Convert flags to argparse in 'run_squad_pytorch.py'

run_squad_pytorch.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team 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.
+"""Run BERT on SQuAD."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import json
+import math
+import os
+import modeling
+import optimization
+import tokenization
+import six
+import tensorflow as tf
+import argparse
+
+parser = argparse.ArgumentParser()
+
+## Required parameters
+parser.add_argument("--bert_config_file", default=None, type=str, required=True,
+                    help="The config json file corresponding to the pre-trained BERT model. " 
+                         "This specifies the model architecture.")
+parser.add_argument("--vocab_file", default=None, type=str, required=True,
+                    help="The vocabulary file that the BERT model was trained on.")
+parser.add_argument("--output_dir", default=None, type=str, required=True,
+                    help="The output directory where the model checkpoints will be written.")
+
+## Other parameters
+parser.add_argument("--train_file", default=None, type=str, help="SQuAD json for training. E.g., train-v1.1.json")
+parser.add_argument("--predict_file", default=None, type=str,
+                    help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
+parser.add_argument("--init_checkpoint", default=None, type=str,
+                    help="Initial checkpoint (usually from a pre-trained BERT model).")
+parser.add_argument("--do_lower_case", default=True, type=bool,
+                    help="Whether to lower case the input text. Should be True for uncased "
+                         "models and False for cased models.")
+parser.add_argument("--max_seq_length", default=384, type=int,
+                    help="The maximum total input sequence length after WordPiece tokenization. "
+                         "Sequences longer than this will be truncated, and sequences shorter than this will be padded.")
+parser.add_argument("--doc_stride", default=128, type=int,
+                    help="When splitting up a long document into chunks, how much stride to take between chunks.")
+parser.add_argument("--max_query_length", default=64, type=int,
+                    help="The maximum number of tokens for the question. Questions longer than this will "
+                         "be truncated to this length.")
+parser.add_argument("--do_train", default=False, type=bool, help="Whether to run training.")
+parser.add_argument("--do_predict", default=False, type=bool, help="Whether to run eval on the dev set.")
+parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.")
+parser.add_argument("--predict_batch_size", default=8, type=int, help="Total batch size for predictions.")
+parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.")
+parser.add_argument("--warmup_proportion", default=0.1, type=float,
+                    help="Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10% "
+                         "of training.")
+parser.add_argument("--save_checkpoints_steps", default=1000, type=int, help="How often to save the model checkpoint.")
+parser.add_argument("--iterations_per_loop", default=1000, type=int,
+                    help="How many steps to make in each estimator call.")
+parser.add_argument("--n_best_size", default=20, type=int,
+                    help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
+parser.add_argument("--max_answer_length", default=30, type=int,
+                    help="The maximum length of an answer that can be generated. This is needed because the start "
+                         "and end predictions are not conditioned on one another.")
+
+### BEGIN - TO DELETE EVENTUALLY --> NO SENSE IN PYTORCH ###
+parser.add_argument("--use_tpu", default=False, type=bool, help="Whether to use TPU or GPU/CPU.")
+parser.add_argument("--tpu_name", default=None, type=str,
+                    help="The Cloud TPU to use for training. This should be either the name used when creating the "
+                         "Cloud TPU, or a grpc://ip.address.of.tpu:8470 url.")
+parser.add_argument("--tpu_zone", default=None, type=str,
+                    help="[Optional] GCE zone where the Cloud TPU is located in. If not specified, we will attempt "
+                         "to automatically detect the GCE project from metadata.")
+parser.add_argument("--gcp_project", default=None, type=str,
+                    help="[Optional] Project name for the Cloud TPU-enabled project. If not specified, we will attempt "
+                         "to automatically detect the GCE project from metadata.")
+parser.add_argument("--master", default=None, type=str, help="[Optional] TensorFlow master URL.")
+parser.add_argument("--num_tpu_cores", default=8, type=int, help="Only used if `use_tpu` is True. "
+                                                                 "Total number of TPU cores to use.")
+### END - TO DELETE EVENTUALLY --> NO SENSE IN PYTORCH ###
+
+parser.add_argument("--verbose_logging", default=False, type=bool,
+                    help="If true, all of the warnings related to data processing will be printed. "
+                         "A number of warnings are expected for a normal SQuAD evaluation.")
