word_edit_model.py

# code adapted from https://github.com/google-research/bert

# 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.
"""BERT finetuning runner."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import csv
import os
import modeling
import modified_modeling #obtains contextual embeddings 
#                         for appends and replacements
#                         for edit factorized architecture
#                         figure 2 in the paper
import optimization
import tokenization
import tensorflow as tf
import numpy as np
import pickle

from itertools import chain
from tensorflow.python.lib.io.file_io import get_matching_files
from modeling import get_shape_list

import wem_utils

flags = tf.flags

FLAGS = flags.FLAGS

## Required parameters
flags.DEFINE_string(
    "data_dir", None,
    "The input data dir. Should contain the .txt files (or other data files) "
    "for the task.")

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(
    "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", 128,
    "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_bool("do_train", False, "Whether to run training.")

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

flags.DEFINE_bool(
    "do_predict", False,
    "Whether to run the model in inference mode on the test set.")

flags.DEFINE_integer("train_batch_size", 64, "Total batch size for training.")

flags.DEFINE_integer("eval_batch_size", 512, "Total batch size for eval.")

flags.DEFINE_integer("predict_batch_size", 512, "Total batch size for predict.")

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_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_float("copy_weight", 1, "weight to copy")

flags.DEFINE_bool("use_bert_more", True, "use bert more exhaustively for logit computation")

flags.DEFINE_string("path_inserts", None, "path to insert pickle")

flags.DEFINE_string("path_multitoken_inserts", None, "path to multitoken_inserts")

flags.DEFINE_bool("subtract_replaced_from_replacement", True, "subtract_replaced_from_replacement")

flags.DEFINE_string("eval_checkpoint", None, "checkpoint to evaluate gec model")

flags.DEFINE_string("predict_checkpoint", None, "checkpoint to use for predictions")

flags.DEFINE_integer("random_seed",0,"random seed for creating random initializations")

flags.DEFINE_bool("create_train_tf_records", True, "whether to create train tf records")

flags.DEFINE_bool("create_predict_tf_records", True, "whether to create predict tf records")

#flags.DEFINE_bool("dump_probs", False, "dump edit probs to numpy file while decoding")

class PaddingInputExample(object):
  """Fake example so the num input examples is a multiple of the batch size.

  When running eval/predict on the TPU, we need to pad the number of examples
  to be a multiple of the batch size, because the TPU requires a fixed batch
  size. The alternative is to drop the last batch, which is bad because it means
  the entire output data won't be generated.

  We use this class instead of `None` because treating `None` as padding
  battches could cause silent errors.
  """

class GECInputExample(object):
  def __init__(self, guid, input_sequence, edit_sequence=None):
    """Constructs a InputExample."""
    self.guid = guid
    self.input_sequence = input_sequence
    self.edit_sequence = edit_sequence


class GECInputFeatures(object):
  def __init__(self, input_sequence, input_mask, segment_ids, edit_sequence):
    self.input_sequence = input_sequence
    self.input_mask = input_mask
    self.segment_ids = segment_ids
    #self.label_id = label_id
    self.edit_sequence = edit_sequence


class DataProcessor(object):
  """Base class for data converters for sequence classification data sets."""

  def get_train_examples(self, data_dir):
    """Gets a collection of `InputExample`s for the train set."""
    raise NotImplementedError()

  def get_dev_examples(self, data_dir):
    """Gets a collection of `InputExample`s for the dev set."""
    raise NotImplementedError()

  def get_test_examples(self, data_dir):
    """Gets a collection of `InputExample`s for prediction."""
    raise NotImplementedError()

