debug_compare_unity_pt.py

#!/usr/bin/env python3
"""
Compare Unity GENMO-exported .pt inputs between two sequences.

Goal: find *input* differences (K/bbox/kp2d/cam vel) that could plausibly cause
incam instability/oscillation for a specific clip.

Run:
  /root/miniconda3/envs/gvhmr/bin/python debug_compare_unity_pt.py \\
    --a 101_biboo_birthday_speech_explosion_2 \\
    --b 107_biboo_birthday_speech_explosion_8
"""

from __future__ import annotations

import argparse
from dataclasses import dataclass
from pathlib import Path

import numpy as np
import torch

from genmo.utils.pylogger import Log
from third_party.GVHMR.hmr4d.utils.geo.hmr_cam import (
    compute_bbox_info_bedlam,
    get_bbx_xys,
    get_a_pred_cam,
    safely_render_x3d_K,
)
from third_party.GVHMR.hmr4d.utils.smplx_utils import make_smplx


@dataclass
class SeqStats:
    vid: str
    L: int
    W: int
    H: int
    fx_mean: float
    fx_std: float
    bbx_saved_c_std: tuple[float, float]
    bbx_saved_s_std: float
    bbx_gt_c_std: tuple[float, float]
    bbx_gt_s_std: float
    bbx_delta_c_mean: tuple[float, float]
    bbx_delta_c_std: tuple[float, float]
    bbx_delta_s_mean: float
    bbx_delta_s_std: float
    fcliff_saved_std: tuple[float, float, float]
    fcliff_gt_std: tuple[float, float, float]
    fcliff_delta_std: tuple[float, float, float]
    kp2d_conf_gt05_frac: float
    kp2d_oof_conf_gt05_frac: float
    gt_pred_cam_std: tuple[float, float, float]
    gt_pred_cam_d_std: tuple[float, float, float]
    gt_pred_cam_d_p95: tuple[float, float, float]
    gt_pred_cam_norm_std: float
    gt_pred_cam_d_norm_p95: float


def _as_np(x: torch.Tensor) -> np.ndarray:
    return x.detach().cpu().numpy()


def _infer_wh_from_K(K_fullimg: np.ndarray) -> tuple[int, int]:
    cx = float(np.median(K_fullimg[:, 0, 2]))
    cy = float(np.median(K_fullimg[:, 1, 2]))
    W = int(round(cx * 2.0))
    H = int(round(cy * 2.0))
    return max(W, 1), max(H, 1)


def compute_seq_stats(pt_path: Path, smplx_model) -> SeqStats:
    data = torch.load(pt_path, map_location="cpu", weights_only=False)
    vid = pt_path.stem
    bbx_saved = _as_np(data["bbx_xys"]).astype(np.float64)  # (L,3)
    K = _as_np(data["K_fullimg"]).astype(np.float64)  # (L,3,3)
    kp2d = _as_np(data.get("kp2d", torch.zeros((bbx_saved.shape[0], 17, 3)))).astype(
        np.float64
    )
    transl_c = _as_np(data["smpl_params_c"]["transl"]).astype(np.float64)  # (L,3)

    L = int(bbx_saved.shape[0])
    W, H = _infer_wh_from_K(K)
    fx = K[:, 0, 0]

    # Compute GT-projected bbox from verts (same logic used during training when bbx is missing).
    smpl_params_c = data["smpl_params_c"]
    with torch.no_grad():
        out = smplx_model(
            global_orient=smpl_params_c["global_orient"].float(),
            body_pose=smpl_params_c["body_pose"].float(),
            betas=smpl_params_c["betas"].float(),
            transl=smpl_params_c["transl"].float(),
        )
        verts = out.vertices  # (L, V, 3)

    verts_b = verts[None]  # (1,L,V,3)
    K_b = torch.from_numpy(K).float()[None]  # (1,L,3,3)
    i_x2d = safely_render_x3d_K(verts_b, K_b, thr=0.3)  # (1,L,V,2)
    bbx_gt = get_bbx_xys(i_x2d, do_augment=False)[0].detach().cpu().numpy().astype(np.float64)  # (L,3)

