run_yesno.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 1.1 and SQuAD 2.0."""

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 numpy as np
import tensorflow as tf

flags = tf.flags

FLAGS = flags.FLAGS

## Required parameters
flags.DEFINE_string(
    "bert_config_file", None,
    "The config json file corresponding to the pre-trained BERT model. "
    "This specifies the model architecture.")

flags.DEFINE_string("vocab_file", None,
                    "The vocabulary file that the BERT model was trained on.")

flags.DEFINE_string(
    "output_dir", None,
    "The output directory where the model checkpoints will be written.")

## Other parameters
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(
    "init_checkpoint", None,
    "Initial checkpoint (usually from a pre-trained BERT model).")

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_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("do_train", False, "Whether to run training.")

flags.DEFINE_bool("do_predict", False, "Whether to run eval on the dev set.")

flags.DEFINE_integer("train_batch_size", 32, "Total batch size for training.")

flags.DEFINE_integer("predict_batch_size", 8,
                     "Total batch size for predictions.")

flags.DEFINE_float("learning_rate", 5e-5, "The initial learning rate for Adam.")

flags.DEFINE_float("num_train_epochs", 3.0,
                   "Total number of training epochs to perform.")

flags.DEFINE_float(
    "warmup_proportion", 0.1,
    "Proportion of training to perform linear learning rate warmup for. "
    "E.g., 0.1 = 10% of training.")

flags.DEFINE_integer("save_checkpoints_steps", 1000,
                     "How often to save the model checkpoint.")

flags.DEFINE_integer("iterations_per_loop", 1000,
                     "How many steps to make in each estimator call.")

flags.DEFINE_integer(
    "n_best_size", 20,
    "The total number of n-best predictions to generate in the "
    "nbest_predictions.json output file.")

flags.DEFINE_integer(
    "max_answer_length", 30,
    "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.")

flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.")

tf.flags.DEFINE_string(
    "tpu_name", None,
    "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.")

tf.flags.DEFINE_string(
    "tpu_zone", None,
    "[Optional] GCE zone where the Cloud TPU is located in. If not "
    "specified, we will attempt to automatically detect the GCE project from "
    "metadata.")

tf.flags.DEFINE_string(
    "gcp_project", None,
    "[Optional] Project name for the Cloud TPU-enabled project. If not "
    "specified, we will attempt to automatically detect the GCE project from "
    "metadata.")

tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.")

flags.DEFINE_integer(
    "num_tpu_cores", 8,
    "Only used if `use_tpu` is True. Total number of TPU cores to use.")

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.DEFINE_bool(
    "version_2_with_negative", False,
    "If true, the SQuAD examples contain some that do not have an answer.")

flags.DEFINE_float(
    "null_score_diff_threshold", 0.0,
    "If null_score - best_non_null is greater than the threshold predict null.")

tf.flags.DEFINE_string("bioasq_version", None, "[Optional] TensorFlow master URL.")
tf.flags.DEFINE_string("bioasq_batch", None, "[Optional] TensorFlow master URL.")
tf.flags.DEFINE_string("bioasq_task", None, "[Optional] TensorFlow master URL.")
tf.flags.DEFINE_string("bioasq_snippet", None, "[Optional] TensorFlow master URL.")


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,
               answer):
    self.qas_id = qas_id
    self.question_text = question_text
    self.doc_tokens = doc_tokens
    self.answer = answer

  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.answer:
      s += ", answer: %r" % (self.answer)
    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,
               target):
    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.target = target


def read_squad_examples(input_file, is_training):
  """Read a SQuAD json file into a list of SquadExample."""
  is_bioasq=True # for BioASQ

  with tf.gfile.Open(input_file, "r") as reader:
    #if is_bioasq:
      #input_data = [{u'paragraphs':json.load(reader)["questions"], u'title':'bioASQ'}] # to fit the shape of squad code
    #else:
    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
      if is_bioasq:
        paragraph_text.replace('/',' ')  # need review
      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"]
        answer = None
        if is_training:
          assert (qa["is_impossible"] == True) != (qa["answers"] == "yes")
          assert qa["answers"] in ["yes", "no"]
          answer = 1 if qa["answers"] == 'yes' else 0

        example = SquadExample(
            qas_id=qas_id,
            question_text=question_text,
            doc_tokens=doc_tokens,
            answer=answer
        )
        examples.append(example)

  target_cnt = 500
  if is_training:
    pos_cnt = sum([1 for example in examples if example.answer == 1])
    neg_cnt = sum([1 for example in examples if example.answer == 0])
    print()
    print('Imbalance btw {} vs {}'.format(pos_cnt, neg_cnt))
    # rng = random.Random(12345)
    random.shuffle(examples)