  @classmethod
  def _read_file(cls, input_file):
    """Reads a tab separated value file."""
    with tf.gfile.Open(input_file, "r") as f:
      return (line for line in f)


class GECProcessor(DataProcessor):
  def get_train_examples(self, data_dir):
    """See base class."""
    train_incorr = self._read_file(os.path.join(data_dir, "train_incorr.txt"))
    train_labels = self._read_file(os.path.join(data_dir, "train_labels.txt"))
    return self._create_examples(train_incorr, train_labels, "train")

  def get_dev_examples(self, data_dir):
    """See base class."""
    dev_incorr = self._read_file(os.path.join(data_dir, "dev_incorr.txt"))
    dev_labels = self._read_file(os.path.join(data_dir, "dev_labels.txt"))
    return self._create_examples(dev_incorr, dev_labels, "dev")

  def get_test_examples(self, data_dir):
    """See base class."""
    test_incorr = self._read_file(os.path.join(data_dir, "test_incorr.txt"))
    #test_labels = self._read_file(os.path.join(data_dir, "test_labels.txt"))
    test_labels = None
    return self._create_examples(test_incorr, test_labels, "test")

  def _create_examples(self, incorr_lines, labels_lines, set_type):
    """Creates examples for the training and dev sets."""
    if set_type != "test":
      for (i, (incorr_line, labels_line)) in enumerate(zip(incorr_lines, labels_lines)):
        guid = "%s-%s" % (set_type, i)
        input_sequence = incorr_line
        edit_sequence = labels_line
        yield GECInputExample(guid, input_sequence, edit_sequence)
    else:
      for (i, incorr_line) in enumerate(incorr_lines):
        guid = "%s-%s" % (set_type, i)
        input_sequence = incorr_line
        edit_sequence = None
        yield GECInputExample(guid, input_sequence, edit_sequence)

def gec_convert_single_example(ex_index, example, max_seq_length):
  """Converts a single `InputExample` into a single `InputFeatures`."""
  if isinstance(example, PaddingInputExample):
    return GECInputFeatures(
        input_sequence=[0] * max_seq_length,
        input_mask=[0] * max_seq_length,
        segment_ids=[0] * max_seq_length,
        edit_sequence=[0] * max_seq_length)

  input_sequence = list(map(int, example.input_sequence.strip().split()))
  if len(input_sequence) > max_seq_length:
    input_sequence = input_sequence[0:(max_seq_length)]
  
  if example.edit_sequence:
    edit_sequence = list(map(int, example.edit_sequence.strip().split()))
    if len(edit_sequence) > max_seq_length:
      edit_sequence = edit_sequence[0:(max_seq_length)]
    
    if len(input_sequence) != len(edit_sequence):
      print("This should ideally not happen")
      exit(1)
  else:
    edit_sequence = None

  input_mask = [1] * len(input_sequence)
  segment_ids = [0] * len(input_sequence)

  # Zero-pad up to the sequence length.
  while len(input_sequence) < max_seq_length:
    input_sequence.append(0)
    if edit_sequence:
      edit_sequence.append(0)
    input_mask.append(0)
    segment_ids.append(0)

  if not edit_sequence:
    edit_sequence = [-1] * max_seq_length

  assert len(input_sequence) == max_seq_length
  assert len(input_mask) == max_seq_length
  assert len(segment_ids) == max_seq_length
  assert len(edit_sequence) == max_seq_length

  if ex_index < 5:
    tf.logging.info("*** Example ***")
    tf.logging.info("guid: %s" % (example.guid))
    tf.logging.info("input_sequence: %s" % " ".join([str(x) for x in input_sequence]))
    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]))
    tf.logging.info("edit_sequence: %s" % " ".join([str(x) for x in edit_sequence]))

  feature = GECInputFeatures(
      input_sequence=input_sequence,
      input_mask=input_mask,
      segment_ids=segment_ids,
      edit_sequence=edit_sequence)
  return feature

def gec_file_based_convert_examples_to_features(
    examples, max_seq_length, output_dir, mode, num_examples):
  """Convert a set of `InputExample`s to a TFRecord file."""
  num_writers = 0
  writer = None 
  for (ex_index, example) in enumerate(examples):
    if ex_index%10000==0:
      tf.logging.info("Writing example %d of %d" % (ex_index, num_examples))
    if ex_index % 500000000000 == 0:
      if writer:
      	writer.close()
      	del writer
      output_file = os.path.join(output_dir, "{}_{}.tf_record".format(mode,num_writers))
      writer = tf.python_io.TFRecordWriter(output_file)
      num_writers += 1

    feature = gec_convert_single_example(ex_index, example, max_seq_length)

    def create_int_feature(values):
      f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))
      return f

    features = collections.OrderedDict()
    features["input_sequence"] = create_int_feature(feature.input_sequence)
    features["input_mask"] = create_int_feature(feature.input_mask)
    features["segment_ids"] = create_int_feature(feature.segment_ids)
    features["edit_sequence"] = create_int_feature(feature.edit_sequence)