    # bbox stats
    bbx_saved_c = bbx_saved[:, :2]
    bbx_saved_s = bbx_saved[:, 2]
    bbx_gt_c = bbx_gt[:, :2]
    bbx_gt_s = bbx_gt[:, 2]

    bbx_delta_c = bbx_saved_c - bbx_gt_c
    bbx_delta_s = bbx_saved_s - bbx_gt_s

    # f_cliffcam stats (this is what the network sees)
    bbx_saved_t = torch.from_numpy(bbx_saved).float()
    bbx_gt_t = torch.from_numpy(bbx_gt).float()
    K_t = torch.from_numpy(K).float()
    fcliff_saved = compute_bbox_info_bedlam(bbx_saved_t, K_t).numpy().astype(np.float64)
    fcliff_gt = compute_bbox_info_bedlam(bbx_gt_t, K_t).numpy().astype(np.float64)
    fcliff_delta = fcliff_saved - fcliff_gt

    # Conditioning target used by incam translation loss: gt_pred_cam (s,tx,ty)
    # (see `third_party/.../hmr_cam.py:get_a_pred_cam`).
    gt_pred_cam = get_a_pred_cam(
        torch.from_numpy(transl_c).float(),
        bbx_saved_t,
        K_t,
    ).numpy().astype(np.float64)  # (L,3)
    d_gt_pred_cam = np.diff(gt_pred_cam, axis=0)
    gt_pred_cam_std = gt_pred_cam.std(axis=0)
    gt_pred_cam_norm_std = float(np.linalg.norm(gt_pred_cam - gt_pred_cam.mean(axis=0), axis=1).std())
    gt_pred_cam_d_std = d_gt_pred_cam.std(axis=0)
    gt_pred_cam_d_p95 = np.percentile(np.abs(d_gt_pred_cam), 95, axis=0)
    gt_pred_cam_d_norm_p95 = float(np.percentile(np.linalg.norm(d_gt_pred_cam, axis=1), 95))

    # kp2d sanity: how much of provided kp2d is confidently in-frame?
    conf = kp2d[..., 2]
    x = kp2d[..., 0]
    y = kp2d[..., 1]
    conf_gt05 = conf > 0.5
    conf_gt05_frac = float(conf_gt05.mean()) if conf.size else 0.0
    oof = (x < 0.0) | (x > (W - 1.0)) | (y < 0.0) | (y > (H - 1.0))
    oof_conf = oof & conf_gt05
    oof_conf_frac = float(oof_conf.sum() / max(conf_gt05.sum(), 1.0))

    return SeqStats(
        vid=vid,
        L=L,
        W=W,
        H=H,
        fx_mean=float(fx.mean()),
        fx_std=float(fx.std()),
        bbx_saved_c_std=(float(bbx_saved_c[:, 0].std()), float(bbx_saved_c[:, 1].std())),
        bbx_saved_s_std=float(bbx_saved_s.std()),
        bbx_gt_c_std=(float(bbx_gt_c[:, 0].std()), float(bbx_gt_c[:, 1].std())),
        bbx_gt_s_std=float(bbx_gt_s.std()),
        bbx_delta_c_mean=(
            float(bbx_delta_c[:, 0].mean()),
            float(bbx_delta_c[:, 1].mean()),
        ),
        bbx_delta_c_std=(float(bbx_delta_c[:, 0].std()), float(bbx_delta_c[:, 1].std())),
        bbx_delta_s_mean=float(bbx_delta_s.mean()),
        bbx_delta_s_std=float(bbx_delta_s.std()),
        fcliff_saved_std=(
            float(fcliff_saved[:, 0].std()),
            float(fcliff_saved[:, 1].std()),
            float(fcliff_saved[:, 2].std()),
        ),
        fcliff_gt_std=(
            float(fcliff_gt[:, 0].std()),
            float(fcliff_gt[:, 1].std()),
            float(fcliff_gt[:, 2].std()),
        ),
        fcliff_delta_std=(
            float(fcliff_delta[:, 0].std()),
            float(fcliff_delta[:, 1].std()),
            float(fcliff_delta[:, 2].std()),
        ),
        kp2d_conf_gt05_frac=conf_gt05_frac,
        kp2d_oof_conf_gt05_frac=oof_conf_frac,
        gt_pred_cam_std=(float(gt_pred_cam_std[0]), float(gt_pred_cam_std[1]), float(gt_pred_cam_std[2])),
        gt_pred_cam_d_std=(float(gt_pred_cam_d_std[0]), float(gt_pred_cam_d_std[1]), float(gt_pred_cam_d_std[2])),
        gt_pred_cam_d_p95=(float(gt_pred_cam_d_p95[0]), float(gt_pred_cam_d_p95[1]), float(gt_pred_cam_d_p95[2])),
        gt_pred_cam_norm_std=gt_pred_cam_norm_std,
        gt_pred_cam_d_norm_p95=gt_pred_cam_d_norm_p95,
    )


