def get_examples(data_dir, output_data_dir, mode, task_id, shard_id):
"""Reads NQ data, does sling entity linking and returns augmented data."""
file_path = nq_data_utils.get_sharded_filename(data_dir, mode, task_id,
shard_id, 'jsonl.gz')
output_file_path = nq_data_utils.get_sharded_filename(
output_data_dir, mode, task_id, shard_id, 'tsv')
print(file_path)
tf.logging.info("Reading file: %s" % (file_path))
if not os.path.exists(file_path):
return None, None
nq_data, entities = extract_nq_data(file_path, output_file_path)
tf.logging.info("NQ data Size: " + str(len(nq_data.keys())))
return nq_data, entities
def get_output_fp(shard_counter):
output_file = nq_data_utils.get_sharded_filename(FLAGS.output_data_dir,
FLAGS.split,
FLAGS.task_id,
shard_counter, "jsonl.gz")
op = gzip.GzipFile(fileobj=tf.gfile.Open(output_file, "w"))
return op
def main(_):
# input_file = nq_data_utils.get_sharded_filename(FLAGS.input_data_dir,
# FLAGS.split, FLAGS.task_id,
# FLAGS.shard_split_id,
# "jsonl.gz")
# print("Reading file %s", input_file)
# output_file = nq_data_utils.get_sharded_filename(FLAGS.output_data_dir,
# FLAGS.split, FLAGS.task_id,
# FLAGS.shard_split_id,
# "jsonl")
# count = nq_jsonl_to_tsv(input_file, output_file)
# print(count)
input_data_dir = '/remote/bones/user/vbalacha/fact_augmented_text/fat/fat_bert_nq/generated_files/nq_t5_data_new/'
stats = {'train':(50,7), 'dev':(4,16)}
for split in ['train', 'dev']:
if split == 'train':
output_file = os.path.join(input_data_dir, "all_train.jsonl")
else:
output_file = os.path.join(input_data_dir, "all_dev.jsonl")
op = open(output_file, 'w')
count = 0
for task in range(stats[split][0]):
for shard_split in range(stats[split][1]):
input_file = nq_data_utils.get_sharded_filename(input_data_dir,
split, task,
shard_split,
"jsonl")
if not os.path.exists(input_file):
continue
print("Reading file %s", input_file)
fp = open(input_file)
for line in fp:
count += 1
op.write(line+"\n")
print(count)
def main(_):
examples_processed = 0
instances_processed = 0
num_examples_with_correct_context = 0
if FLAGS.is_training:
creator_fn = run_nq.CreateTFExampleFn(is_training=FLAGS.is_training)
instances = []
input_file = nq_data_utils.get_sharded_filename(FLAGS.input_data_dir,
FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id,
"jsonl.gz")
tf.logging.info("Reading file %s", input_file)
for example in nq_data_utils.get_nq_examples(input_file):
for instance in creator_fn.process(example):
instances.append(instance)
instances_processed += 1
if example["has_correct_context"]:
num_examples_with_correct_context += 1
if examples_processed % 100 == 0:
tf.logging.info("Examples processed: %d", examples_processed)
tf.logging.info("Instances processed: %d", instances_processed)
examples_processed += 1
if FLAGS.max_examples > 0 and examples_processed >= FLAGS.max_examples:
break
tf.logging.info("Examples with correct context retained: %d of %d",
num_examples_with_correct_context, examples_processed)
random.shuffle(instances)
tf.logging.info("Total no: of instances in current shard: %d",
len(instances))
output_file = nq_data_utils.get_sharded_filename(FLAGS.output_data_dir,
FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id,
"tf-record")
with tf.python_io.TFRecordWriter(output_file) as writer:
for instance in instances:
writer.write(instance)
# For eval - First process every shard in parallel
elif not FLAGS.is_training and not FLAGS.merge_eval:
input_file = nq_data_utils.get_sharded_filename(FLAGS.input_data_dir,
FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id,
"jsonl.gz")
tf.logging.info("Reading file %s", input_file)
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)
eval_examples = run_nq.read_nq_examples(
input_file=input_file, is_training=False)
output_file = nq_data_utils.get_sharded_filename(FLAGS.output_data_dir,
FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id,
"tf-record")
eval_writer = run_nq.FeatureWriter(filename=output_file, is_training=False)
eval_features = []
examples_processed = 0
def append_feature(feature):
eval_features.append(feature)
eval_writer.process_feature(feature)
examples_processed = len(eval_features)
if examples_processed % 100 == 0:
tf.logging.info("Examples processed: %d", examples_processed)
_ = run_nq.convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
is_training=False,
output_fn=append_feature)
eval_writer.close()
# For eval - Fianlly merge all shards into 1
else:
instances = []
for task in range(FLAGS.max_dev_tasks):
for shard_split in range(FLAGS.max_dev_shard_splits):
input_file = nq_data_utils.get_sharded_filename(FLAGS.input_data_dir,
FLAGS.split, task,
shard_split,
"tf-record")
tf.logging.info("Reading file %s", input_file)
instances.extend([
tf.train.Example.FromString(r)
for r in tf.python_io.tf_record_iterator(input_file)
])
output_file = os.path.join(FLAGS.output_data_dir, "eval.tf-record")
with tf.python_io.TFRecordWriter(output_file) as writer:
for instance in instances:
writer.write(instance.SerializeToString())
'Read nq data to extract entities')
flags.DEFINE_string('relation_emb_file', '',
'Read nq data to extract entities')
#flags.DEFINE_string('relations_file', None, 'input relations dict')
#flags.DEFINE_string('rel2id_file', None, 'input relations dict')
flags.DEFINE_string('output_file', None, '')
flags.DEFINE_integer("shard_id", None,
"Train and dev shard to read from and write to.")