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

def gec_file_based_input_fn_builder(output_dir, mode, seq_length,
                                is_training, drop_remainder):
  """Creates an `input_fn` closure to be passed to TPUEstimator."""
  #output_dir_parts = PurePath(output_dir).parts
  #output_dir_path = "/home/awasthiabhijeet05/mnt_bucket/" + "/".join(output_dir_parts[2:])
  #print(output_dir_path+"/"+"{}_*.tf_record".format(mode))
  input_files = get_matching_files(output_dir+"/"+"{}_*.tf_record".format(mode))
  print("INPUT_FILES: " + " AND ".join(input_files))

  name_to_features = {
      "input_sequence": tf.FixedLenFeature([seq_length], tf.int64),
      "input_mask": tf.FixedLenFeature([seq_length], tf.int64),
      "segment_ids": tf.FixedLenFeature([seq_length], tf.int64),
      "edit_sequence": tf.FixedLenFeature([seq_length], 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_files)
    if is_training:
      d = d.repeat()
      d = d.shuffle(buffer_size=5000)

    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

def edit_word_embedding_lookup(embedding_table, input_ids,
 use_one_hot_embeddings, vocab_size, embedding_size):
  if use_one_hot_embeddings:
    one_hot_input_ids = tf.one_hot(input_ids, depth=vocab_size)
    output = tf.matmul(one_hot_input_ids, embedding_table)
  else:
    output = tf.nn.embedding_lookup(embedding_table, input_ids)
  return output


def gec_create_model(bert_config, is_training, input_sequence, 
  input_mask, segment_ids, edit_sequence, 
  use_one_hot_embeddings, mode, 
  copy_weight, 
  use_bert_more, 
  insert_ids,
  multitoken_insert_ids,
  subtract_replaced_from_replacement):
  """Creates a classification model."""
  # insert_ids: word ids of unigram inserts (list)
  # multitoken_insert_ids: word_ids of bigram inserts (list of tuples of length 2)
  # Defining the space of all possible edits: 
  # unk, sos and eos are dummy edits mapped to 0, 1 and 2 respectively
  # copy is mapped to 3
  # del is mapped to 4
  num_appends = len(insert_ids) + len(multitoken_insert_ids)
  num_replaces = num_appends # appends and replacements come from the same set (inserts and multitoken_inserts)
  append_begin = 5 # First append edit (mapped to 5)
  append_end = append_begin + num_appends - 1 #Last append edit
  rep_begin = append_end + 1 # First replace edit
  rep_end = rep_begin + num_replaces - 1 #Last replace edit  
  num_suffix_transforms = 58 #num of transformation edits
  num_labels = 5 + num_appends + num_replaces + num_suffix_transforms # total number of edits
  print("************ num of labels : {} ***************".format(num_labels))

  config = bert_config
  input_sequence_shape = modeling.get_shape_list(input_sequence,2)
  batch_size = input_sequence_shape[0]
  seq_len = input_sequence_shape[1]

  if not use_bert_more:  #default use of bert (without logit factorisation)
    model = modeling.BertModel(
        config=bert_config,
        is_training=is_training,
        input_ids=input_sequence,
        input_mask=input_mask,
        token_type_ids=segment_ids,
        use_one_hot_embeddings=use_one_hot_embeddings)

    output_layer = model.get_sequence_output()
  else:                 # LOGIT FACTORISATION is On!
    model = modified_modeling.BertModel(
        config=bert_config,
        is_training=is_training,
        input_ids=input_sequence,
        input_mask=input_mask,
        token_type_ids=segment_ids,
        use_one_hot_embeddings=use_one_hot_embeddings)

    output_layer = model.get_sequence_output()
    replace_layer = output_layer[:,seq_len:2*seq_len,:]  #representation of replacement slots as described in paper
    append_layer = output_layer[:,2*seq_len:3*seq_len,:] #representation of append slots as described in paper
    output_layer = output_layer[:,0:seq_len,:]

