# SPDX-License-Identifier: Apache-2.0 """Agreement of this package with the fp32 torch CPU reference on real hand images. Run the same list of (image, hand box, is_right) through one implementation and save the outputs: # fp32 torch CPU reference (hamer.reference.Hamer with the released checkpoint; no device) python eval_agreement.py run --impl cpu --samples samples.json --root --out ref.npz # this package through the Python API on the chip (published configuration) python eval_agreement.py run --impl tt --samples samples.json --root --out tt.npz # this package through a running server (POST /predict, what a client receives) python eval_agreement.py run --impl http --url http://127.0.0.1:20000 --samples samples.json --root --out http.npz and compare two output files: python eval_agreement.py compare --ref ref.npz --test http.npz --json agreement.json ``samples.json`` is a list of ``{"id", "set", "image" (path relative to --root), "bbox": [x1, y1, x2, y2], "is_right"}``. Every implementation does the same host steps (decode, square crop of 2x the box, resize, mirror for a left hand) with the package's own preprocessing; the CPU reference runs on the normalized float crop and the chip on the uint8 crop, as served. Metrics per set (``compare``): PCC of the 157-value regression output (16 rotation matrices, 10 betas, crop camera), the geodesic angle between corresponding rotation matrices (global orientation and the 15 MANO joint rotations, in degrees), the absolute betas error, the full-frame camera translation error ``|cam_t_full - ref|`` in millimetres (MANO units are metres) and its relative depth error. Without a MANO model file the outputs are MANO parameters, so joint and vertex errors are not computed here. """ from __future__ import annotations import argparse import base64 import json import os import sys import time from pathlib import Path import numpy as np def load_samples(path): return json.loads(Path(path).read_text()) def run(a): samples = load_samples(a.samples) if a.limit: samples = samples[: a.limit] root = Path(a.root) n = len(samples) out = {"rotmats": np.zeros((n, 16, 3, 3)), "betas": np.zeros((n, 10)), "cam": np.zeros((n, 3)), "cam_t_full": np.zeros((n, 3)), "time_ms": np.zeros(n)} t_start = time.time() if a.impl == "http": import http.client import urllib.parse u = urllib.parse.urlparse(a.url) conn = http.client.HTTPConnection(u.hostname, u.port or 80, timeout=120) cache = {} for i, s in enumerate(samples): p = root / s["image"] if p not in cache: cache = {p: base64.b64encode(p.read_bytes()).decode()} body = json.dumps({"image": cache[p], "bbox": s["bbox"], "is_right": bool(s["is_right"])}).encode() t0 = time.perf_counter() conn.request("POST", "/predict", body=body, headers={"Content-Type": "application/json"}) r = conn.getresponse() raw = r.read() out["time_ms"][i] = (time.perf_counter() - t0) * 1e3 if r.status != 200: raise SystemExit(f"{s['id']}: HTTP {r.status} {raw[:300]!r}") d = json.loads(raw) for k in ("rotmats", "betas", "cam", "cam_t_full"): out[k][i] = np.asarray(d[k], dtype=np.float64) else: import torch from PIL import Image from hamer import HamerModel, load_reference if a.threads: torch.set_num_threads(a.threads) if a.impl == "cpu": ref = load_reference() model = HamerModel.from_reference(ref, backend="cpu", warmup_variants=None) else: model = HamerModel.from_pretrained(device_id=0, warmup_variants=None) print(f"[eval] impl={a.impl} info={json.dumps({k: model.info.get(k) for k in ('backend', 'dispatch', 'grid', 'warmup')})}", flush=True) img_cache = {} for i, s in enumerate(samples): p = root / s["image"] if p not in img_cache: img_cache = {p: Image.open(p).convert("RGB")} t0 = time.perf_counter() h = model(img_cache[p], bbox=[float(v) for v in s["bbox"]], is_right=bool(s["is_right"])) out["time_ms"][i] = (time.perf_counter() - t0) * 1e3 out["rotmats"][i], out["betas"][i], out["cam"][i], out["cam_t_full"][i] = h.rotmats, h.betas, h.cam, h.cam_t_full if (i + 1) % 50 == 0 or i + 1 == n: el = time.time() - t_start