flags.DEFINE_string(
"mode", "train",
"Train and dev split to read from and write to. Accepted values: ['train', 'dev', 'test']"
)
input_file = nq_data_utils.get_sharded_filename(FLAGS.nq_dir, FLAGS.mode,
FLAGS.task_id, FLAGS.shard_id,
'jsonl.gz')
embeddings_file = "/remote/bones/user/vbalacha/datasets/glove/glove.6B.300d.txt"
#with gzip.GzipFile(fileobj=tf.gfile.Open(FLAGS.rel2id_file, 'rb')) as op4:
# rel2id = json.load(op4)
# op4.close()
#id2rel = {str(idx): ent for ent, idx in rel2id.items()}
dim = 300
apr_obj = ApproximatePageRank(mode='train',
task_id=FLAGS.task_id,
shard_id=FLAGS.shard_id)
#apr_obj = ApproximatePageRank()
#fp = open("/remote/bones/user/vbalacha/google-research/fat/fat/fat_bert_nq/generated_files/tmpdir_unk-1_/unique_id_list_0000.txt", "w")
name_to_features = {
"unique_ids": tf.FixedLenFeature([], tf.int64),
"input_ids": tf.FixedLenFeature([512], tf.int64),
"input_mask": tf.FixedLenFeature([512], tf.int64),
"segment_ids": tf.FixedLenFeature([512], tf.int64),
}
#instances = []
train_count = 0
for task in range(max_train_tasks):
for shard in range(max_shard_splits):
#for task in [0]:
# for shard in [0]:
input_file = nq_utils.get_sharded_filename(input_data_dir, "train",
task, shard, 'tf-record')
print("Reading file %s", input_file)
if not os.path.exists(input_file):
continue
#instances.extend([
# tf.train.Example.FromString(r)
# for r in tf.python_io.tf_record_iterator(input_file)
#])
#example = tf.parse_single_example(record, name_to_features)
for record in tf.python_io.tf_record_iterator(input_file):
#example = tf.parse_single_example(record, name_to_features)
example = tf.train.Example.FromString(record)
unique_id = example.features.feature[
'unique_ids'].int64_list.value[0]
fp.write(str(unique_id) + "\n")
#print("Training size: "+str(count))
shard_id, 'jsonl.gz')
print(file_path)
tf.logging.info("Reading file: %s" % (file_path))
if not os.path.exists(file_path):
return None, None
nq_data, entities = extract_nq_data(file_path)
tf.logging.info("NQ data Size: " + str(len(nq_data.keys())))
return nq_data, entities
if __name__ == '__main__':
print(FLAGS.full_wiki)
print(FLAGS.decompose_ppv)
print(FLAGS.apr_files_dir)
question_emb_file = nq_data_utils.get_sharded_filename(
FLAGS.question_emb_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, 'pkl')
question_embeddings = pkl.load(open(question_emb_file, 'rb'))
relation_embeddings = pkl.load(open(FLAGS.relation_emb_file, 'rb'))
max_tasks = {"train": 50, "dev": 5}
max_shards = {"train": 7, "dev": 17}
apr = ApproximatePageRank(FLAGS.split, FLAGS.task_id, FLAGS.shard_split_id)
empty_ents = 0
for mode in [FLAGS.split]:
# Parse all shards in each mode
# Currently sequentially, can be parallelized later
for task_id in [FLAGS.task_id]: #range(0, max_tasks[mode]):
for shard_id in [FLAGS.shard_split_id
]: #range(0, max_shards[mode]):
# if task_id == 0 and shard_id in range(0, 16):
# print("skipping finished job")
flags.DEFINE_string('nq_dir',
'/remote/bones/user/vbalacha/datasets/ent_linked_nq_new/',
'Read nq data to extract entities')
#flags.DEFINE_string('output_dir', '/remote/bones/user/vbalacha/datasets/nq_question_embeddings/', 'Read nq data to extract entities')
flags.DEFINE_integer("shard_id", None,
"Train and dev shard to read from and write to.")