  output_layer_shape = modeling.get_shape_list(output_layer,3)
  hidden_size = output_layer_shape[-1]

  flattened_output_layer = tf.reshape(output_layer,[-1, hidden_size])

  h_edit = flattened_output_layer

  if use_bert_more:
    h_word = flattened_output_layer
    flattened_replace_layer = tf.reshape(replace_layer,[-1, hidden_size])
    flattened_append_layer = tf.reshape(append_layer, [-1, hidden_size])

    m_replace = flattened_replace_layer    
    m_append = flattened_append_layer

    
    with tf.variable_scope("cls/predictions"):
      with tf.variable_scope("transform"):
        h_word = tf.layers.dense(
            h_word,
            units=bert_config.hidden_size,
            activation=modeling.get_activation(bert_config.hidden_act),
            kernel_initializer=modeling.create_initializer(
                bert_config.initializer_range))
        h_word = modeling.layer_norm(h_word)

    with tf.variable_scope("cls/predictions",reuse=True):
      with tf.variable_scope("transform",reuse=True):
        m_replace = tf.layers.dense(
            m_replace,
            units=bert_config.hidden_size,
            activation=modeling.get_activation(bert_config.hidden_act),
            kernel_initializer=modeling.create_initializer(
                bert_config.initializer_range))
        m_replace = modeling.layer_norm(m_replace)

    with tf.variable_scope("cls/predictions",reuse=True):
      with tf.variable_scope("transform",reuse=True):
        m_append = tf.layers.dense(
            m_append,
            units=bert_config.hidden_size,
            activation=modeling.get_activation(bert_config.hidden_act),
            kernel_initializer=modeling.create_initializer(
                bert_config.initializer_range))
        m_append = modeling.layer_norm(m_append)
    
    word_embedded_input = model.word_embedded_input
    flattened_word_embedded_input = tf.reshape(word_embedded_input, [-1, hidden_size])    

  labels = edit_sequence
  
  edit_weights = tf.get_variable(
      "edit_weights", [num_labels, hidden_size],
      initializer=tf.truncated_normal_initializer(stddev=0.02))

  if is_training:
    h_edit = tf.nn.dropout(h_edit, keep_prob=0.9) 

  if use_bert_more:
      # append/replace weight vector for a given append or replace operation
      # correspond to word embedding for its token argument
      # for multitoken append/replace (e.g. has been)
      # weight vector is sum of word embeddings of token arguments

      append_weights = edit_word_embedding_lookup(model.embedding_table, insert_ids,
       use_one_hot_embeddings, config.vocab_size, config.hidden_size)
      replace_weights = append_weights #tokens in replace and append vocab are same 
                                       #(i.e. inserts and multitoken_inserts)

      multitoken_append_weights = wem_utils.edit_embedding_loopkup(model.embedding_table, multitoken_insert_ids,
                        use_one_hot_embeddings, config.vocab_size, config.hidden_size)
      multitoken_replace_weights = multitoken_append_weights #tokens in replace and append vocab are same 
                                                             #(i.e. inserts and multitoken_inserts)

      append_weights = tf.concat([append_weights, multitoken_append_weights],0)
      replace_weights = tf.concat([replace_weights, multitoken_replace_weights],0)

  with tf.variable_scope("loss"):
    edit_logits = tf.matmul(h_edit, edit_weights, transpose_b=True) #first term in eq3 in paper
    logits = edit_logits
    if use_bert_more:

      #=============== inplace_word_logits==============# #2nd term in eq3 in paper
      inplace_logit = tf.reduce_sum(h_word * flattened_word_embedded_input, axis=1, keepdims=True) #copy
      #inplace_logit = tf.reduce_sum(m_replace * flattened_word_embedded_input, axis=1, keepdims=True) #copy
      inplace_logit_appends = tf.tile(inplace_logit,[1,num_appends])
      inplace_logit_transforms = tf.tile(inplace_logit,[1,num_suffix_transforms])
      zero_3_logits = tf.zeros([batch_size*seq_len,3]) #unk sos eos 
      zero_1_logits = tf.zeros([batch_size*seq_len,1]) # del
      zero_replace_logits = tf.zeros([batch_size*seq_len,num_replaces])

      concat_list = [zero_3_logits, inplace_logit, zero_1_logits]\
                  + [inplace_logit_appends]\
                  + [zero_replace_logits]\
                  + [inplace_logit_transforms]

      inplace_word_logits = tf.concat(concat_list,1)