+
+args = parser.parse_args()
+
+
+class SquadExample(object):
+    """A single training/test example for simple sequence classification."""
+
+    def __init__(self,
+                 qas_id,
+                 question_text,
+                 doc_tokens,
+                 orig_answer_text=None,
+                 start_position=None,
+                 end_position=None):
+        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
+
+    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)
+        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):
+        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
+
+
+def read_squad_examples(input_file, is_training):
+    """Read a SQuAD json file into a list of SquadExample."""
+    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
+                if is_training:
+                    if len(qa["answers"]) != 1:
+                        raise ValueError(
+                            "For training, each question should have exactly 1 answer.")
+                    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
+
+                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)
+                examples.append(example)
+    return examples
+
+
+def convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                 doc_stride, max_query_length, is_training):
+    """Loads a data file into a list of `InputBatch`s."""
+
+    unique_id = 1000000000
+
+    features = []
+    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:
+            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:
+                # 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
+                if (example.start_position < doc_start or
+                        example.end_position < doc_start or
+                        example.start_position > doc_end or example.end_position > doc_end):
+                    continue
+
+                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 example_index < 20:
+                tf.logging.info("*** Example ***")
+                tf.logging.info("unique_id: %s" % (unique_id))
+                tf.logging.info("example_index: %s" % (example_index))
+                tf.logging.info("doc_span_index: %s" % (doc_span_index))
+                tf.logging.info("tokens: %s" % " ".join(
+                    [tokenization.printable_text(x) for x in tokens]))
+                tf.logging.info("token_to_orig_map: %s" % " ".join(
+                    ["%d:%d" % (x, y) for (x, y) in six.iteritems(token_to_orig_map)]))
+                tf.logging.info("token_is_max_context: %s" % " ".join([
+                    "%d:%s" % (x, y) for (x, y) in six.iteritems(token_is_max_context)
+                ]))
+                tf.logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+                tf.logging.info(
+                    "input_mask: %s" % " ".join([str(x) for x in input_mask]))
+                tf.logging.info(
+                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
+                if is_training:
+                    answer_text = " ".join(tokens[start_position:(end_position + 1)])
+                    tf.logging.info("start_position: %d" % (start_position))
+                    tf.logging.info("end_position: %d" % (end_position))
+                    tf.logging.info(
+                        "answer: %s" % (tokenization.printable_text(answer_text)))
+
+            features.append(
+                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))
+            unique_id += 1
+
+    return features
+
+
+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 _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 create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
+                 use_one_hot_embeddings):
+    """Creates a classification model."""
+    model = modeling.BertModel(
+        config=bert_config,
+        is_training=is_training,
+        input_ids=input_ids,
+        input_mask=input_mask,
+        token_type_ids=segment_ids,
+        use_one_hot_embeddings=use_one_hot_embeddings)
+
+    final_hidden = model.get_sequence_output()
+
+    final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
+    batch_size = final_hidden_shape[0]
+    seq_length = final_hidden_shape[1]
+    hidden_size = final_hidden_shape[2]
+
+    output_weights = tf.get_variable(
+        "cls/squad/output_weights", [2, hidden_size],
+        initializer=tf.truncated_normal_initializer(stddev=0.02))
+
+    output_bias = tf.get_variable(
+        "cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
+
+    final_hidden_matrix = tf.reshape(final_hidden,
+                                     [batch_size * seq_length, hidden_size])
+    logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
+    logits = tf.nn.bias_add(logits, output_bias)
+
+    logits = tf.reshape(logits, [batch_size, seq_length, 2])
+    logits = tf.transpose(logits, [2, 0, 1])
+
+    unstacked_logits = tf.unstack(logits, axis=0)
+
+    (start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
+
+    return (start_logits, end_logits)
+
+
+def model_fn_builder(bert_config, init_checkpoint, learning_rate,
+                     num_train_steps, num_warmup_steps, use_tpu,
+                     use_one_hot_embeddings):
+    """Returns `model_fn` closure for TPUEstimator."""
