"""Reference-free confidence of recognizer outputs: mean negative log-likelihood of the (loop-cut) predicted
transliteration under the recognizer that produced it, given the image and the recognition prompt.

Input: a prediction file with {"id", "image", "pred", ...}. Output: one line per row,
{"id", "nll", "n_tokens"}; nll is null for empty predictions. Reuses the training example builder
(train_lora.Examples) with the prediction as the target and without the end-of-turn token. Resumable.
"""
import argparse
import json
import math
import os
import sys

import torch

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from qwen_common import load_model  # noqa: E402
from train_lora import Examples  # noqa: E402


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", required=True)
    ap.add_argument("--adapter", required=True)
    ap.add_argument("--input", required=True)
    ap.add_argument("--image_root", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--device", default="cuda")
    ap.add_argument("--limit", type=int, default=0)  # smoke tests
    args = ap.parse_args()
    rows = [json.loads(l) for l in open(args.input) if l.strip()]
    done = set()
    if os.path.exists(args.out):
        done = {json.loads(l)["id"] for l in open(args.out) if l.strip()}
    todo = [dict(r, target=r["pred"] or "") for r in rows if r["id"] not in done]
    if args.limit:
        todo = todo[:args.limit]
    print(f"rows {len(rows)} done {len(done)} todo {len(todo)}", flush=True)
    if not todo:
        return
    processor, model = load_model(args.model, adapter=args.adapter)
    model.to(args.device).eval()
    ds = Examples(todo, processor, "ocr", args.image_root)
    ds.end = ""  # score the transliteration only, not the end-of-turn token
    fout = open(args.out, "a")
    for i, r in enumerate(todo):
        rec = {"id": r["id"], "nll": None, "n_tokens": 0}
        if r["target"].strip():
            try:
                item = ds[i]
                n = int((item["labels"] != -100).sum())
                batch = {k: v.to(args.device) for k, v in item.items()}
                with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16):
                    loss = model(**batch).loss.item()
                rec.update(nll=loss if math.isfinite(loss) else None, n_tokens=n)
            except Exception as e:  # keep going; the row is reported as unscored
                rec["error"] = repr(e)[:200]
        fout.write(json.dumps(rec) + "\n")
        fout.flush()
        if (i + 1) % 50 == 0:
            print(f"{i + 1}/{len(todo)}", flush=True)
    print("DONE", flush=True)


if __name__ == "__main__":
    main()