      #======additional (insert,replace) logits ====# #3rd term in eqn3 in paper
      zero_5_logits = tf.zeros([batch_size*seq_len,5])      
      append_logits = tf.matmul(m_append, append_weights, transpose_b=True)

      if subtract_replaced_from_replacement:
        replace_logits = replacement_minus_replaced_logits(m_replace, 
          flattened_word_embedded_input, replace_weights)
      else:
        replace_logits = tf.matmul(m_replace, replace_weights, transpose_b=True)
      
      suffix_logits  = tf.zeros([batch_size*seq_len,num_suffix_transforms])
      
      concat_list = [zero_5_logits, append_logits, replace_logits, suffix_logits]
      additional_logits = tf.concat(concat_list,1)
      #====================================================#

      logits = edit_logits + inplace_word_logits + additional_logits
      logits_bias = tf.get_variable("output_bias", shape=[num_labels], initializer=tf.zeros_initializer())
      logits += logits_bias
    
    logits = tf.reshape(logits, [output_layer_shape[0], output_layer_shape[1], num_labels])
    log_probs = tf.nn.log_softmax(logits, axis=-1)
    probs = tf.nn.softmax(logits,axis=-1)
    one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
    per_token_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
    per_token_loss = per_token_loss * tf.to_float(input_mask)
    mask = copy_weight*tf.to_float(tf.equal(labels,3)) +  tf.to_float(tf.not_equal(labels,3))
    masked_per_token_loss = per_token_loss * mask
    per_example_loss = tf.reduce_sum(masked_per_token_loss, axis=-1)
    loss = tf.reduce_mean(per_example_loss)            

    return (loss, per_example_loss, logits, probs)


def replacement_minus_replaced_logits(replace_layer, word_embedded_input, weights):
  result_1 = tf.matmul(replace_layer, weights, transpose_b=True)
  result_2 = replace_layer * word_embedded_input
  result_2 = tf.reduce_sum(result_2,1)
  result_2 = tf.expand_dims(result_2,-1)
  return result_1 - result_2

def gec_model_fn_builder(bert_config, init_checkpoint, learning_rate,
                     num_train_steps, num_warmup_steps, use_tpu,
                     use_one_hot_embeddings, copy_weight, 
                     use_bert_more, 
                     inserts, insert_ids,
                     multitoken_inserts, multitoken_insert_ids,
                     subtract_replaced_from_replacement):
  """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))

    input_sequence = features["input_sequence"]
    input_mask = features["input_mask"]
    segment_ids = features["segment_ids"]
    edit_sequence = features["edit_sequence"]

    is_training = (mode == tf.estimator.ModeKeys.TRAIN)

    (total_loss, per_example_loss, logits, probabilities) = gec_create_model(
        bert_config, is_training, input_sequence,
         input_mask, segment_ids, edit_sequence,
        use_one_hot_embeddings, mode, 
        copy_weight, 
        use_bert_more, 
        insert_ids,
        multitoken_insert_ids,
        subtract_replaced_from_replacement)

    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)

    output_spec = None
    if mode == tf.estimator.ModeKeys.TRAIN:
      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.EVAL:
      def metric_fn(per_example_loss, edit_sequence, logits):
        predictions = tf.argmax(logits[:,:,3:], axis=-1, output_type=tf.int32) + 3 
        mask = tf.equal(edit_sequence,0)
        mask = tf.logical_or(mask, tf.equal(edit_sequence,1))
        mask = tf.logical_or(mask, tf.equal(edit_sequence,2))
        mask = tf.logical_or(mask, tf.equal(edit_sequence,3))
        mask = tf.to_float(tf.logical_not(mask))
        accuracy = tf.metrics.accuracy(edit_sequence, predictions, mask)
        loss = tf.metrics.mean(per_example_loss)
        result_dict = {}
        result_dict["eval_accuracy"] = accuracy
        result_dict["eval_loss"] = loss      
        return {
            "eval_accuracy": accuracy,
            "eval_loss": loss,
        }

