# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved. # 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. from __future__ import absolute_import from __future__ import division from __future__ import print_function import collections import json import math import os import random import modeling import optimization import tokenization import six import tensorflow as tf import horovod.tensorflow as hvd import time flags = tf.flags FLAGS = None def extract_flags(): flags.DEFINE_integer( "max_seq_length", 384, "The maximum total input sequence length after WordPiece tokenization. " "Sequences longer than this will be truncated, and sequences shorter " "than this will be padded.") flags.DEFINE_integer( "doc_stride", 128, "When splitting up a long document into chunks, how much stride to " "take between chunks.") flags.DEFINE_integer( "max_query_length", 64, "The maximum number of tokens for the question. Questions longer than " "this will be truncated to this length.") flags.DEFINE_bool( "version_2_with_negative", False, "If true, the SQuAD examples contain some that do not have an answer.") flags.DEFINE_string("train_file", None, "SQuAD json for training. E.g., train-v1.1.json") flags.DEFINE_string( "predict_file", None, "SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json") flags.DEFINE_string( "squad_dir", None, "The output directory where the model checkpoints will be written.") flags.DEFINE_string("vocab_file", None, "The vocabulary file that the BERT model was trained on.") flags.DEFINE_bool( "do_lower_case", True, "Whether to lower case the input text. Should be True for uncased " "models and False for cased models.") flags.DEFINE_bool( "verbose_logging", False, "If true, all of the warnings related to data processing will be printed. " "A number of warnings are expected for a normal SQuAD evaluation.") flags.mark_flag_as_required("train_file") flags.mark_flag_as_required("predict_file") flags.mark_flag_as_required("squad_dir") flags.mark_flag_as_required("vocab_file") return flags.FLAGS class SquadExample(object): """A single training/test example for simple sequence classification. For examples without an answer, the start and end position are -1. """ def __init__(self, qas_id, question_text, doc_tokens, orig_answer_text=None, start_position=None, end_position=None, is_impossible=False): self.qas_id = qas_id self.question_text = question_text self.doc_tokens = doc_tokens self.orig_answer_text = orig_answer_text self.start_position = start_position self.end_position = end_position self.is_impossible = is_impossible def __str__(self): return self.__repr__() def __repr__(self): s = "" s += "qas_id: %s" % (tokenization.printable_text(self.qas_id)) s += ", question_text: %s" % ( tokenization.printable_text(self.question_text)) s += ", doc_tokens: [%s]" % (" ".join(self.doc_tokens)) if self.start_position: s += ", start_position: %d" % (self.start_position) if self.start_position: s += ", end_position: %d" % (self.end_position) if self.start_position: s += ", is_impossible: %r" % (self.is_impossible) return s class InputFeatures(object): """A single set of features of data.""" def __init__(self, unique_id, example_index, doc_span_index, tokens, token_to_orig_map, token_is_max_context, input_ids, input_mask, segment_ids, start_position=None, end_position=None, is_impossible=None): self.unique_id = unique_id self.example_index = example_index self.doc_span_index = doc_span_index self.tokens = tokens self.token_to_orig_map = token_to_orig_map self.token_is_max_context = token_is_max_context self.input_ids = input_ids self.input_mask = input_mask self.segment_ids = segment_ids self.start_position = start_position self.end_position = end_position self.is_impossible = is_impossible def read_squad_examples(input_file, is_training, version_2_with_negative=False, input_data=None): """Return list of SquadExample from input_data or input_file (SQuAD json file)""" if input_data is None: with tf.gfile.Open(input_file, "r") as reader: input_data = json.load(reader)["data"] def is_whitespace(c): if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F: return True return False examples = [] for entry in input_data: for paragraph in entry["paragraphs"]: paragraph_text = paragraph["context"] doc_tokens = [] char_to_word_offset = [] prev_is_whitespace = True for c in paragraph_text: if is_whitespace(c): prev_is_whitespace = True else: if prev_is_whitespace: doc_tokens.append(c) else: doc_tokens[-1] += c prev_is_whitespace = False char_to_word_offset.append(len(doc_tokens) - 1) for qa in paragraph["qas"]: qas_id = qa["id"] question_text = qa["question"] start_position = None end_position = None orig_answer_text = None is_impossible = False if is_training: if version_2_with_negative: is_impossible = qa["is_impossible"] if (len(qa["answers"]) != 1) and (not is_impossible): raise ValueError( "For training, each question should have exactly 1 answer.") if not is_impossible: answer = qa["answers"][0] orig_answer_text = answer["text"] answer_offset = answer["answer_start"] answer_length = len(orig_answer_text) start_position = char_to_word_offset[answer_offset] end_position = char_to_word_offset[answer_offset + answer_length - 1] # Only add answers where the text can be exactly recovered from the # document. If this CAN'T happen it's likely due to weird Unicode # stuff so we will just skip the example. # # Note that this means for training mode, every example is NOT # guaranteed to be preserved. actual_text = " ".join( doc_tokens[start_position:(end_position + 1)]) cleaned_answer_text = " ".join( tokenization.whitespace_tokenize(orig_answer_text)) if actual_text.find(cleaned_answer_text) == -1: tf.logging.warning("Could not find answer: '%s' vs. '%s'", actual_text, cleaned_answer_text) continue else: start_position = -1 end_position = -1 orig_answer_text = "" example = SquadExample( qas_id=qas_id, question_text=question_text, doc_tokens=doc_tokens, orig_answer_text=orig_answer_text, start_position=start_position, end_position=end_position, is_impossible=is_impossible) examples.append(example) return examples def _check_is_max_context(doc_spans, cur_span_index, position): """Check if this is the 'max context' doc span for the token.""" # Because of the sliding window approach taken to scoring documents, a single # token can appear in multiple documents. E.g. # Doc: the man went to the store and bought a gallon of milk # Span A: the man went to the # Span B: to the store and bought # Span C: and bought a gallon of # ... # # Now the word 'bought' will have two scores from spans B and C. We only # want to consider the score with "maximum context", which we define as # the *minimum* of its left and right context (the *sum* of left and # right context will always be the same, of course). # # In the example the maximum context for 'bought' would be span C since # it has 1 left context and 3 right context, while span B has 4 left context # and 0 right context. best_score = None best_span_index = None for (span_index, doc_span) in enumerate(doc_spans): end = doc_span.start + doc_span.length - 1 if position < doc_span.start: continue if position > end: continue num_left_context = position - doc_span.start num_right_context = end - position score = min(num_left_context, num_right_context) + 0.01 * doc_span.length if best_score is None or score > best_score: best_score = score best_span_index = span_index return cur_span_index == best_span_index def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text): """Returns tokenized answer spans that better match the annotated answer.""" # The SQuAD annotations are character based. We first project them to # whitespace-tokenized words. But then after WordPiece tokenization, we can # often find a "better match". For example: # # Question: What year was John Smith born? # Context: The leader was John Smith (1895-1943). # Answer: 1895 # # The original whitespace-tokenized answer will be "(1895-1943).". However # after tokenization, our tokens will be "( 1895 - 1943 ) .". So we can match # the exact answer, 1895. # # However, this is not always possible. Consider the following: # # Question: What country is the top exporter of electornics? # Context: The Japanese electronics industry is the lagest in the world. # Answer: Japan # # In this case, the annotator chose "Japan" as a character sub-span of # the word "Japanese". Since our WordPiece tokenizer does not split # "Japanese", we just use "Japanese" as the annotation. This is fairly rare # in