+
+    def model_fn(features, labels, mode, params):  # pylint: disable=unused-argument
+        """The `model_fn` for TPUEstimator."""
+
+        tf.logging.info("*** Features ***")
+        for name in sorted(features.keys()):
+            tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))
+
+        unique_ids = features["unique_ids"]
+        input_ids = features["input_ids"]
+        input_mask = features["input_mask"]
+        segment_ids = features["segment_ids"]
+
+        is_training = (mode == tf.estimator.ModeKeys.TRAIN)
+
+        (start_logits, end_logits) = create_model(
+            bert_config=bert_config,
+            is_training=is_training,
+            input_ids=input_ids,
+            input_mask=input_mask,
+            segment_ids=segment_ids,
+            use_one_hot_embeddings=use_one_hot_embeddings)
+
+        tvars = tf.trainable_variables()
+
+        initialized_variable_names = {}
+        scaffold_fn = None
+        if init_checkpoint:
+            (assignment_map,
+             initialized_variable_names) = modeling.get_assigment_map_from_checkpoint(
+                tvars, init_checkpoint)
+            if use_tpu:
+
+                def tpu_scaffold():
+                    tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
+                    return tf.train.Scaffold()
+
+                scaffold_fn = tpu_scaffold
+            else:
+                tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
+
+        tf.logging.info("**** Trainable Variables ****")
+        for var in tvars:
+            init_string = ""
+            if var.name in initialized_variable_names:
+                init_string = ", *INIT_FROM_CKPT*"
+            tf.logging.info("  name = %s, shape = %s%s", var.name, var.shape,
+                            init_string)
+
+        output_spec = None
+        if mode == tf.estimator.ModeKeys.TRAIN:
+            seq_length = modeling.get_shape_list(input_ids)[1]
+
+            def compute_loss(logits, positions):
+                one_hot_positions = tf.one_hot(
+                    positions, depth=seq_length, dtype=tf.float32)
+                log_probs = tf.nn.log_softmax(logits, axis=-1)
+                loss = -tf.reduce_mean(
+                    tf.reduce_sum(one_hot_positions * log_probs, axis=-1))
+                return loss
+
+            start_positions = features["start_positions"]
+            end_positions = features["end_positions"]
+
+            start_loss = compute_loss(start_logits, start_positions)
+            end_loss = compute_loss(end_logits, end_positions)
+
+            total_loss = (start_loss + end_loss) / 2.0
+
+            train_op = optimization.create_optimizer(
+                total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)
+
+            output_spec = tf.contrib.tpu.TPUEstimatorSpec(
+                mode=mode,
+                loss=total_loss,
+                train_op=train_op,
+                scaffold_fn=scaffold_fn)
+        elif mode == tf.estimator.ModeKeys.PREDICT:
+            predictions = {
+                "unique_ids": unique_ids,
+                "start_logits": start_logits,
+                "end_logits": end_logits,
+            }
+            output_spec = tf.contrib.tpu.TPUEstimatorSpec(
+                mode=mode, predictions=predictions, scaffold_fn=scaffold_fn)
+        else:
+            raise ValueError(
+                "Only TRAIN and PREDICT modes are supported: %s" % (mode))
+
+        return output_spec
+
+    return model_fn
+
+
+def input_fn_builder(features, seq_length, is_training, drop_remainder):
+    """Creates an `input_fn` closure to be passed to TPUEstimator."""
+
+    all_unique_ids = []
+    all_input_ids = []
+    all_input_mask = []
+    all_segment_ids = []
+    all_start_positions = []
+    all_end_positions = []
+
+    for feature in features:
+        all_unique_ids.append(feature.unique_id)
+        all_input_ids.append(feature.input_ids)
+        all_input_mask.append(feature.input_mask)
+        all_segment_ids.append(feature.segment_ids)
+        if is_training:
+            all_start_positions.append(feature.start_position)
+            all_end_positions.append(feature.end_position)
+
+    def input_fn(params):
+        """The actual input function."""
+        batch_size = params["batch_size"]
+
+        num_examples = len(features)
+
+        # This is for demo purposes and does NOT scale to large data sets. We do
+        # not use Dataset.from_generator() because that uses tf.py_func which is
+        # not TPU compatible. The right way to load data is with TFRecordReader.