print(f"[eval] {i + 1}/{n} ({el:.0f}s, {el / (i + 1):.2f}s/hand)", flush=True) model.close() np.savez(a.out, ids=np.array([s["id"] for s in samples]), sets=np.array([s["set"] for s in samples]), impl=a.impl, wall_s=time.time() - t_start, **out) print(f"[eval] wrote {a.out}: {n} hands in {time.time() - t_start:.0f}s") def pcc(u, v): u = u - u.mean() v = v - v.mean() return float(u @ v / max(np.linalg.norm(u) * np.linalg.norm(v), 1e-12)) def geodesic_deg(ra, rb): """Rotation angle between (..., 3, 3) rotation matrices ra and rb, degrees. Uses |ra - rb|_F = 2 sqrt(2) sin(theta / 2), which stays accurate for small angles (arccos of the trace does not: a 1e-6 trace error alone reads as about 0.08 degrees).""" f = np.linalg.norm(ra - rb, axis=(-2, -1)) return np.degrees(2.0 * np.arcsin(np.clip(f / (2.0 * np.sqrt(2.0)), 0.0, 1.0))) def stats(v): v = np.asarray(v, dtype=np.float64).reshape(-1) return {"mean": float(v.mean()), "median": float(np.median(v)), "p99": float(np.percentile(v, 99)), "max": float(v.max()), "min": float(v.min()), "n": int(v.size)} def compare(a): r, t = np.load(a.ref), np.load(a.test) if list(r["ids"]) != list(t["ids"]): # align on ids (the test run may cover a prefix) idx = {k: i for i, k in enumerate(r["ids"])} sel = [idx[k] for k in t["ids"]] else: sel = list(range(len(r["ids"]))) R = {k: r[k][sel] for k in ("rotmats", "betas", "cam", "cam_t_full")} T = {k: t[k] for k in ("rotmats", "betas", "cam", "cam_t_full")} sets = t["sets"] res = {"ref": str(a.ref), "test": str(a.test), "ref_impl": str(r["impl"]), "test_impl": str(t["impl"]), "sets": {}} for name in ["all"] + sorted(set(sets.tolist())): m = np.ones(len(sets), bool) if name == "all" else sets == name vr = np.concatenate([R["rotmats"][m].reshape(-1, 144), R["betas"][m], R["cam"][m]], 1) vt = np.concatenate([T["rotmats"][m].reshape(-1, 144), T["betas"][m], T["cam"][m]], 1) p = np.array([pcc(x, y) for x, y in zip(vt, vr)]) ang = geodesic_deg(T["rotmats"][m], R["rotmats"][m]) # (N, 16) dt = np.linalg.norm(T["cam_t_full"][m] - R["cam_t_full"][m], axis=1) * 1e3 dz = np.abs(T["cam_t_full"][m][:, 2] - R["cam_t_full"][m][:, 2]) / np.abs(R["cam_t_full"][m][:, 2]) * 100 ds = np.abs(T["cam"][m][:, 0] - R["cam"][m][:, 0]) / np.abs(R["cam"][m][:, 0]) * 100 res["sets"][name] = { "hands": int(m.sum()), "pcc157": {**stats(p), "below_0.9999": int((p < 0.9999).sum()), "below_0.999": int((p < 0.999).sum())}, "rot_err_deg_all16": stats(ang), "rot_err_deg_global": stats(ang[:, 0]), "rot_err_deg_hand_pose": stats(ang[:, 1:]), "rot_err_deg_per_hand_mean": stats(ang.mean(1)), "betas_abs_err": stats(np.abs(T["betas"][m] - R["betas"][m])), "cam_t_full_err_mm": stats(dt), "cam_t_full_depth_rel_err_pct": stats(dz), "cam_scale_rel_err_pct": stats(ds), "ref_cam_t_full_depth_m": stats(R["cam_t_full"][m][:, 2]), } s = res["sets"]["all"] print(f"[compare] {s['hands']} hands: PCC mean {s['pcc157']['mean']:.6f} min {s['pcc157']['min']:.6f}; " f"rotation error mean {s['rot_err_deg_all16']['mean']:.3f} deg max {s['rot_err_deg_all16']['max']:.3f}; " f"cam_t_full error mean {s['cam_t_full_err_mm']['mean']:.2f} mm max {s['cam_t_full_err_mm']['max']:.2f}") if a.json: Path(a.json).write_text(json.dumps(res, indent=1)) return res def main(): ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) sub = ap.add_subparsers(dest="cmd", required=True) r = sub.add_parser("run") r.add_argument("--impl", choices=["cpu", "tt", "http"], required=True) r.add_argument("--samples", required=True) r.add_argument("--root", required=True) r.add_argument("--out", required=True) r.add_argument("--url", default="http://127.0.0.1:20000") r.add_argument("--limit", type=int, default=0) r.add_argument("--threads", type=int, default=0, help="torch CPU threads (cpu impl)") c = sub.add_parser("compare") c.add_argument("--ref", required=True) c.add_argument("--test", required=True) c.add_argument("--json", default=None) a = ap.parse_args() if a.cmd == "run": run(a) else: compare(a) if __name__ == "__main__": sys.exit(main())