| from argparse import ArgumentParser |
| from transformers import AutoTokenizer |
| import os |
| from tqdm import tqdm |
| from multiprocessing import Pool |
| from dense.processor import SimpleCollectionProcessor |
|
|
| parser = ArgumentParser() |
| parser.add_argument('--tokenizer_name', required=True) |
| parser.add_argument('--truncate', type=int, default=128) |
| parser.add_argument('--file', required=True) |
| parser.add_argument('--save_to', required=True) |
| parser.add_argument('--n_splits', type=int, default=10) |
|
|
| args = parser.parse_args() |
|
|
| tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=True) |
| processor = SimpleCollectionProcessor(tokenizer=tokenizer, max_length=args.truncate) |
|
|
| with open(args.file, 'r') as f: |
| lines = f.readlines() |
|
|
| n_lines = len(lines) |
| if n_lines % args.n_splits == 0: |
| split_size = int(n_lines / args.n_splits) |
| else: |
| split_size = int(n_lines / args.n_splits) + 1 |
|
|
|
|
| os.makedirs(args.save_to, exist_ok=True) |
| with Pool() as p: |
| for i in range(args.n_splits): |
| with open(os.path.join(args.save_to, f'split{i:02d}.json'), 'w') as f: |
| pbar = tqdm(lines[i*split_size: (i+1)*split_size]) |
| pbar.set_description(f'split - {i:02d}') |
| for jitem in p.imap(processor.process_line, pbar, chunksize=500): |
| f.write(jitem + '\n') |
|
|
|
|
|
|