def _print_stats(s: SeqStats) -> None:
    Log.info(f"=== {s.vid} ===")
    Log.info(f"L={s.L}  W×H={s.W}×{s.H}  fx_mean/std={s.fx_mean:.3f}/{s.fx_std:.3f}")
    Log.info(
        "bbx_saved std: center(x/y)=(%.2f,%.2f) size=%.2f"
        % (*s.bbx_saved_c_std, s.bbx_saved_s_std)
    )
    Log.info(
        "bbx_gt    std: center(x/y)=(%.2f,%.2f) size=%.2f"
        % (*s.bbx_gt_c_std, s.bbx_gt_s_std)
    )
    Log.info(
        "bbx(saved-gt) mean: center(x/y)=(%.2f,%.2f) size=%.2f"
        % (*s.bbx_delta_c_mean, s.bbx_delta_s_mean)
    )
    Log.info(
        "bbx(saved-gt) std : center(x/y)=(%.2f,%.2f) size=%.2f"
        % (*s.bbx_delta_c_std, s.bbx_delta_s_std)
    )
    Log.info(
        "f_cliff std saved=(%.4f,%.4f,%.4f) gt=(%.4f,%.4f,%.4f) delta_std=(%.4f,%.4f,%.4f)"
        % (
            *s.fcliff_saved_std,
            *s.fcliff_gt_std,
            *s.fcliff_delta_std,
        )
    )
    Log.info(
        "kp2d conf>0.5 frac=%.3f  oof|conf>0.5 frac=%.3f"
        % (s.kp2d_conf_gt05_frac, s.kp2d_oof_conf_gt05_frac)
    )
    Log.info(
        "gt_pred_cam std(s/tx/ty)=(%.4f,%.4f,%.4f)  norm_std=%.4f"
        % (*s.gt_pred_cam_std, s.gt_pred_cam_norm_std)
    )
    Log.info(
        "d(gt_pred_cam) std(s/tx/ty)=(%.4f,%.4f,%.4f)  p95_abs(s/tx/ty)=(%.4f,%.4f,%.4f)  p95_norm=%.4f"
        % (*s.gt_pred_cam_d_std, *s.gt_pred_cam_d_p95, s.gt_pred_cam_d_norm_p95)
    )


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--root", default="processed_dataset/genmo_features")
    ap.add_argument("--a", required=True)
    ap.add_argument("--b", required=True)
    args = ap.parse_args()

    root = Path(args.root)
    pt_a = root / f"{args.a}.pt"
    pt_b = root / f"{args.b}.pt"
    if not pt_a.exists():
        raise FileNotFoundError(pt_a)
    if not pt_b.exists():
        raise FileNotFoundError(pt_b)

    smplx_model = make_smplx("supermotion").eval()

    s_a = compute_seq_stats(pt_a, smplx_model)
    s_b = compute_seq_stats(pt_b, smplx_model)

    _print_stats(s_a)
    _print_stats(s_b)

    Log.info("=== delta (A - B) ===")
    Log.info(
        "fcliff_delta_std A vs B: (%.4f,%.4f,%.4f) vs (%.4f,%.4f,%.4f)"
        % (*s_a.fcliff_delta_std, *s_b.fcliff_delta_std)
    )
    Log.info(
        "bbx(saved-gt) center std A vs B: (%.2f,%.2f) vs (%.2f,%.2f)"
        % (*s_a.bbx_delta_c_std, *s_b.bbx_delta_c_std)
    )


if __name__ == "__main__":
    main()