import pickle
import os
import tensorflow as tf
import numpy as np
from collections import defaultdict
from tqdm import tqdm
from capreolus.extractor import Extractor
from capreolus import Dependency, ConfigOption, get_logger
from capreolus.utils.common import padlist
from capreolus.utils.exceptions import MissingDocError
from capreolus.tokenizer.punkt import PunktTokenizer
[docs]logger = get_logger(__name__)
[docs]@Extractor.register
class BertPassage(Extractor):
"""
Extracts passages from the document to be later consumed by a BERT based model.
Does NOT use all the passages. The first passages is always used. Use the `prob` config to control the probability
of a passage being selected
Gotcha: In Tensorflow the train tfrecords have shape (batch_size, maxseqlen) while dev tf records have the shape
(batch_size, num_passages, maxseqlen). This is because during inference, we want to pool over the scores of the
passages belonging to a doc
"""
[docs] module_name = "bertpassage"
[docs] dependencies = [
Dependency(
key="index", module="index", name="anserini", default_config_overrides={"indexstops": True, "stemmer": "none"}
),
Dependency(key="tokenizer", module="tokenizer", name="berttokenizer"),
]
[docs] config_spec = [
ConfigOption("maxseqlen", 256, "Maximum input length for BERT"),
ConfigOption("usecache", False, "Should the extracted features be cached?"),
ConfigOption("passagelen", 150, "Length of the extracted passage"),
ConfigOption("stride", 100, "Stride"),
ConfigOption("sentences", False, "Use a sentence tokenizer to form passages"),
ConfigOption("numpassages", 16, "Number of passages per document"),
ConfigOption(
"prob",
0.1,
"The probability that a passage from the document will be used for training (the first passage is always used)",
),
]
[docs] def load_state(self, qids, docids):
cache_fn = self.get_state_cache_file_path(qids, docids)
logger.debug("loading state from: %s", cache_fn)
with open(cache_fn, "rb") as f:
state_dict = pickle.load(f)
self.qid2toks = state_dict["qid2toks"]
self.docid2passages = state_dict["docid2passages"]
[docs] def cache_state(self, qids, docids):
os.makedirs(self.get_cache_path(), exist_ok=True)
with open(self.get_state_cache_file_path(qids, docids), "wb") as f:
state_dict = {"qid2toks": self.qid2toks, "docid2passages": self.docid2passages}
pickle.dump(state_dict, f, protocol=-1)
[docs] def get_tf_feature_description(self):
feature_description = {
"pos_bert_input": tf.io.FixedLenFeature([], tf.string),
"pos_mask": tf.io.FixedLenFeature([], tf.string),
"pos_seg": tf.io.FixedLenFeature([], tf.string),
"neg_bert_input": tf.io.FixedLenFeature([], tf.string),
"neg_mask": tf.io.FixedLenFeature([], tf.string),
"neg_seg": tf.io.FixedLenFeature([], tf.string),
"label": tf.io.FixedLenFeature([], tf.string),
}
return feature_description
[docs] def create_tf_train_feature(self, sample):
"""
Returns a set of features from a doc.
Of the num_passages passages that are present in a document, we use only a subset of it.
params:
sample - A dict where each entry has the shape [batch_size, num_passages, maxseqlen]
Returns a list of features. Each feature is a dict, and each value in the dict has the shape [batch_size, maxseqlen].
Yes, the output shape is different to the input shape because we sample from the passages.
"""
num_passages = self.config["numpassages"]
def _bytes_feature(value):
"""Returns a bytes_list from a string / byte. Our features are multi-dimensional tensors."""