    new_examples = []
    new_pos_cnt = 0
    new_neg_cnt = 0
    for example in examples:
      if example.answer == 1 and new_pos_cnt >= target_cnt:
        continue
      if example.answer == 0 and new_neg_cnt >= target_cnt:
        continue
      else:
        new_examples.append(example)
        new_pos_cnt += (1 if example.answer == 1 else 0)
        new_neg_cnt += (1 if example.answer == 0 else 0)

    pos_cnt = sum([1 for example in new_examples if example.answer == 1])
    neg_cnt = sum([1 for example in new_examples if example.answer == 0])
    random.shuffle(new_examples)
    print('Balanced as {} vs {}'.format(pos_cnt, neg_cnt))
    print('Sample: {}'.format(new_examples[0]))
    return new_examples
  else:
    return examples


def convert_examples_to_features(examples, tokenizer, max_seq_length,
                                 doc_stride, max_query_length, is_training,
                                 output_fn):
  """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)

    target = example.answer

    # 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

      if example_index < 1:
        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:
          tf.logging.info(
              "target: %d" % (target))

      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,
          target=target)

      # Run callback
      output_fn(feature)
      unique_id += 1


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]

  # Standard weight (for start/end)
  '''
  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())

  '''
  # Sigmoid weight
  output_weights_sgm = tf.get_variable(
      "cls/squad/output_weights_sgm", [1, hidden_size],
      initializer=tf.truncated_normal_initializer(stddev=0.02))
  output_bias_sgm = tf.get_variable(
      "cls/squad/output_bias_sgm", [1], initializer=tf.zeros_initializer())

  # Get first CLS token
  cls_hidden = final_hidden[:,0,:]
  # cls_hidden = tf.layers.dropout(cls_hidden, rate=0.5, training=is_training)

  # Standard logit
  '''
  logits = tf.matmul(cls_hidden, output_weights, transpose_b=True)
  logits = tf.nn.bias_add(logits, output_bias)
  '''

  # Sigmoid logit
  logits = tf.matmul(cls_hidden, output_weights_sgm, transpose_b=True)
  logits = tf.nn.bias_add(logits, output_bias_sgm)

  return 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)

    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_assignment_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)
      break

    output_spec = None
    if mode == tf.estimator.ModeKeys.TRAIN:
      seq_length = modeling.get_shape_list(input_ids)[1]

      def compute_loss(logits, targets):
        '''
        one_hot_positions = tf.one_hot(
            targets, depth=2, 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))
        '''
        logits = tf.squeeze(logits, axis=1)
        loss = tf.nn.sigmoid_cross_entropy_with_logits(
            labels=tf.cast(targets, tf.float32),
            logits=logits
        )
        loss = tf.reduce_mean(loss)
        return loss

      targets = features["target"]
      total_loss = compute_loss(logits, targets)

      train_op = optimization.create_optimizer(
          total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)
      logging_hook = tf.train.LoggingTensorHook({"loss" : total_loss}, every_n_iter=50)

      output_spec = tf.contrib.tpu.TPUEstimatorSpec(
          mode=mode,
          loss=total_loss,
          train_op=train_op,
          scaffold_fn=scaffold_fn,
          training_hooks=[logging_hook])
      print("new model!!")
    elif mode == tf.estimator.ModeKeys.PREDICT:
      predictions = {
          "unique_ids": unique_ids,
          "logits": 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(input_file, seq_length, is_training, drop_remainder):
  """Creates an `input_fn` closure to be passed to TPUEstimator."""

  name_to_features = {
      "unique_ids": tf.FixedLenFeature([], tf.int64),
      "input_ids": tf.FixedLenFeature([seq_length], tf.int64),
      "input_mask": tf.FixedLenFeature([seq_length], tf.int64),
      "segment_ids": tf.FixedLenFeature([seq_length], tf.int64),
  }

  if is_training:
    name_to_features["target"] = tf.FixedLenFeature([], tf.int64)

  def _decode_record(record, name_to_features):
    """Decodes a record to a TensorFlow example."""
    example = tf.parse_single_example(record, name_to_features)

    # tf.Example only supports tf.int64, but the TPU only supports tf.int32.
    # So cast all int64 to int32.
    for name in list(example.keys()):
      t = example[name]
      if t.dtype == tf.int64:
        t = tf.to_int32(t)
      example[name] = t

    return example

  def input_fn(params):
    """The actual input function."""
    batch_size = params["batch_size"]

