def write_examples(job_id, args):
"""A single process creating and writing out pre-processed examples."""
job_tmp_dir = os.path.join(args.data_dir, "tmp", "job_" + str(job_id))
owt_dir = os.path.join(args.data_dir, "openwebtext")
def log(*args):
msg = " ".join(map(str, args))
print("Job {}:".format(job_id), msg)
log("Creating example writer")
example_writer = build_pretraining_dataset.ExampleWriter(
job_id=job_id,
vocab_file=os.path.join(args.data_dir, "vocab.txt"),
output_dir=os.path.join(args.data_dir, "pretrain_small_tfrecords"),
max_seq_length=args.max_seq_length,
num_jobs=args.num_processes,
blanks_separate_docs=False,
do_lower_case=args.do_lower_case)
log("Writing tf examples")
fnames = sorted(tf.io.gfile.listdir(owt_dir))
fnames = [
f for (i, f) in enumerate(fnames) if i % args.num_processes == job_id
]
random.shuffle(fnames)
start_time = time.time()
count = 0
for file_no, fname in enumerate(fnames):
if count >= MAX_DATA_ROW:
break
count = count + 1
if file_no > 0 and file_no % 10 == 0:
elapsed = time.time() - start_time
log("processed {:}/{:} files ({:.1f}%), ELAPSED: {:}s, ETA: {:}s, "
"{:} examples written".format(
file_no, len(fnames), 100.0 * file_no / len(fnames),
int(elapsed),
int((len(fnames) - file_no) / (file_no / elapsed)),
example_writer.n_written))
utils.rmkdir(job_tmp_dir)
with tarfile.open(os.path.join(owt_dir, fname)) as f:
f.extractall(job_tmp_dir)
extracted_files = tf.io.gfile.listdir(job_tmp_dir)
random.shuffle(extracted_files)
for txt_fname in extracted_files:
example_writer.write_examples(os.path.join(job_tmp_dir, txt_fname))
example_writer.finish()
log("Done!")
def write_examples(job_id, args):
"""A single process creating and writing out pre-processed examples."""
job_tmp_dir = os.path.join(args.data_dir, "tmp", "job_" + str(job_id))
owt_dir = os.path.join(args.data_dir, "wiki")
def log(*args):
msg = " ".join(map(str, args))
print("Job {}:".format(job_id), msg)
log("Creating example writer")
example_writer = build_pretraining_dataset.ExampleWriter(
job_id=job_id,
model_file=os.path.join(args.model_dir, "wiki-ja.model"),
vocab_file=os.path.join(args.model_dir, "wiki-ja.vocab"),
output_dir=os.path.join(args.model_dir, "pretrain_tfrecords"),
max_seq_length=args.max_seq_length,
num_jobs=args.num_processes,
blanks_separate_docs=False,
do_lower_case=args.do_lower_case)
log("Writing tf examples")
fnames = tf.io.gfile.listdir(owt_dir)
fnames = [f for f in fnames if '.' not in f]
fnames = sorted(fnames)
fnames = [
f for (i, f) in enumerate(fnames) if i % args.num_processes == job_id
]
random.shuffle(fnames)
for file_no, fname in enumerate(fnames):
print('file number : {} of job_id: {}'.format(file_no, job_id))
utils.rmkdir(job_tmp_dir)
copy_tree(os.path.join(owt_dir, fname), job_tmp_dir)
list_files = tf.io.gfile.listdir(job_tmp_dir)
list_files = [fi for fi in list_files if fi != 'all.txt']
for file_name in list_files:
example_writer.write_examples(os.path.join(job_tmp_dir, file_name))
example_writer.finish()
log("Done!")
