representation.
#     """
#
#     # Tokenize the input tokens
#     utils.check_equal(len(query_text), FLAGS.batch_size)
#     question_token_ids = self.tokenizer.batch_encode_plus(
#         query_text)["input_ids"]
#     question_token_ids = tf.cast(
#         question_token_ids.merge_dims(1, 2).to_tensor(), tf.int32)
#
#     # Add a CLS token at the start of the input, and a SEP token at the end
#     cls_ids = tf.fill((question_token_ids.shape[0], 1), self.cls_token_id)
#     sep_ids = tf.fill((question_token_ids.shape[0], 1), self.sep_token_id)
#     question_token_ids = tf.concat((cls_ids, question_token_ids, sep_ids), 1)
#     utils.check_equal(question_token_ids.shape[0], FLAGS.batch_size)
#
#     with utils.log_duration(LOGGER, "retrieve_multi", "Encode the query"):
#       question_emb = self.query_encoder.signatures["projected"](
#           input_ids=question_token_ids,
#           input_mask=tf.ones_like(question_token_ids),
#           segment_ids=tf.zeros_like(question_token_ids))["default"]
#     LOGGER.debug("question_emb.shape: %s", question_emb.shape)
#     utils.check_equal(question_emb.shape[0], FLAGS.batch_size)
#
#     with utils.log_duration(LOGGER, "retrieve_multi", "search with ScaNN"):
#       retrieved_block_ids, _ = self.scann_searcher.search_batched(
#       question_emb)
#       utils.check_equal(retrieved_block_ids.shape, (
#           FLAGS.batch_size, self.scann_config.num_neighbors))
#
#     # Gather the embeddings
#     # [batch_size, retriever_beam_size, projection_size]
#     retrieved_block_ids = retrieved_block_ids.astype(np.int64)
#     retrieved_block_emb = tf.gather(self.block_emb, retrieved_block_ids)
#     utils.check_equal(retrieved_block_emb.shape[:2], (
#         FLAGS.batch_size, self.scann_config.num_neighbors))
#
#     # Actually retrieve the text
#     retrieved_blocks = tf.gather(self.blocks, retrieved_block_ids)
#     utils.check_equal(retrieved_blocks.shape, (
#         FLAGS.batch_size, self.scann_config.num_neighbors
#     ))
#     return retrieved_blocks

#
# class FullyCachedRetriever:
#   def __init__(
#       self, db_path: str, block_records_path: str, num_block_records: int
#   ):
#     """Uses a file where all the retrievals have been made in advance.
#
#     Uses the exact retrievals from query_cacher_tfrecord.py, which have been made in
#     advance, as the questions don't change. The retrievals are made by
#     fetching
#     the pre-made retrievals by using the question-id in a lookup table.
#     The inner products are also present in the file; they are used to sample
#     from the pre-made retrievals to teach the model to adapt to having a wider
#     variety of retrievals each epoch.
#
#     Args:
#       db_path: Path to the hdf5 file that was generated with
#       `query_cacher_tfrecord.py`,
#         that contains the pre-made retrievals for all questions.
#       block_records_path: Path to the file with the reference
#       (often wikipedia)
#         text, that gets retrieved.
#       num_block_records: Number of entries in the reference db.
#     """
#     # Load the db
#
#     input_file = h5py.File(tf.io.gfile.GFile(db_path, "rb"), "r")
#     self._keys = ["train", "eval", "test"]
#
#     LOGGER.debug("Building the hash table")
#
#     self._indices_by_ids = {}
#     for split in self._keys:
#       self._indices_by_ids[split] = (
#           tf.lookup.StaticHashTable(
#               tf.lookup.KeyValueTensorInitializer(
#                   input_file[split]["sample_ids"],
#                   tf.range(input_file[split]["retrieval"][
#                   "indices"].shape[0])
#               ), 1))
#
#     LOGGER.debug("Building the self._distances_by_h5_index")
#     self._distances_by_h5_index = {
#         split: tf.constant(input_file[split]["retrieval"]["distances"][:])
#         for split in self._keys
#     }
#
#     LOGGER.debug("Building the self._db_entry_by_h5_index")
#     self._db_entry_by_h5_index = {
#         split: tf.constant(input_file[split]["retrieval"]["indices"][:])
#         for split in self._keys
#     }
#
#     with utils.log_duration(
#         LOGGER, "FullyCachedRetriever.__init__", "Load the textual dataset"
#     ):
#       # Extract the appropriate text
#       # The buffer_size is taken from the original ORQA code.
#       blocks_dataset = tf.data.TFRecordDataset(
#           block_records_path, buffer_size=512 * 1024 * 1024
#       )
#       blocks_dataset = blocks_dataset.batch(
#           num_block_records, drop_remainder=True
#       )
#       self._blocks = tf.data.experimental.get_single_element(blocks_dataset)
#
#   @tf.function
#   def retrieve(
#       self, ds_split: str, question_ids: tf.Tensor, temperature: float, k: int
#   ) -> tf.Tensor:
#     """Does the retrieving.
#
#     Args:
#       ds_split:
#         The h5 files are split per dataset split "train", "eval", "test". This
#         argument tells us which one to use.
#       question_ids: Id of the question. To be used to get the
#       cached retrievals.
#       temperature: Temperature to be used when sampling from the neighbors.
#       k: Number of neighbors to use.
#
#     Returns:
#       A dict with the logits and the retrieved reference text blocks.
#     """
#
#     indices = self._indices_by_ids[ds_split].lookup(question_ids)
#     distances = tf.gather(self._distances_by_h5_index[ds_split], indices)
#     db_indices = tf.gather(self._db_entry_by_h5_index[ds_split], indices)
#
#     # pick block ids
#     logits = distances / temperature
#     selections = tf_utils.sample_without_replacement(logits, k)
#     final_indices = tf.gather(db_indices, selections, batch_dims=-1)
#     # final_logits = tf.gather(logits, selections, batch_dims=-1)
#
#     retrieved_blocks = tf.gather(self._blocks, final_indices)
#     # utils.check_equal(final_logits.shape, final_indices.shape)
#     return retrieved_blocks


# RetrieverType = Union[
  # BERTScaNNRetriever,
  # FullyCachedRetriever
# ]