    # For training, we want a lot of parallel reading and shuffling.
    # For eval, we want no shuffling and parallel reading doesn't matter.
    d = tf.data.TFRecordDataset(input_file)
    if is_training:
      d = d.repeat(None)
      # d = d.shuffle(buffer_size=100)

    d = d.apply(
        tf.contrib.data.map_and_batch(
            lambda record: _decode_record(record, name_to_features),
            batch_size=batch_size,
            drop_remainder=drop_remainder))

    return d

  return input_fn


RawResult = collections.namedtuple("RawResult",
                                   ["unique_id", "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, output_null_log_odds_file):
  """Write final predictions to the json file and log-odds of null if needed."""
  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", "answer", "logit"])

  all_predictions = collections.OrderedDict()
  all_nbest_json = collections.OrderedDict()
  scores_diff_json = collections.OrderedDict()

  for (example_index, example) in enumerate(all_examples):
    features = example_index_to_features[example_index]

    prelim_predictions = []
    # keep track of the minimum score of null start+end of position 0
    score_null = 1000000  # large and positive
    for (feature_index, feature) in enumerate(features):
      result = unique_id_to_result[feature.unique_id]
      # logits = result.logits[0]
      logits = result.logits
      answer = 'yes' if logits[0] > 0.5 else 'no'
      # answer = 'yes' if np.argmax(logits) == 1 else 'no'

      prelim_predictions.append(
          _PrelimPrediction(
              feature_index=feature_index,
              answer=answer,
              # logit=max(logits)))
              logit=logits))

      break

    assert len(prelim_predictions) == 1
    # Not used..
    # probs = _compute_softmax(logits)
    probs = logits
    for (i, entry) in enumerate(prelim_predictions):
      output = collections.OrderedDict()
      output["text"] = entry.answer
      output["probability"] = probs
      output["logit"] = entry.logit

    all_predictions[example.qas_id] = [prelim_predictions[0].answer, probs]

  with tf.gfile.GFile(output_prediction_file, "w") as writer:
    writer.write(json.dumps(all_predictions, 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 FLAGS.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 FLAGS.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 FLAGS.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 FLAGS.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 _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


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.num_pos_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
    self.num_pos_features += 1 if feature.target == 1 else 0

    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["target"] = create_int_feature([feature.target])

    tf_example = tf.train.Example(features=tf.train.Features(feature=features))
    self._writer.write(tf_example.SerializeToString())

  def close(self):
    self._writer.close()


def validate_flags_or_throw(bert_config):
  """Validate the input FLAGS or throw an exception."""
  tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case,
                                                FLAGS.init_checkpoint)

  if not FLAGS.do_train and not FLAGS.do_predict:
    raise ValueError("At least one of `do_train` or `do_predict` must be True.")

  if FLAGS.do_train:
    if not FLAGS.train_file:
      raise ValueError(
          "If `do_train` is True, then `train_file` must be specified.")
  if FLAGS.do_predict:
    if not FLAGS.predict_file:
      raise ValueError(
          "If `do_predict` is True, then `predict_file` must be specified.")

  if FLAGS.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" %
        (FLAGS.max_seq_length, bert_config.max_position_embeddings))

  if FLAGS.max_seq_length <= FLAGS.max_query_length + 3:
    raise ValueError(
        "The max_seq_length (%d) must be greater than max_query_length "
        "(%d) + 3" % (FLAGS.max_seq_length, FLAGS.max_query_length))


def main(_):
  #seed = 1111
  #random.seed(seed)
  #np.random.seed(seed)
  #tf.set_random_seed(seed)
  #tf.logging.set_verbosity(tf.logging.INFO)

  bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file)

  validate_flags_or_throw(bert_config)

  tf.gfile.MakeDirs(FLAGS.output_dir)

  tokenizer = tokenization.FullTokenizer(
      vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)

  tpu_cluster_resolver = None
  if FLAGS.use_tpu and FLAGS.tpu_name:
    tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
        FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)

  is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
  run_config = tf.contrib.tpu.RunConfig(
      cluster=tpu_cluster_resolver,
      master=FLAGS.master,
      model_dir=FLAGS.output_dir,
      save_checkpoints_steps=FLAGS.save_checkpoints_steps,
      tpu_config=tf.contrib.tpu.TPUConfig(
          iterations_per_loop=FLAGS.iterations_per_loop,
          num_shards=FLAGS.num_tpu_cores,
          per_host_input_for_training=is_per_host))

  num_train_steps = None
  num_warmup_steps = None
  if FLAGS.do_train:
    train_examples = read_squad_examples(
        input_file=FLAGS.train_file, is_training=True)
    num_train_steps = int(math.ceil(len(train_examples) / FLAGS.train_batch_size) * FLAGS.num_train_epochs)
    # num_train_steps = FLAGS.save_checkpoints_steps
    num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)