+        feature_map = {
+            "unique_ids":
+                tf.constant(all_unique_ids, shape=[num_examples], dtype=tf.int32),
+            "input_ids":
+                tf.constant(
+                    all_input_ids, shape=[num_examples, seq_length],
+                    dtype=tf.int32),
+            "input_mask":
+                tf.constant(
+                    all_input_mask,
+                    shape=[num_examples, seq_length],
+                    dtype=tf.int32),
+            "segment_ids":
+                tf.constant(
+                    all_segment_ids,
+                    shape=[num_examples, seq_length],
+                    dtype=tf.int32),
+        }
+        if is_training:
+            feature_map["start_positions"] = tf.constant(
+                all_start_positions, shape=[num_examples], dtype=tf.int32)
+            feature_map["end_positions"] = tf.constant(
+                all_end_positions, shape=[num_examples], dtype=tf.int32)
+
+        d = tf.data.Dataset.from_tensor_slices(feature_map)
+
+        if is_training:
+            d = d.repeat()
+            d = d.shuffle(buffer_size=100)
+
+        d = d.batch(batch_size=batch_size, drop_remainder=drop_remainder)
+        return d
+
+    return input_fn
+
+
+RawResult = collections.namedtuple("RawResult",
+                                   ["unique_id", "start_logits", "end_logits"])
+
+
+def write_predictions(all_examples, all_features, all_results, n_best_size,
+                      max_answer_length, do_lower_case, output_prediction_file,
+                      output_nbest_file):
+    """Write final predictions to the json file."""
+    tf.logging.info("Writing predictions to: %s" % (output_prediction_file))
+    tf.logging.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index]))
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True)
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"])
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+
+            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+            orig_doc_start = feature.token_to_orig_map[pred.start_index]
+            orig_doc_end = feature.token_to_orig_map[pred.end_index]
+            orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+            tok_text = " ".join(tok_tokens)
+
+            # De-tokenize WordPieces that have been split off.
+            tok_text = tok_text.replace(" ##", "")
+            tok_text = tok_text.replace("##", "")
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            final_text = get_final_text(tok_text, orig_text, do_lower_case)
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        all_predictions[example.qas_id] = nbest_json[0]["text"]
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with tf.gfile.GFile(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with tf.gfile.GFile(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+
+def get_final_text(pred_text, orig_text, do_lower_case):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #   pred_text = steve smith
+    #   orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heruistic between
+    # `pred_text` and `orig_text` to get a character-to-charcter alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for (i, c) in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = tokenization.BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        if args.verbose_logging:
+            tf.logging.info(
+                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        if args.verbose_logging:
+            tf.logging.info("Length not equal after stripping spaces: '%s' vs '%s'",
+                            orig_ns_text, tok_ns_text)
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for (i, tok_index) in six.iteritems(tok_ns_to_s_map):
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        if args.verbose_logging:
+            tf.logging.info("Couldn't map start position")
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        if args.verbose_logging:
+            tf.logging.info("Couldn't map end position")
+        return orig_text
+
+    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs
+
+
+def main(_):
+    tf.logging.set_verbosity(tf.logging.INFO)
+
+    if not args.do_train and not args.do_predict:
+        raise ValueError("At least one of `do_train` or `do_predict` must be True.")
+
+    if args.do_train:
+        if not args.train_file:
+            raise ValueError(
+                "If `do_train` is True, then `train_file` must be specified.")
+    if args.do_predict:
+        if not args.predict_file:
+            raise ValueError(
+                "If `do_predict` is True, then `predict_file` must be specified.")