#flags.DEFINE_integer("task_id", None,
# "Train and dev shard to read from and write to.")
flags.DEFINE_string(
"mode", "train",
"Train and dev split to read from and write to. Accepted values: ['train', 'dev', 'test']"
)
flags.DEFINE_bool("filter_sw", False, "")
questions_file = nq_data_utils.get_sharded_filename(FLAGS.nq_dir, FLAGS.mode,
FLAGS.task_id,
FLAGS.shard_id, 'jsonl.gz')
embeddings_file = "/remote/bones/user/vbalacha/datasets/glove/glove.6B.300d.txt"
output_file = nq_data_utils.get_sharded_filename(FLAGS.output_dir, FLAGS.mode,
FLAGS.task_id, FLAGS.shard_id,
'pkl')
dim = 300
stopwords = set(SW.words("english"))
stopwords.add("'s")
word_to_question = {}
question_lens = {}
def extract_keywords(text):
"""Remove wh-words and stop words from text."""
def main(_):
examples_processed = 0
instances_processed = 0
examples_with_instances = 0
num_examples_with_correct_context = 0
# if FLAGS.create_pretrain_data or FLAGS.create_fact_annotation_data:
# pretrain_file = open(nq_data_utils.get_sharded_filename(FLAGS.pretrain_data_dir,
# FLAGS.split, FLAGS.task_id,
# FLAGS.shard_split_id,
# "txt"), 'w')
# else:
# pretrain_file = None
pretrain_file = open(
nq_data_utils.get_sharded_filename(FLAGS.pretrain_data_dir,
FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "txt"), 'w')
if FLAGS.is_training:
fixed_train_list = None
if FLAGS.use_fixed_training_data:
fp = open(FLAGS.fixed_training_data_filepath)
fixed_train_list = []
for line in fp:
fixed_train_list.append(int(line))
creator_fn = run_nq.CreateTFExampleFn(is_training=FLAGS.is_training)
instances = []
input_file = nq_data_utils.get_sharded_filename(
FLAGS.input_data_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "jsonl.gz")
print("Reading file %s", input_file)
file_stats_counter = {
'example_count': 0,
'sp_recall_sum': 0,
'answer_reach_counter': 0,
'single_answer_reach_counter': 0,
'multi_answer_recall': 0,
'single_answer_counter': 0,
'multi_answer_counter': 0,
'multi_answer_size_counter': 0
}
for example in nq_data_utils.get_nq_examples(input_file):
ins_count = 0
stats_count = None
for instance, stats_count in creator_fn.process(
example, pretrain_file, fixed_train_list):
instances.append(instance)
instances_processed += 1
ins_count = 1
if FLAGS.use_passage_rw_facts_in_shortest_path or FLAGS.use_question_to_passage_facts_in_shortest_path:
file_stats_counter['example_count'] += 1
file_stats_counter['sp_recall_sum'] += stats_count[
'fact_recall_counter']
if len(stats_count['answers_reached']) > 0:
file_stats_counter['answer_reach_counter'] += 1
if len(stats_count['answer_entity_ids']) > 1:
file_stats_counter[
'multi_answer_recall'] += stats_count[
'answer_recall_counter']
else:
file_stats_counter[
'single_answer_reach_counter'] += 1
if len(stats_count['answer_entity_ids']) > 1:
file_stats_counter['multi_answer_counter'] += 1
file_stats_counter['multi_answer_size_counter'] += len(
stats_count['answer_entity_ids'])
else:
file_stats_counter['single_answer_counter'] += 1
if stats_count is not None and FLAGS.use_question_rw_facts_in_shortest_path:
file_stats_counter['example_count'] += 1
file_stats_counter['sp_recall_sum'] += stats_count[
'fact_recall_counter']
if len(stats_count['answers_reached']) > 0:
file_stats_counter['answer_reach_counter'] += 1
if len(stats_count['answer_entity_ids']) > 1:
file_stats_counter[
'multi_answer_recall'] += stats_count[
'answer_recall_counter']
else:
file_stats_counter['single_answer_reach_counter'] += 1
if len(stats_count['answer_entity_ids']) > 1:
file_stats_counter['multi_answer_counter'] += 1
file_stats_counter['multi_answer_size_counter'] += len(
stats_count['answer_entity_ids'])
else:
file_stats_counter['single_answer_counter'] += 1
if example["has_correct_context"]:
num_examples_with_correct_context += 1
if examples_processed % 100 == 0:
print("Examples processed: %d", examples_processed)
print("Instances processed: %d", instances_processed)
examples_processed += 1