if isinstance(value, type(tf.constant(0))): # if value ist tensor
value = value.numpy() # get value of tensor
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
posdoc, negdoc, negdoc_id = sample["pos_bert_input"], sample["neg_bert_input"], sample["negdocid"]
posdoc_mask, posdoc_seg, negdoc_mask, negdoc_seg = (
sample["pos_mask"],
sample["pos_seg"],
sample["neg_mask"],
sample["neg_seg"],
)
label = sample["label"]
features = []
for i in range(num_passages):
# Always use the first passage, then sample from the remaining passages
if i > 0 and self.rng.random() > self.config["prob"]:
continue
bert_input_line = posdoc[i]
bert_input_line = " ".join(self.tokenizer.bert_tokenizer.convert_ids_to_tokens(list(bert_input_line)))
passage = bert_input_line.split("[SEP]")[-2]
# Ignore empty passages as well
if passage.strip() == "[PAD]":
continue
feature = {
"pos_bert_input": _bytes_feature(tf.io.serialize_tensor(posdoc[i])),
"pos_mask": _bytes_feature(tf.io.serialize_tensor(posdoc_mask[i])),
"pos_seg": _bytes_feature(tf.io.serialize_tensor(posdoc_seg[i])),
"neg_bert_input": _bytes_feature(tf.io.serialize_tensor(negdoc[i])),
"neg_mask": _bytes_feature(tf.io.serialize_tensor(negdoc_mask[i])),
"neg_seg": _bytes_feature(tf.io.serialize_tensor(negdoc_seg[i])),
"label": _bytes_feature(tf.io.serialize_tensor(label[i])),
}
features.append(feature)
return features
[docs] def create_tf_dev_feature(self, sample):
"""
Unlike the train feature, the dev set uses all passages. Both the input and the output are dicts with the shape
[batch_size, num_passages, maxseqlen]
"""
posdoc, negdoc, negdoc_id = sample["pos_bert_input"], sample["neg_bert_input"], sample["negdocid"]
posdoc_mask, posdoc_seg, negdoc_mask, negdoc_seg = (
sample["pos_mask"],
sample["pos_seg"],
sample["neg_mask"],
sample["neg_seg"],
)
label = sample["label"]
def _bytes_feature(value):
"""Returns a bytes_list from a string / byte."""
if isinstance(value, type(tf.constant(0))): # if value ist tensor
value = value.numpy() # get value of tensor
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
feature = {
"pos_bert_input": _bytes_feature(tf.io.serialize_tensor(posdoc)),
"pos_mask": _bytes_feature(tf.io.serialize_tensor(posdoc_mask)),
"pos_seg": _bytes_feature(tf.io.serialize_tensor(posdoc_seg)),
"neg_bert_input": _bytes_feature(tf.io.serialize_tensor(negdoc)),
"neg_mask": _bytes_feature(tf.io.serialize_tensor(negdoc_mask)),
"neg_seg": _bytes_feature(tf.io.serialize_tensor(negdoc_seg)),
"label": _bytes_feature(tf.io.serialize_tensor(label)),
}
return [feature]
[docs] def parse_tf_train_example(self, example_proto):
feature_description = self.get_tf_feature_description()
parsed_example = tf.io.parse_example(example_proto, feature_description)
def parse_tensor_as_int(x):
parsed_tensor = tf.io.parse_tensor(x, tf.int64)
parsed_tensor.set_shape([self.config["maxseqlen"]])
return parsed_tensor
def parse_label_tensor(x):
parsed_tensor = tf.io.parse_tensor(x, tf.float32)
parsed_tensor.set_shape([2])
return parsed_tensor
pos_bert_input = tf.map_fn(parse_tensor_as_int, parsed_example["pos_bert_input"], dtype=tf.int64)
pos_mask = tf.map_fn(parse_tensor_as_int, parsed_example["pos_mask"], dtype=tf.int64)
pos_seg = tf.map_fn(parse_tensor_as_int, parsed_example["pos_seg"], dtype=tf.int64)
neg_bert_input = tf.map_fn(parse_tensor_as_int, parsed_example["neg_bert_input"], dtype=tf.int64)
neg_mask = tf.map_fn(parse_tensor_as_int, parsed_example["neg_mask"], dtype=tf.int64)
neg_seg = tf.map_fn(parse_tensor_as_int, parsed_example["neg_seg"], dtype=tf.int64)
label = tf.map_fn(parse_label_tensor, parsed_example["label"], dtype=tf.float32)