    # Re-define run_config for training
    run_config = tf.contrib.tpu.RunConfig(
        cluster=tpu_cluster_resolver,
        master=FLAGS.master,
        model_dir=FLAGS.output_dir,
        save_checkpoints_steps=int(num_train_steps/FLAGS.num_train_epochs),
        tpu_config=tf.contrib.tpu.TPUConfig(
            iterations_per_loop=int(num_train_steps/FLAGS.num_train_epochs),
            num_shards=FLAGS.num_tpu_cores,
            per_host_input_for_training=is_per_host))

    rng = random.Random(12345)
    rng.shuffle(train_examples)

  model_fn = model_fn_builder(
      bert_config=bert_config,
      init_checkpoint=FLAGS.init_checkpoint,
      learning_rate=FLAGS.learning_rate,
      num_train_steps=num_train_steps,
      num_warmup_steps=num_warmup_steps,
      use_tpu=FLAGS.use_tpu,
      use_one_hot_embeddings=FLAGS.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=FLAGS.use_tpu,
      model_fn=model_fn,
      config=run_config,
      train_batch_size=FLAGS.train_batch_size,
      predict_batch_size=FLAGS.predict_batch_size)

  if FLAGS.do_train:
    # We write to a temporary file to avoid storing very large constant tensors
    # in memory.
    train_writer = FeatureWriter(
        filename=os.path.join(FLAGS.output_dir, "train.tf_record"),
        is_training=True)

    # Bootstrap
    for epoch_idx in range(int(FLAGS.num_train_epochs)):
        train_examples = read_squad_examples(
            input_file=FLAGS.train_file, 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)

    train_writer.close()
    tf.logging.info("***** Running training *****")
    tf.logging.info("  Num orig examples = %d", len(train_examples) * int(FLAGS.num_train_epochs))
    tf.logging.info("  Num split examples = %d", train_writer.num_features)
    tf.logging.info("  Num pos examples = %d", train_writer.num_pos_features)
    tf.logging.info("  Batch size = %d", FLAGS.train_batch_size)
    tf.logging.info("  Num steps = %d", num_train_steps)
    tf.logging.info("  Num warmup steps = %d", num_warmup_steps)

    train_input_fn = input_fn_builder(
        input_file=train_writer.filename,
        seq_length=FLAGS.max_seq_length,
        is_training=True,
        drop_remainder=True)

    # estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)
    estimator.train(input_fn=train_input_fn, steps=num_train_steps)

  if FLAGS.do_predict:
    eval_examples = read_squad_examples(
        input_file=FLAGS.predict_file, is_training=False)

    eval_writer = FeatureWriter(
        filename=os.path.join(FLAGS.output_dir, "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)
    eval_writer.close()

    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", FLAGS.predict_batch_size)

    all_results = []

    predict_input_fn = input_fn_builder(
        input_file=eval_writer.filename,
        seq_length=FLAGS.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) % 100 == 0:
        tf.logging.info("Processing example: %d" % (len(all_results)))
      unique_id = int(result["unique_ids"])
      logits = [float(x) for x in result["logits"].flat]
      all_results.append(
          RawResult(
              unique_id=unique_id,
              logits=logits))

    output_prediction_file = os.path.join(FLAGS.output_dir, "predictions.json")
    #output_prediction_file = os.path.join(FLAGS.output_dir, "predictions_{}b{}_{}{}.json".format(
    #    FLAGS.bioasq_version, FLAGS.bioasq_batch, FLAGS.bioasq_task, FLAGS.bioasq_snippet))
    output_nbest_file = os.path.join(FLAGS.output_dir, "nbest_predictions.json")
    output_null_log_odds_file = os.path.join(FLAGS.output_dir, "null_odds.json")

    write_predictions(eval_examples, eval_features, all_results,
                      FLAGS.n_best_size, FLAGS.max_answer_length,
                      FLAGS.do_lower_case, output_prediction_file,
                      output_nbest_file, output_null_log_odds_file)


if __name__ == "__main__":
  flags.mark_flag_as_required("vocab_file")
  flags.mark_flag_as_required("bert_config_file")
  flags.mark_flag_as_required("output_dir")
  tf.app.run()