inference.py

#!/usr/bin/env python3
"""Run the MelDynamics Polish Metrical HTR model on a single line crop.

Example:
    python inference.py masked_line_crop.png --device cuda
"""
from __future__ import annotations

import argparse

import torch
from PIL import Image
from transformers import TrOCRProcessor, VisionEncoderDecoderModel


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("image", help="one segmented/masked text-line crop")
    ap.add_argument("--model", default="meldynamics/polish-metrical-htr-experimental")
    ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
    ap.add_argument("--max-new-tokens", type=int, default=96)
    args = ap.parse_args()

    # The trained checkpoint includes the compatible processor/tokenizer.  Force the
    # non-square line geometry; the model otherwise falls back to a ViT square size.
    processor = TrOCRProcessor.from_pretrained(args.model, size={"height": 192, "width": 1024})
    model = VisionEncoderDecoderModel.from_pretrained(args.model).to(args.device).eval()
    image = Image.open(args.image).convert("RGB")
    pixels = processor(images=image, return_tensors="pt").pixel_values.to(args.device)
    with torch.inference_mode():
        ids = model.generate(
            pixels,
            max_new_tokens=args.max_new_tokens,
            num_beams=1,
            interpolate_pos_encoding=True,
        )
    print(processor.batch_decode(ids, skip_special_tokens=True)[0])


if __name__ == "__main__":
    main()