examples_with_instances += ins_count
if FLAGS.max_examples > 0 and examples_processed >= FLAGS.max_examples:
break
#time.sleep(5)
print("Examples with correct context retained: %d of %d",
num_examples_with_correct_context, examples_processed)
random.shuffle(instances)
print("Total no: of instances in current shard: %d", len(instances))
output_file = nq_data_utils.get_sharded_filename(
FLAGS.output_data_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "tf-record")
stats_file = nq_data_utils.get_sharded_filename(
FLAGS.output_data_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "stats.txt")
with tf.python_io.TFRecordWriter(output_file) as writer:
for instance in instances:
writer.write(instance)
with open(stats_file, 'w') as fp:
print("Example count: %d", file_stats_counter['example_count'])
print("Fact Recall sum: %d", file_stats_counter['sp_recall_sum'])
print("Count with answers reached: %d",
file_stats_counter['answer_reach_counter'])
print("Single Answer Example count: %d",
file_stats_counter['single_answer_counter'])
print("Single Answer Reached count: %d",
file_stats_counter['single_answer_reach_counter'])
print("Multi Answer Example count: %d",
file_stats_counter['multi_answer_counter'])
print("Multi Answer recall sum: %d",
file_stats_counter['multi_answer_recall'])
print("Multi Answer Size counter: %d",
file_stats_counter['multi_answer_size_counter'])
fp.write("Example count: " +
str(file_stats_counter['example_count']) + "\n")
fp.write("Fact Recall sum: " +
str(file_stats_counter['sp_recall_sum']) + "\n")
fp.write("Count with answers reached: " +
str(file_stats_counter['answer_reach_counter']) + "\n")
fp.write("Single Answer Example count: " +
str(file_stats_counter['single_answer_counter']) + "\n")
fp.write("Single Answer Reached count: " +
str(file_stats_counter['single_answer_reach_counter']) +
"\n")
fp.write("Multi Answer Example count: " +
str(file_stats_counter['multi_answer_counter']) + "\n")
fp.write("Multi Answer recall sum: " +
str(file_stats_counter['multi_answer_recall']) + "\n")
fp.write("Multi Answer Size counter: " +
str(file_stats_counter['multi_answer_size_counter']) +
"\n")
fp.write("Examples with instances Processed: " +
str(examples_with_instances) + "\n")
# For eval - First process every shard in parallel
elif not FLAGS.is_training and not FLAGS.merge_eval:
input_file = nq_data_utils.get_sharded_filename(
FLAGS.input_data_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "jsonl.gz")
print("Reading file %s", input_file)
tokenizer = tokenization.FullTokenizer(
vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)
eval_examples = run_nq.read_nq_examples(input_file=input_file,
is_training=False)
output_file = nq_data_utils.get_sharded_filename(
FLAGS.output_data_dir, FLAGS.split, FLAGS.task_id,
FLAGS.shard_split_id, "tf-record")
eval_writer = run_nq.FeatureWriter(filename=output_file,
is_training=False)
eval_features = []
examples_processed = 0
def append_feature(feature):
eval_features.append(feature)
eval_writer.process_feature(feature)
examples_processed = len(eval_features)
if examples_processed % 10 == 0:
print("Examples processed: %d", examples_processed)
_ = run_nq.convert_examples_to_features(examples=eval_examples,
tokenizer=tokenizer,
is_training=False,
output_fn=append_feature,
pretrain_file=pretrain_file)
eval_writer.close()
# For eval - Fianlly merge all shards into 1
else:
instances = []
for task in range(FLAGS.max_dev_tasks):
for shard_split in range(FLAGS.max_dev_shard_splits):
input_file = nq_data_utils.get_sharded_filename(
FLAGS.input_data_dir, FLAGS.split, task, shard_split,
"tf-record")
print("Reading file %s", input_file)
instances.extend([
tf.train.Example.FromString(r)
for r in tf.python_io.tf_record_iterator(input_file)
])
output_file = os.path.join(FLAGS.output_data_dir, "eval.tf-record")
with tf.python_io.TFRecordWriter(output_file) as writer:
for instance in instances:
writer.write(instance.SerializeToString())
