-
Notifications
You must be signed in to change notification settings - Fork 1
/
preprocess.py
39 lines (34 loc) · 1.73 KB
/
preprocess.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
# This script launches the jobs defined in `preprocessing`.
import argparse
import logging
import threading
import multiprocessing
multiprocessing.set_start_method('fork') # only on unix
from dataset import *
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--dest-prefix', type=str, required=True, help='Export the processed files to that S3 directory.')
parser.add_argument('--job', type=str, required=True, help='Name of the preprocessing job to run.')
# Optional
parser.add_argument('--bucket', type=str, default="gopilot", help='Name of the S3 bucket to download the dataset files from.')
parser.add_argument('--source-prefix', type=str, help='Feed all files in that S3 directory to the job.')
parser.add_argument('--cache-dir', type=str, default=".cache", help='Local mirror of the S3 bucket. Files are not downloaded if they already exist in the cache.')
args = parser.parse_args()
# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(message)s')
job_name = args.job
delattr(args, "job")
pool = multiprocessing.Pool()
if job_name == "tokenize-with-huggingface":
job = TokenizeWithHuggingFaceJob(**vars(args))
elif job_name == "tokenize-with-gopilot":
job = TokenizeWithGopilotJob(**vars(args))
elif job_name == "train-huggingface-tokenizer":
job = TrainHuggingFaceTokenizerJob(**vars(args))
elif job_name == "train-gopilot-tokenizer":
job = TrainGopilotTokenizerJob(**vars(args))
elif job_name == "upload-the-stack":
job = UploadTheStackJob(**vars(args))
else:
raise ValueError(f"Unknown job {args.job}")
job.run()