|
""" |
|
Script to generate WebDataset tar shards from images. |
|
""" |
|
|
|
import argparse |
|
import pickle |
|
import json |
|
import tarfile |
|
from pathlib import Path |
|
from io import BytesIO |
|
|
|
import torch |
|
from torch.utils.data import DataLoader |
|
from torchvision.transforms import ToPILImage |
|
from PIL import Image |
|
from tqdm import tqdm |
|
|
|
from custom_datasets import OnlineFontSquare, TextSampler |
|
|
|
def parse_arguments(): |
|
"""Parses command-line arguments.""" |
|
parser = argparse.ArgumentParser(description='Generate WebDataset tar shards from images') |
|
parser.add_argument('--output_dir', type=str, default='/home/vpippi/font-square-v2/tars/fine_tune', help='Output directory') |
|
parser.add_argument('--fonts', type=str, default='files/font_square/clean_fonts', help='Fonts path') |
|
parser.add_argument('--backgrounds', type=str, default='files/font_square/backgrounds', help='Backgrounds path') |
|
parser.add_argument('--renderers', type=str, help='Renderers path') |
|
parser.add_argument('--db_multiplier', type=int, default=1, help='Dataset multiplier') |
|
parser.add_argument('--dataloader_num_workers', type=int, default=15, help='Dataloader num workers') |
|
parser.add_argument('--shard_size', type=int, default=4000, help='Samples per tar shard') |
|
return parser.parse_args() |
|
|
|
def setup_dataset(args): |
|
"""Initializes dataset and sampler.""" |
|
sampler = TextSampler(4, 128, (1, 32)) |
|
|
|
renderers = None |
|
if args.renderers: |
|
with open(args.renderers, 'rb') as f: |
|
renderers = pickle.load(f) |
|
|
|
dataset = OnlineFontSquare(args.fonts, args.backgrounds, sampler, renderers=renderers) |
|
dataset.length *= args.db_multiplier |
|
return dataset |
|
|
|
def add_bytes_to_tar(tar, filename, data_bytes): |
|
"""Adds a file to the tar archive.""" |
|
ti = tarfile.TarInfo(name=filename) |
|
ti.size = len(data_bytes) |
|
tar.addfile(ti, BytesIO(data_bytes)) |
|
|
|
def process_samples(loader, args): |
|
"""Processes dataset samples and writes them into tar shards.""" |
|
output_dir = Path(args.output_dir) |
|
output_dir.mkdir(parents=True, exist_ok=True) |
|
to_pil = ToPILImage() |
|
|
|
shard_idx, sample_in_shard, total_samples = 0, 0, 0 |
|
tar_path = output_dir / f'{shard_idx:06d}.tar' |
|
tar = tarfile.open(tar_path, mode='w') |
|
|
|
try: |
|
for sample in tqdm(loader, desc="Processing samples"): |
|
text = sample['text'][0].strip() |
|
rgb_img_tensor = sample['img'][0] |
|
bw_img_tensor = sample['text_img'][0] |
|
writer_id = sample['writer'][0].item() |
|
|
|
key = f'{sample_in_shard:06d}' |
|
rgb_pil = to_pil((rgb_img_tensor + 1) / 2) |
|
bw_pil = to_pil((bw_img_tensor + 1) / 2) |
|
|
|
rgb_bytes_io = BytesIO() |
|
rgb_pil.save(rgb_bytes_io, format='PNG') |
|
rgb_bytes = rgb_bytes_io.getvalue() |
|
|
|
bw_bytes_io = BytesIO() |
|
bw_pil.save(bw_bytes_io, format='PNG') |
|
bw_bytes = bw_bytes_io.getvalue() |
|
|
|
metadata = json.dumps({"text": text, "writer_id": writer_id}).encode('utf-8') |
|
|
|
add_bytes_to_tar(tar, f'{key}.rgb.png', rgb_bytes) |
|
add_bytes_to_tar(tar, f'{key}.bw.png', bw_bytes) |
|
add_bytes_to_tar(tar, f'{key}.json', metadata) |
|
|
|
sample_in_shard += 1 |
|
total_samples += 1 |
|
|
|
if sample_in_shard >= args.shard_size: |
|
tar.close() |
|
shard_idx += 1 |
|
sample_in_shard = 0 |
|
tar_path = output_dir / f'{shard_idx:06d}.tar' |
|
tar = tarfile.open(tar_path, mode='w') |
|
|
|
except KeyboardInterrupt: |
|
print("Interrupted by user.") |
|
finally: |
|
tar.close() |
|
|
|
print(f"Finished writing {total_samples} samples in {shard_idx+1} tar shards into {output_dir}") |
|
|
|
def main(): |
|
args = parse_arguments() |
|
dataset = setup_dataset(args) |
|
loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=args.dataloader_num_workers) |
|
process_samples(loader, args) |
|
|
|
if __name__ == '__main__': |
|
main() |
|
|