return (pos_bert_input, pos_mask, pos_seg, neg_bert_input, neg_mask, neg_seg), label
[docs] def parse_tf_dev_example(self, example_proto):
feature_description = self.get_tf_feature_description()
parsed_example = tf.io.parse_example(example_proto, feature_description)
def parse_tensor_as_int(x):
parsed_tensor = tf.io.parse_tensor(x, tf.int64)
parsed_tensor.set_shape([self.config["numpassages"], self.config["maxseqlen"]])
return parsed_tensor
def parse_label_tensor(x):
parsed_tensor = tf.io.parse_tensor(x, tf.float32)
parsed_tensor.set_shape([self.config["numpassages"], 2])
return parsed_tensor
pos_bert_input = tf.map_fn(parse_tensor_as_int, parsed_example["pos_bert_input"], dtype=tf.int64)
pos_mask = tf.map_fn(parse_tensor_as_int, parsed_example["pos_mask"], dtype=tf.int64)
pos_seg = tf.map_fn(parse_tensor_as_int, parsed_example["pos_seg"], dtype=tf.int64)
neg_bert_input = tf.map_fn(parse_tensor_as_int, parsed_example["neg_bert_input"], dtype=tf.int64)
neg_mask = tf.map_fn(parse_tensor_as_int, parsed_example["neg_mask"], dtype=tf.int64)
neg_seg = tf.map_fn(parse_tensor_as_int, parsed_example["neg_seg"], dtype=tf.int64)
label = tf.map_fn(parse_label_tensor, parsed_example["label"], dtype=tf.float32)
return (pos_bert_input, pos_mask, pos_seg, neg_bert_input, neg_mask, neg_seg), label
[docs] def get_passages_for_doc(self, doc):
"""
Extract passages from the doc.
If there are too many passages, keep the first and the last one and sample from the rest.
If there are not enough packages, pad.
"""
tokenize = self.tokenizer.tokenize
numpassages = self.config["numpassages"]
passages = []
for i in range(0, len(doc), self.config["stride"]):
if i >= len(doc):
assert len(passages) > 0, f"no passage can be built from empty document {doc}"
break
else:
passage = doc[i : i + self.config["passagelen"]]
passages.append(tokenize(" ".join(passage)))
n_actual_passages = len(passages)
# If we have a more passages than required, keep the first and last, and sample from the rest
if n_actual_passages > numpassages:
if numpassages > 1:
passages = [passages[0]] + list(self.rng.choice(passages[1:-1], numpassages - 2, replace=False)) + [passages[-1]]
else:
passages = [passages[0]]
else:
# Pad until we have the required number of passages
for _ in range(numpassages - n_actual_passages):
passages.append(["[PAD]"])
assert len(passages) == self.config["numpassages"]
return passages
# from https://github.com/castorini/birch/blob/2dd0401ebb388a1c96f8f3357a064164a5db3f0e/src/utils/doc_utils.py#L73
def _chunk_sent(self, sent, max_len):
words = self.tokenizer.tokenize(sent)
if len(words) <= max_len:
return [words]
chunked_sents = []
size = int(len(words) / max_len)
for i in range(0, size):
seq = words[i * max_len : (i + 1) * max_len]
chunked_sents.append(seq)
return chunked_sents
def _build_passages_from_sentences(self, docids):
punkt = PunktTokenizer()
for docid in tqdm(docids, "extract passages"):
passages = []
numpassages = self.config["numpassages"]
for sentence in punkt.tokenize(self.index.get_doc(docid)):
if len(passages) >= numpassages:
break
passages.extend(self._chunk_sent(sentence, self.config["passagelen"]))
if numpassages != 0:
passages = passages[:numpassages]
n_actual_passages = len(passages)
for _ in range(numpassages - n_actual_passages):
# randomly use one of previous passages when the document is exhausted
# idx = random.randint(0, n_actual_passages - 1)
# passages.append(passages[idx])
# append empty passages
passages.append([""])
assert len(passages) == self.config["numpassages"]
self.docid2passages[docid] = sorted(passages, key=len)
def _build_vocab(self, qids, docids, topics):
if self.is_state_cached(qids, docids) and self.config["usecache"]:
self.load_state(qids, docids)
logger.info("Vocabulary loaded from cache")
elif self.config["sentences"]:
self.docid2passages = {}
self._build_passages_from_sentences(docids)
self.qid2toks = {qid: self.tokenizer.tokenize(topics[qid]) for qid in tqdm(qids, desc="querytoks")}
self.cache_state(qids, docids)
else:
logger.info("Building bertpassage vocabulary")
self.docid2passages = {}
for docid in tqdm(docids, "extract passages"):
# Naive tokenization based on white space
doc = self.index.get_doc(docid).split()
passages = self.get_passages_for_doc(doc)
self.docid2passages[docid] = passages
self.qid2toks = {qid: self.tokenizer.tokenize(topics[qid]) for qid in tqdm(qids, desc="querytoks")}
self.cache_state(qids, docids)
[docs] def exist(self):
return hasattr(self, "docid2passages") and len(self.docid2passages)
[docs] def preprocess(self, qids, docids, topics):
if self.exist():
return
self.index.create_index()
self.qid2toks = defaultdict(list)
self.docid2passages = None
self._build_vocab(qids, docids, topics)
[docs] def id2vec(self, qid, posid, negid=None, label=None):
"""
See parent class for docstring
"""
assert label is not None
tokenizer = self.tokenizer
maxseqlen = self.config["maxseqlen"]
query_toks = self.qid2toks[qid]
pos_bert_inputs = []
pos_bert_masks = []
pos_bert_segs = []
# N.B: The passages in self.docid2passages are not bert tokenized
pos_passages = self.docid2passages[posid]
for tokenized_passage in pos_passages:
input_line = ["[CLS]"] + query_toks + ["[SEP]"] + tokenized_passage + ["[SEP]"]
if len(input_line) > maxseqlen:
input_line = input_line[:maxseqlen]
input_line[-1] = "[SEP]"
padded_input_line = padlist(input_line, padlen=self.config["maxseqlen"], pad_token=self.pad_tok)
pos_bert_masks.append([1] * len(input_line) + [0] * (len(padded_input_line) - len(input_line)))
pos_bert_segs.append([0] * (len(query_toks) + 2) + [1] * (len(padded_input_line) - len(query_toks) - 2))
pos_bert_inputs.append(tokenizer.convert_tokens_to_ids(padded_input_line))
# TODO: Rename the posdoc key in the below dict to 'pos_bert_input'
data = {
"qid": qid,
"posdocid": posid,
"pos_bert_input": np.array(pos_bert_inputs, dtype=np.long),
"pos_mask": np.array(pos_bert_masks, dtype=np.long),
"pos_seg": np.array(pos_bert_segs, dtype=np.long),
"negdocid": "",
"neg_bert_input": np.zeros((self.config["numpassages"], self.config["maxseqlen"]), dtype=np.long),
"neg_mask": np.zeros((self.config["numpassages"], self.config["maxseqlen"]), dtype=np.long),
"neg_seg": np.zeros((self.config["numpassages"], self.config["maxseqlen"]), dtype=np.long),
"label": np.repeat(np.array([label], dtype=np.float32), self.config["numpassages"], 0),
}
if negid:
neg_bert_inputs = []
neg_bert_masks = []
neg_bert_segs = []
neg_passages = self.docid2passages[negid]
for tokenized_passage in neg_passages:
input_line = ["[CLS]"] + query_toks + ["[SEP]"] + tokenized_passage + ["[SEP]"]
if len(input_line) > maxseqlen:
input_line = input_line[:maxseqlen]
input_line[-1] = "[SEP]"
padded_input_line = padlist(input_line, padlen=self.config["maxseqlen"], pad_token=self.pad_tok)
neg_bert_masks.append([1] * len(input_line) + [0] * (len(padded_input_line) - len(input_line)))
neg_bert_segs.append([0] * (len(query_toks) + 2) + [1] * (len(padded_input_line) - len(query_toks) - 2))
neg_bert_inputs.append(tokenizer.convert_tokens_to_ids(padded_input_line))
if not neg_bert_inputs:
raise MissingDocError(qid, negid)
data["negdocid"] = negid
data["neg_bert_input"] = np.array(neg_bert_inputs, dtype=np.long)
data["neg_mask"] = np.array(neg_bert_masks, dtype=np.long)
data["neg_seg"] = np.array(neg_bert_segs, dtype=np.long)
return data