+
+    bert_config = modeling.BertConfig.from_json_file(args.bert_config_file)
+
+    if args.max_seq_length > bert_config.max_position_embeddings:
+        raise ValueError(
+            "Cannot use sequence length %d because the BERT model "
+            "was only trained up to sequence length %d" %
+            (args.max_seq_length, bert_config.max_position_embeddings))
+
+    tf.gfile.MakeDirs(args.output_dir)
+
+    tokenizer = tokenization.FullTokenizer(
+        vocab_file=args.vocab_file, do_lower_case=args.do_lower_case)
+
+    tpu_cluster_resolver = None
+    if args.use_tpu and args.tpu_name:
+        tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
+            args.tpu_name, zone=args.tpu_zone, project=args.gcp_project)
+
+    is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
+    run_config = tf.contrib.tpu.RunConfig(
+        cluster=tpu_cluster_resolver,
+        master=args.master,
+        model_dir=args.output_dir,
+        save_checkpoints_steps=args.save_checkpoints_steps,
+        tpu_config=tf.contrib.tpu.TPUConfig(
+            iterations_per_loop=args.iterations_per_loop,
+            num_shards=args.num_tpu_cores,
+            per_host_input_for_training=is_per_host))
+
+    train_examples = None
+    num_train_steps = None
+    num_warmup_steps = None
+    if args.do_train:
+        train_examples = read_squad_examples(
+            input_file=args.train_file, is_training=True)
+        num_train_steps = int(
+            len(train_examples) / args.train_batch_size * args.num_train_epochs)
+        num_warmup_steps = int(num_train_steps * args.warmup_proportion)
+
+    model_fn = model_fn_builder(
+        bert_config=bert_config,
+        init_checkpoint=args.init_checkpoint,
+        learning_rate=args.learning_rate,
+        num_train_steps=num_train_steps,
+        num_warmup_steps=num_warmup_steps,
+        use_tpu=args.use_tpu,
+        use_one_hot_embeddings=args.use_tpu)
+
+    # If TPU is not available, this will fall back to normal Estimator on CPU
+    # or GPU.
+    estimator = tf.contrib.tpu.TPUEstimator(
+        use_tpu=args.use_tpu,
+        model_fn=model_fn,
+        config=run_config,
+        train_batch_size=args.train_batch_size,
+        predict_batch_size=args.predict_batch_size)
+
+    if args.do_train:
+        train_features = convert_examples_to_features(
+            examples=train_examples,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=True)
+        tf.logging.info("***** Running training *****")
+        tf.logging.info("  Num orig examples = %d", len(train_examples))
+        tf.logging.info("  Num split examples = %d", len(train_features))
+        tf.logging.info("  Batch size = %d", args.train_batch_size)
+        tf.logging.info("  Num steps = %d", num_train_steps)
+        train_input_fn = input_fn_builder(
+            features=train_features,
+            seq_length=args.max_seq_length,
+            is_training=True,
+            drop_remainder=True)
+        estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
+
+    if args.do_predict:
+        eval_examples = read_squad_examples(
+            input_file=args.predict_file, is_training=False)
+        eval_features = convert_examples_to_features(
+            examples=eval_examples,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=False)
+
+        tf.logging.info("***** Running predictions *****")
+        tf.logging.info("  Num orig examples = %d", len(eval_examples))
+        tf.logging.info("  Num split examples = %d", len(eval_features))
+        tf.logging.info("  Batch size = %d", args.predict_batch_size)
+
+        all_results = []
+
+        predict_input_fn = input_fn_builder(
+            features=eval_features,
+            seq_length=args.max_seq_length,
+            is_training=False,
+            drop_remainder=False)
+
+        # If running eval on the TPU, you will need to specify the number of
+        # steps.
+        all_results = []
+        for result in estimator.predict(
+                predict_input_fn, yield_single_examples=True):
+            if len(all_results) % 1000 == 0:
+                tf.logging.info("Processing example: %d" % (len(all_results)))
+            unique_id = int(result["unique_ids"])
+            start_logits = [float(x) for x in result["start_logits"].flat]
+            end_logits = [float(x) for x in result["end_logits"].flat]
+            all_results.append(
+                RawResult(
+                    unique_id=unique_id,
+                    start_logits=start_logits,
+                    end_logits=end_logits))
+
+        output_prediction_file = os.path.join(args.output_dir, "predictions.json")
+        output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
+        write_predictions(eval_examples, eval_features, all_results,
+                          args.n_best_size, args.max_answer_length,
+                          args.do_lower_case, output_prediction_file,
+                          output_nbest_file)
+
+
+if __name__ == "__main__":
+    tf.app.run()