import json from pathlib import Path import cv2 import numpy as np import onnxruntime as ort from indra.indra_ocr.vocab import Vocab from indra.indra_ocr.rec_preproc import pad_batch from tools.exp.common import collapseIds, lev, norm home = Path.home() data = home / "ocr/ocr-deva-ft/data/chunks-v4/test" manifest = Path("results/stratified_1500/sample_manifest.jsonl") model = Path("models/indra-ocr-v3.1.1/rec.onnx") outdir = Path("results/padding_variants") outdir.mkdir(parents=True, exist_ok=True) if not manifest.is_file(): raise FileNotFoundError(f"Missing baseline manifest: {manifest}") rows = [json.loads(line) for line in manifest.read_text(encoding="utf-8").splitlines() if line.strip()] if len(rows) != 1500: raise RuntimeError(f"Expected 1500 manifest rows, got {len(rows)}") session = ort.InferenceSession(str(model), providers=["CPUExecutionProvider"]) vocab = Vocab.multi() variants = { "baseline": (0.00, 0.00), "small_pad": (0.05, 0.10), "moderate_pad": (0.20, 0.30), } results = [] for i, row in enumerate(rows, 1): path = data / row["image"] gray = cv2.imread(str(path), cv2.IMREAD_GRAYSCALE) if gray is None: raise RuntimeError(f"Cannot read {path}") h, w = gray.shape for name, (pad_y, pad_x) in variants.items(): py, px = round(h * pad_y), round(h * pad_x) padded = cv2.copyMakeBorder( gray, py, py, px, px, cv2.BORDER_CONSTANT, value=255 ) ph, pw = padded.shape target_w = max(8, min(4096, round(pw * 32 / ph))) interp = cv2.INTER_AREA if ph > 32 else cv2.INTER_LINEAR resized = cv2.resize(padded, (target_w, 32), interpolation=interp) ink = 1.0 - resized.astype(np.float32) / 255.0 x, widths = pad_batch([ink]) ids, conf, lengths = session.run( ["ids", "conf", "lengths"], {"x": x, "widths": widths.astype(np.int64)} ) seq = ids[0, :int(lengths[0])].tolist() pred = vocab.decode(collapseIds(seq)) ref, hyp = norm(row["text"]), norm(pred) results.append({ "image": row["image"], "edition": row["edition"], "variant": name, "ground_truth": row["text"], "prediction": pred, "edits": lev(ref, hyp), "reference_chars": len(ref), "exact_match": int(ref == hyp), }) if i % 100 == 0: print(f"Processed {i}/{len(rows)} images") output = outdir / "predictions.jsonl" with output.open("w", encoding="utf-8") as f: for r in results: f.write(json.dumps(r, ensure_ascii=False) + "\n") print("\n=== PADDING VARIANT RESULTS ===") for edition in ["2082-03-24", "2082-08-12", "2082-12-25", "ALL"]: print(f"\nEdition: {edition}") for name in variants: subset = [r for r in results if r["variant"] == name and (edition == "ALL" or r["edition"] == edition)] edits = sum(r["edits"] for r in subset) chars = sum(r["reference_chars"] for r in subset) exact = sum(r["exact_match"] for r in subset) print(f" {name}: exact={exact}/{len(subset)} ({100*exact/max(len(subset),1):.2f}%), CER={100*edits/max(chars,1):.4f}% ({edits}/{chars})") print("\nPredictions:", output) print("Providers:", session.get_providers())