SQuAD, but does happen. tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text)) for new_start in range(input_start, input_end + 1): for new_end in range(input_end, new_start - 1, -1): text_span = " ".join(doc_tokens[new_start:(new_end + 1)]) if text_span == tok_answer_text: return (new_start, new_end) return (input_start, input_end) def convert_examples_to_features(examples, tokenizer, max_seq_length, doc_stride, max_query_length, is_training, output_fn, verbose_logging=False): """Loads a data file into a list of `InputBatch`s.""" unique_id = 1000000000 for (example_index, example) in enumerate(examples): query_tokens = tokenizer.tokenize(example.question_text) if len(query_tokens) > max_query_length: query_tokens = query_tokens[0:max_query_length] tok_to_orig_index = [] orig_to_tok_index = [] all_doc_tokens = [] for (i, token) in enumerate(example.doc_tokens): orig_to_tok_index.append(len(all_doc_tokens)) sub_tokens = tokenizer.tokenize(token) for sub_token in sub_tokens: tok_to_orig_index.append(i) all_doc_tokens.append(sub_token) tok_start_position = None tok_end_position = None if is_training and example.is_impossible: tok_start_position = -1 tok_end_position = -1 if is_training and not example.is_impossible: tok_start_position = orig_to_tok_index[example.start_position] if example.end_position < len(example.doc_tokens) - 1: tok_end_position = orig_to_tok_index[example.end_position + 1] - 1 else: tok_end_position = len(all_doc_tokens) - 1 (tok_start_position, tok_end_position) = _improve_answer_span( all_doc_tokens, tok_start_position, tok_end_position, tokenizer, example.orig_answer_text) # The -3 accounts for [CLS], [SEP] and [SEP] max_tokens_for_doc = max_seq_length - len(query_tokens) - 3 # We can have documents that are longer than the maximum sequence length. # To deal with this we do a sliding window approach, where we take chunks # of the up to our max length with a stride of `doc_stride`. _DocSpan = collections.namedtuple( # pylint: disable=invalid-name "DocSpan", ["start", "length"]) doc_spans = [] start_offset = 0 while start_offset < len(all_doc_tokens): length = len(all_doc_tokens) - start_offset if length > max_tokens_for_doc: length = max_tokens_for_doc doc_spans.append(_DocSpan(start=start_offset, length=length)) if start_offset + length == len(all_doc_tokens): break start_offset += min(length, doc_stride) for (doc_span_index, doc_span) in enumerate(doc_spans): tokens = [] token_to_orig_map = {} token_is_max_context = {} segment_ids = [] tokens.append("[CLS]") segment_ids.append(0) for token in query_tokens: tokens.append(token) segment_ids.append(0) tokens.append("[SEP]") segment_ids.append(0) for i in range(doc_span.length): split_token_index = doc_span.start + i token_to_orig_map[len(tokens)] = tok_to_orig_index[split_token_index] is_max_context = _check_is_max_context(doc_spans, doc_span_index, split_token_index) token_is_max_context[len(tokens)] = is_max_context tokens.append(all_doc_tokens[split_token_index]) segment_ids.append(1) tokens.append("[SEP]") segment_ids.append(1) input_ids = tokenizer.convert_tokens_to_ids(tokens) # The mask has 1 for real tokens and 0 for padding tokens. Only real # tokens are attended to. input_mask = [1] * len(input_ids) # Zero-pad up to the sequence length. while len(input_ids) < max_seq_length: input_ids.append(0) input_mask.append(0) segment_ids.append(0) assert len(input_ids) == max_seq_length assert len(input_mask) == max_seq_length assert len(segment_ids) == max_seq_length start_position = None end_position = None if is_training and not example.is_impossible: # For training, if our document chunk does not contain an annotation # we throw it out, since there is nothing to predict. doc_start = doc_span.start doc_end = doc_span.start + doc_span.length - 1 out_of_span = False if not (tok_start_position >= doc_start and tok_end_position <= doc_end): out_of_span = True if out_of_span: start_position = 0 end_position = 0 else: doc_offset = len(query_tokens) + 2 start_position = tok_start_position - doc_start + doc_offset end_position = tok_end_position - doc_start + doc_offset if is_training and example.is_impossible: start_position = 0 end_position = 0 if verbose_logging and example_index < 20: tf.compat.v1.logging.info("*** Example ***") tf.compat.v1.logging.info("unique_id: %s" % (unique_id)) tf.compat.v1.logging.info("example_index: %s" % (example_index)) tf.compat.v1.logging.info("doc_span_index: %s" % (doc_span_index)) tf.compat.v1.logging.info("tokens: %s" % " ".join( [tokenization.printable_text(x) for x in tokens])) tf.compat.v1.logging.info("token_to_orig_map: %s" % " ".join( ["%d:%d" % (x, y) for (x, y) in six.iteritems(token_to_orig_map)])) tf.compat.v1.logging.info("token_is_max_context: %s" % " ".join([ "%d:%s" % (x, y) for (x, y) in six.iteritems(token_is_max_context) ])) tf.compat.v1.logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) tf.compat.v1.logging.info( "input_mask: %s" % " ".join([str(x) for x in input_mask])) tf.compat.v1.logging.info( "segment_ids: %s" % " ".join([str(x) for x in segment_ids])) if is_training and example.is_impossible: tf.compat.v1.logging.info("impossible example") if is_training and not example.is_impossible: answer_text = " ".join(tokens[start_position:(end_position + 1)]) tf.compat.v1.logging.info("start_position: %d" % (start_position)) tf.compat.v1.logging.info("end_position: %d" % (end_position)) tf.compat.v1.logging.info( "answer: %s" % (tokenization.printable_text(answer_text))) feature = InputFeatures( unique_id=unique_id, example_index=example_index, doc_span_index=doc_span_index, tokens=tokens, token_to_orig_map=token_to_orig_map, token_is_max_context=token_is_max_context, input_ids=input_ids, input_mask=input_mask, segment_ids=segment_ids, start_position=start_position, end_position=end_position, is_impossible=example.is_impossible) # Run callback output_fn(feature) unique_id += 1 class FeatureWriter(object): """Writes InputFeature to TF example file.""" def __init__(self, filename, is_training): self.filename = filename self.is_training = is_training self.num_features = 0 self._writer = tf.python_io.TFRecordWriter(filename) def process_feature(self, feature): """Write a InputFeature to the TFRecordWriter as a tf.train.Example.""" self.num_features += 1 def create_int_feature(values): feature = tf.train.Feature( int64_list=tf.train.Int64List(value=list(values))) return feature features = collections.OrderedDict() features["unique_ids"] = create_int_feature([feature.unique_id]) features["input_ids"] = create_int_feature(feature.input_ids) features["input_mask"] = create_int_feature(feature.input_mask) features["segment_ids"] = create_int_feature(feature.segment_ids) if self.is_training: features["start_positions"] = create_int_feature([feature.start_position]) features["end_positions"] = create_int_feature([feature.end_position]) impossible = 0 if feature.is_impossible: impossible = 1 features["is_impossible"] = create_int_feature([impossible]) tf_example = tf.train.Example(features=tf.train.Features(feature=features)) self._writer.write(tf_example.SerializeToString()) def close(self): self._writer.close() def main(): FLAGS = extract_flags() tokenizer = tokenization.FullTokenizer( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case) tf.gfile.MakeDirs(FLAGS.squad_dir + "/final_tfrecords_sharded") # We write to a temporary file to avoid storing very large constant tensors # in memory. train_examples = read_squad_examples( input_file=FLAGS.train_file, is_training=True, version_2_with_negative=FLAGS.version_2_with_negative) train_writer = FeatureWriter( filename=os.path.join(FLAGS.squad_dir, "final_tfrecords_sharded/train.tf_record"), is_training=True) convert_examples_to_features( examples=train_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=True, output_fn=train_writer.process_feature, verbose_logging=FLAGS.verbose_logging) train_writer.close() eval_examples = read_squad_examples( input_file=FLAGS.predict_file, is_training=False, version_2_with_negative=FLAGS.version_2_with_negative) eval_writer = FeatureWriter( filename=os.path.join(FLAGS.squad_dir, "final_tfrecords_sharded/eval.tf_record"), is_training=False) eval_features = [] def append_feature(feature): eval_features.append(feature) eval_writer.process_feature(feature) convert_examples_to_features( examples=eval_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=False, output_fn=append_feature, verbose_logging=FLAGS.verbose_logging) eval_writer.close() if __name__ == "__main__": main()