scripts/compare_debug_with_dataset.py

#!/usr/bin/env python3
from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path

import torch

REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))
GVHMR_ROOT = REPO_ROOT / "third_party" / "GVHMR"
if GVHMR_ROOT.is_dir() and str(GVHMR_ROOT) not in sys.path:
    sys.path.insert(0, str(GVHMR_ROOT))

from genmo.utils.eval_utils import compute_camcoord_metrics
from genmo.utils.geo_transform import compute_cam_angvel, compute_cam_tvel, normalize_T_w2c
from genmo.utils.rotation_conversions import axis_angle_to_matrix, matrix_to_axis_angle
from third_party.GVHMR.hmr4d.utils.smplx_utils import make_smplx


def _to_tensor(x):
    if isinstance(x, torch.Tensor):
        return x
    return torch.as_tensor(x)


def _slice(x: torch.Tensor, n: int) -> torch.Tensor:
    return x[:n].clone()


def _compare_arrays(a: torch.Tensor, b: torch.Tensor):
    if a.shape != b.shape:
        return {"shape_a": list(a.shape), "shape_b": list(b.shape)}
    diff = a - b
    return {
        "mae": float(diff.abs().mean().item()),
        "rmse": float((diff.pow(2).mean().sqrt()).item()),
    }


def _load_smplx_tools(device: torch.device):
    smplx = make_smplx("supermotion").to(device).eval()
    smplx2smpl_path = (
        REPO_ROOT / "third_party" / "GVHMR" / "inputs" / "checkpoints" / "body_models" / "smplx2smpl_sparse.pt"
    )
    j_reg_path = (
        REPO_ROOT / "third_party" / "GVHMR" / "inputs" / "checkpoints" / "body_models" / "smpl_neutral_J_regressor.pt"
    )
    smplx2smpl = torch.load(smplx2smpl_path, map_location=device)
    j_reg = torch.load(j_reg_path, map_location=device)
    return smplx, smplx2smpl, j_reg


def _smplx_to_j3d(params: dict, smplx, smplx2smpl, j_reg):
    out = smplx(**params)
    verts = out.vertices if hasattr(out, "vertices") else out[0].vertices
    verts = torch.stack([torch.matmul(smplx2smpl, v) for v in verts])
    j3d = torch.einsum("jv,fvi->fji", j_reg, verts)
    return verts, j3d


def _prepare_params(params: dict, n: int, device: torch.device):
    out = {}
    for k, v in params.items():
        if not isinstance(v, torch.Tensor):
            v = torch.as_tensor(v)
        out[k] = _slice(v, n).to(device)
    return out


def _min_len(*lens: int) -> int:
    lens = [int(x) for x in lens if x is not None]
    return min(lens) if lens else 0


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--debug-io", required=True, help="Path to debug_io.pt")
    parser.add_argument("--dataset-pt", required=True, help="Path to dataset .pt file")
    parser.add_argument("--num-frames", type=int, default=60)
    parser.add_argument("--out", default="outputs/debug_compare.json")
    args = parser.parse_args()

    debug = torch.load(args.debug_io, map_location="cpu", weights_only=False)
    data = torch.load(args.dataset_pt, map_location="cpu", weights_only=False)

    inputs = debug.get("inputs", {})
    outputs = debug.get("outputs", {})

    n = int(args.num_frames)
    results = {"num_frames": n, "inputs": {}, "metrics": {}, "formatted_dataset": True}

    dataset_inputs = dict(data)
    if "T_w2c" in data:
        T_w2c = _to_tensor(data["T_w2c"]).float()
        if T_w2c.ndim == 3:
            normed_T_w2c = normalize_T_w2c(T_w2c)
            dataset_inputs["cam_angvel"] = compute_cam_angvel(normed_T_w2c[:, :3, :3])
            dataset_inputs["cam_tvel"] = compute_cam_tvel(normed_T_w2c[:, :3, 3])

    # Compare input tensors (if present in both).
    input_keys = [
        "bbx_xys",
        "kp2d",
        "K_fullimg",
        "cam_angvel",
        "cam_tvel",
        "T_w2c",
    ]
    for k in input_keys:
        if k in inputs and k in dataset_inputs:
            a_t = _to_tensor(inputs[k])
            b_t = _to_tensor(dataset_inputs[k])
            n_in = _min_len(n, a_t.shape[0], b_t.shape[0])
            a = _slice(a_t, n_in)
            b = _slice(b_t, n_in)
            results["inputs"][k] = _compare_arrays(a, b)

    # Compare incam SMPLX (pred vs GT from dataset).
    pred_key = "smpl_params_incam"
    if pred_key not in outputs:
        pred_key = "pred_smpl_params_incam"

    if pred_key in outputs and "smpl_params_c" in data:
        device = torch.device("cpu")
        smplx, smplx2smpl, j_reg = _load_smplx_tools(device)
        pred_raw = outputs[pred_key]
        gt_raw = data["smpl_params_c"]
        pred_len = int(_to_tensor(pred_raw["global_orient"]).shape[0])
        gt_len = int(_to_tensor(gt_raw["global_orient"]).shape[0])
        n_eval = _min_len(n, pred_len, gt_len)
        pred_params = _prepare_params(pred_raw, n_eval, device)
        gt_params = _prepare_params(gt_raw, n_eval, device)

        pred_verts, pred_j3d = _smplx_to_j3d(pred_params, smplx, smplx2smpl, j_reg)
        gt_verts, gt_j3d = _smplx_to_j3d(gt_params, smplx, smplx2smpl, j_reg)

        metrics = compute_camcoord_metrics(
            {
                "pred_j3d": pred_j3d,
                "target_j3d": gt_j3d,
                "pred_verts": pred_verts,
                "target_verts": gt_verts,
            }
        )
        results["metrics"] = {k: float(v.mean()) for k, v in metrics.items()}

    out_path = Path(args.out)
    out_path.parent.mkdir(parents=True, exist_ok=True)
    with out_path.open("w") as f:
        json.dump(results, f, indent=2)

    print(f"Wrote {out_path}")


if __name__ == "__main__":
    main()