inference.py

#!/usr/bin/env python3
"""Standalone DINOv3 script classifier inference (copied to Hub model repo as ``inference.py``)."""

from __future__ import annotations

import argparse
from pathlib import Path

import torch
import torch.nn as nn
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

DINOV3_MODEL_ID = "facebook/dinov3-vits16-pretrain-lvd1689m"


class DINOv3Classifier(nn.Module):
    def __init__(self, model_id: str, num_classes: int, dropout: float = 0.1):
        super().__init__()
        self.backbone = AutoModel.from_pretrained(model_id)
        hidden = self.backbone.config.hidden_size
        self.head = nn.Sequential(
            nn.LayerNorm(hidden),
            nn.Dropout(dropout),
            nn.Linear(hidden, 128),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(128, num_classes),
        )

    def forward(self, pixel_values):
        out = self.backbone(pixel_values=pixel_values)
        cls = out.last_hidden_state[:, 0, :]
        return self.head(cls)


def _resize_short_edge(img: Image.Image, target: int) -> Image.Image:
    w, h = img.size
    if h <= w:
        new_h = target
        new_w = max(1, int(w * target / h))
    else:
        new_w = target
        new_h = max(1, int(h * target / w))
    return img.resize((new_w, new_h), Image.BICUBIC)


def _center_crop(img: Image.Image, size: int = 224) -> Image.Image:
    img = _resize_short_edge(img, size)
    w, h = img.size
    left = max(0, (w - size) // 2)
    top = max(0, (h - size) // 2)
    crop = img.crop((left, top, left + size, top + size))
    if crop.size != (size, size):
        padded = Image.new("RGB", (size, size), (255, 255, 255))
        padded.paste(crop, (0, 0))
        return padded
    return crop


def apply_preprocess(img: Image.Image, mode: str | None, *, size: int = 224) -> Image.Image:
    if not mode or mode == "none":
        return img
    if mode in ("center_crop", "center_crop_whole_page"):
        return _center_crop(img, size)
    raise ValueError(f"Unknown preprocess mode: {mode!r}")


def label_order(ckpt: dict) -> list[str]:
    idx = ckpt.get("idx_to_label") or {}
    if idx:
        return [str(idx[k]) for k in sorted(idx.keys(), key=lambda x: int(x))]
    raw = ckpt.get("label_to_idx") or {}
    if raw:
        return sorted(raw.keys(), key=lambda k: raw[k])
    raise KeyError("checkpoint missing idx_to_label / label_to_idx")


@torch.no_grad()
def predict(model, processor, image_path: Path, device, *, preprocess: str | None, size: int):
    img = Image.open(image_path).convert("RGB")
    img = apply_preprocess(img, preprocess, size=size)
    pv = processor(images=img, return_tensors="pt")["pixel_values"].to(device)
    logits = model(pv)
    probs = torch.softmax(logits, dim=1).squeeze(0).cpu()
    pred = int(probs.argmax())
    return pred, probs.tolist()


def main() -> None:
    ap = argparse.ArgumentParser(description="DINOv3 Tibetan script page classifier")
    ap.add_argument(
        "--checkpoint",
        type=Path,
        default=Path("final_model.pt"),
        help="Weights file (default: final_model.pt in cwd)",
    )
    ap.add_argument("--image", type=Path, nargs="+", required=True)
    ap.add_argument(
        "--preprocess",
        default="none",
        help="none (full page, DINO processor resize) | center_crop (224 short edge)",
    )
    ap.add_argument("--preprocess-size", type=int, default=224)
    ap.add_argument("--model-id", default=DINOV3_MODEL_ID)
    args = ap.parse_args()

    ckpt = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
    classes = label_order(ckpt)
    idx_to_label = {i: lab for i, lab in enumerate(classes)}

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = DINOv3Classifier(args.model_id, num_classes=len(classes)).to(device)
    model.load_state_dict(ckpt["model_state_dict"])
    model.eval()
    processor = AutoImageProcessor.from_pretrained(args.model_id)
    prep = None if args.preprocess in ("none", "") else args.preprocess

    for path in args.image:
        pred, probs = predict(
            model, processor, path, device, preprocess=prep, size=args.preprocess_size
        )
        name = idx_to_label[pred]
        conf = probs[pred]
        print(f"{path.name}: {name} ({conf:.3f})")
        top3 = sorted(enumerate(probs), key=lambda x: -x[1])[:3]
        for i, p in top3:
            print(f"  {idx_to_label[i]:16s} {p:.3f}")


if __name__ == "__main__":
    main()