| |
| |
| """Write `preprocess-cases.json`: for every image in tests/images/, the SHA-256 of the decoded 8-bit RGB buffer and |
| of the resized buffer, plus the sizes and the soft-token count, computed with the Python reference preprocessing |
| (python/eg2_image.py, itself bit-identical to torchvision's uint8 bicubic). `eg2-cli pixcheck` compares the Swift |
| decoder + resampler against it. Lossy (JPEG) cases are flagged: decoders differ by a few values per sample there. |
| |
| python3 make_preprocess_cases.py [--images images] [--out preprocess-cases.json] |
| """ |
| from __future__ import annotations |
|
|
| import argparse |
| import hashlib |
| import json |
| import sys |
| from pathlib import Path |
|
|
| HERE = Path(__file__).resolve().parent |
| sys.path.insert(0, str(HERE.parent / "python")) |
| import eg2_image |
|
|
|
|
| def main() -> None: |
| ap = argparse.ArgumentParser(description=__doc__.splitlines()[0]) |
| ap.add_argument("--images", default=str(HERE / "images")) |
| ap.add_argument("--out", default=str(HERE / "preprocess-cases.json")) |
| args = ap.parse_args() |
| items = json.loads((Path(args.images) / "manifest.json").read_text()) |
| cases = [] |
| for it in items: |
| rgb = eg2_image.load_rgb(Path(args.images) / it["file"]) |
| th, tw = eg2_image.target_size(rgb.shape[0], rgb.shape[1]) |
| res = eg2_image.resize_bicubic_aa_u8(rgb, th, tw) |
| vi = eg2_image.patchify_cell_major(res) |
| cases.append(dict(file=f"{Path(args.images).name}/{it['file']}", width=int(rgb.shape[1]), height=int(rgb.shape[0]), |
| resized_width=tw, resized_height=th, n_soft=vi.n_soft, lossy=bool(it.get("lossy", False)), |
| decoded_sha256=hashlib.sha256(rgb.tobytes()).hexdigest(), resized_sha256=hashlib.sha256(res.tobytes()).hexdigest())) |
| print(cases[-1]["file"], f"{rgb.shape[1]}x{rgb.shape[0]} -> {tw}x{th}, n={vi.n_soft}") |
| Path(args.out).write_text(json.dumps(cases, indent=1) + "\n") |
| print("wrote", args.out) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