      eval_metrics = (metric_fn, [per_example_loss, edit_sequence, logits])
      output_spec = tf.contrib.tpu.TPUEstimatorSpec(
          mode=mode,
          loss=total_loss,
          eval_metrics=eval_metrics,
          scaffold_fn=scaffold_fn)
    else:
      #first three edit ids unk, sos, eos are dummy. We do not consider them in predictions
      predictions = tf.argmax(logits[:,:,3:], axis=-1, output_type=tf.int32) + 3
      output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                mode=mode,
                predictions={"predictions": predictions, "logits":logits},
                scaffold_fn=scaffold_fn)
    return output_spec

  return model_fn

def get_file_length(file_address):
  num_lines = sum(1 for line in tf.gfile.GFile(file_address,"r"))
  return num_lines


def main(_): 

  tf.logging.set_verbosity(tf.logging.INFO)

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

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

  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))

  tf.gfile.MakeDirs(FLAGS.output_dir)

  processor = GECProcessor()

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

  inserts = pickle.load(tf.gfile.Open(FLAGS.path_inserts,"rb"))
  insert_ids = tokenizer.convert_tokens_to_ids(inserts)

  multitoken_inserts = pickle.load(tf.gfile.Open(FLAGS.path_multitoken_inserts,"rb"))
  multitoken_insert_ids = wem_utils.list_to_ids(multitoken_inserts, tokenizer)

  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,
      keep_checkpoint_max=15,
      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)      
      )

  train_examples = None
  num_train_steps = None
  num_warmup_steps = None

  if FLAGS.do_train:
    tf.set_random_seed(FLAGS.random_seed)
    if FLAGS.create_train_tf_records:
      train_examples = processor.get_train_examples(FLAGS.data_dir)
    num_train_examples = get_file_length(os.path.join(FLAGS.data_dir, "train_labels.txt"))
    print("Number of training examples: {}".format(num_train_examples))
    num_train_steps = int(
        (num_train_examples / FLAGS.train_batch_size) * FLAGS.num_train_epochs)
    num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion)

  model_fn = gec_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, 
      copy_weight=FLAGS.copy_weight,
      use_bert_more=FLAGS.use_bert_more,
      inserts=inserts,
      insert_ids=insert_ids,
      multitoken_inserts=multitoken_inserts,
      multitoken_insert_ids=multitoken_insert_ids,
      subtract_replaced_from_replacement=FLAGS.subtract_replaced_from_replacement,
      )  

  # 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,
      eval_batch_size=FLAGS.eval_batch_size,
      predict_batch_size=FLAGS.predict_batch_size)

  if FLAGS.do_train:
    train_record_dir = FLAGS.output_dir
    if FLAGS.create_train_tf_records:
      gec_file_based_convert_examples_to_features(
          train_examples, FLAGS.max_seq_length, train_record_dir, "train", num_train_examples)

    tf.logging.info("***** Running training *****")
    tf.logging.info("  Num examples = %d", num_train_examples)
    tf.logging.info("  Batch size = %d", FLAGS.train_batch_size)
    tf.logging.info("  Num steps = %d", num_train_steps)   

    train_input_fn = gec_file_based_input_fn_builder(
        output_dir=train_record_dir,
        mode="train",
        seq_length=FLAGS.max_seq_length,
        is_training=True,
        drop_remainder=True)
    #train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, max_steps=num_train_steps)
    estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)

  if FLAGS.do_eval:
    eval_examples = processor.get_dev_examples(FLAGS.data_dir)
    #eval_file = os.path.join(FLAGS.output_dir, "eval.tf_record")
    num_eval_examples = get_file_length(os.path.join(FLAGS.data_dir, "dev_labels.txt"))
    gec_file_based_convert_examples_to_features(
        eval_examples, FLAGS.max_seq_length, FLAGS.output_dir, "eval", num_eval_examples)

    tf.logging.info("***** Running evaluation *****")
    tf.logging.info("  Num examples = %d", num_eval_examples)
    tf.logging.info("  Batch size = %d", FLAGS.eval_batch_size)

    # This tells the estimator to run through the entire set.
    eval_steps = None
    # However, if running eval on the TPU, you will need to specify the
    # number of steps.
    if FLAGS.use_tpu:
      # Eval will be slightly WRONG on the TPU because it will truncate
      # the last batch.
      eval_steps = int(num_eval_examples / FLAGS.eval_batch_size)

    eval_drop_remainder = True if FLAGS.use_tpu else False
    eval_input_fn = gec_file_based_input_fn_builder(
        output_dir=FLAGS.output_dir,
        mode="eval",
        seq_length=FLAGS.max_seq_length,
        is_training=False,
        drop_remainder=eval_drop_remainder)

    result = estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps, checkpoint_path=FLAGS.eval_checkpoint)

    output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
    with tf.gfile.GFile(output_eval_file, "w") as writer:
      tf.logging.info("***** Eval results *****")
      for key in sorted(result.keys()):
        tf.logging.info("  %s = %s", key, str(result[key]))
        writer.write("%s = %s\n" % (key, str(result[key])))

  if FLAGS.do_predict:
    if FLAGS.create_predict_tf_records:
      predict_examples = processor.get_test_examples(FLAGS.data_dir)
    num_test_examples = get_file_length(os.path.join(FLAGS.data_dir, "test_incorr.txt"))
    print("num of test_examples: {}".format(num_test_examples))
    num_actual_predict_examples = num_test_examples

    if FLAGS.create_predict_tf_records:
      if FLAGS.use_tpu:
        # Warning: According to tpu_estimator.py Prediction on TPU is an
        # experimental feature and hence not supported here
        #raise ValueError("Prediction in TPU not supported")
        padded_examples = []
        
        while num_test_examples % FLAGS.predict_batch_size != 0:
          padded_examples.append(PaddingInputExample())
          num_test_examples += 1
        
        iterables = [predict_examples, padded_examples]
        predict_examples = chain()
        for iterable in iterables:
          predict_examples = chain(predict_examples, iterable)


      #predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record")
      gec_file_based_convert_examples_to_features(predict_examples,
                                              FLAGS.max_seq_length,
                                              FLAGS.output_dir, "predict", num_test_examples)

      tf.logging.info("***** Running prediction*****")
      tf.logging.info("  Num examples = %d (%d actual, %d padding)",
                      num_test_examples, num_actual_predict_examples,
                      num_test_examples - num_actual_predict_examples)
      tf.logging.info("  Batch size = %d", FLAGS.predict_batch_size)

    predict_drop_remainder = True if FLAGS.use_tpu else False
    predict_input_fn = gec_file_based_input_fn_builder(
        output_dir=FLAGS.output_dir,
        mode="predict",
        seq_length=FLAGS.max_seq_length,
        is_training=False,
        drop_remainder=predict_drop_remainder)
    #os.path.join(FLAGS.data_dir,"reverse_mtedit_dev_lang8_nucle_fce_wi_locness_output_0.1_ensemble_1")

    result = estimator.predict(input_fn=predict_input_fn, checkpoint_path=FLAGS.predict_checkpoint)
    print("type of result: {}".format(type(result)))

    output_predict_file = os.path.join(FLAGS.output_dir, "test_results.txt")
    output_logits_file = os.path.join(FLAGS.output_dir,"test_logits.npz")
    with tf.gfile.GFile(output_predict_file, "w") as writer:
      num_written_lines = 0
      #start_time = time.time()
      total_time_per_step = 0
      #probs_array=[]
      logits_array = []
      tf.logging.info("***** Predict results *****")
      for i,(elapsed_time,prediction) in enumerate(wem_utils.timer(result)):
        if i >= num_actual_predict_examples:
          continue
        total_time_per_step += elapsed_time
        output_line = " ".join(
            str(edit) for edit in prediction["predictions"] if edit > 0) + "\n"
        #logits = np.array(prediction["logits"])
        #logits_array.append(logits)
        writer.write(output_line)
        num_written_lines += 1
    assert num_written_lines == num_actual_predict_examples
    #with tf.gfile.GFile(output_logits_file,"w") as writer:
      #np.save(writer,np.array(logits_array))
    tf.logging.info("Decoding time: {}".format(total_time_per_step))


if __name__ == "__main__":
  flags.mark_flag_as_required("data_dir")
  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()