kimodo-motion-api / benchmark /evaluate_folder.py
cs686's picture
Deploy Kimodo ZeroGPU motion API
a550c4e verified
Raw
History Blame Contribute Delete
13.6 kB
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
"""
Step (4) of evaluation pipeline.
This script recursively computes metrics for generated and ground-truth motions within a test suite folder tree.
Saves metrics json files per test case and per group of test cases in the folder tree.
"""
import argparse
import json
from itertools import groupby
from pathlib import Path
from typing import Any
import numpy as np
import torch
from tqdm import tqdm
from kimodo.constraints import load_constraints_lst
from kimodo.meta import parse_prompts_from_meta
from kimodo.metrics import (
ContraintFollow,
FootContactConsistency,
FootSkateFromContacts,
FootSkateFromHeight,
FootSkateRatio,
TMR_EmbeddingMetric,
aggregate_metrics,
clear_metrics,
compute_metrics,
compute_tmr_per_sample_retrieval,
)
from kimodo.skeleton import build_skeleton
from kimodo.skeleton.definitions import SOMASkeleton30
from kimodo.tools import load_json, to_torch
DEFAULT_FPS = 30.0
def discover_motion_folders(root: Path) -> list[tuple[Path, Path]]:
root = root.resolve()
if not root.is_dir():
raise FileNotFoundError(f"Folder does not exist: {root}")
out: list[tuple[Path, Path]] = []
for meta_path in root.rglob("meta.json"):
sample_dir = meta_path.parent
if (sample_dir / "motion.npz").is_file() and (sample_dir / "gt_motion.npz").is_file():
rel = sample_dir.relative_to(root)
out.append((sample_dir, rel))
return sorted(out, key=lambda x: str(x[1]))
def group_by_parent(examples: list[tuple[Path, Path]]) -> list[list[tuple[Path, Path]]]:
def parent_key(item: tuple[Path, Path]) -> Path:
return item[1].parent if len(item[1].parts) > 1 else Path(".")
sorted_examples = sorted(examples, key=parent_key)
groups: list[list[tuple[Path, Path]]] = []
for _key, group in groupby(sorted_examples, key=parent_key):
groups.append(list(group))
return groups
def _to_scalar(t: torch.Tensor) -> float:
return float(t.mean().item()) if t.numel() > 0 else float(t.item())
def _to_p95(t: torch.Tensor) -> float:
if t.numel() == 0:
return float("nan")
return float(torch.nanquantile(t, torch.tensor(0.95, device=t.device), dim=0).item())
def _per_sample_metrics_from_saved(metrics_list: list, n: int) -> list[dict[str, float]]:
per_sample: list[dict[str, float]] = [{} for _ in range(n)]
for metric in metrics_list:
for key, lst in metric.saved_metrics.items():
for i, t in enumerate(lst):
if i >= n:
break
per_sample[i][key] = _to_scalar(t)
return per_sample
def _load_pair_embeddings(
sample_dir: Path,
) -> tuple[np.ndarray, np.ndarray, np.ndarray | None] | None:
motion_emb_path = sample_dir / "motion_embedding.npy"
text_emb_path = sample_dir / "text_embedding.npy"
gt_motion_emb_path = sample_dir / "gt_motion_embedding.npy"
if not (motion_emb_path.is_file() and text_emb_path.is_file()):
return None
motion_emb = np.load(motion_emb_path)
text_emb = np.load(text_emb_path)
if motion_emb.ndim == 3 and motion_emb.shape[0] == 1:
motion_emb = motion_emb[0]
if text_emb.ndim == 3 and text_emb.shape[0] == 1:
text_emb = text_emb[0]
gt_motion_emb = None
if gt_motion_emb_path.is_file():
gt_motion_emb = np.load(gt_motion_emb_path)
if gt_motion_emb.ndim == 3 and gt_motion_emb.shape[0] == 1:
gt_motion_emb = gt_motion_emb[0]
return motion_emb, text_emb, gt_motion_emb
def _load_npz_motion(
npz_path: Path,
device: str,
soma30_skel: SOMASkeleton30 | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Load posed_joints and foot_contacts from an NPZ, upscaling SOMA30 to SOMA77 if needed."""
data = np.load(npz_path)
posed_joints = to_torch(data["posed_joints"], device=device)
foot_contacts = to_torch(data["foot_contacts"], device=device)
if posed_joints.shape[-2] == 30 and soma30_skel is not None:
local_rot_mats = to_torch(data["local_rot_mats"], device=device)
root_positions = to_torch(data["root_positions"], device=device)
out77 = soma30_skel.output_to_SOMASkeleton77(
{"local_rot_mats": local_rot_mats, "root_positions": root_positions, "foot_contacts": foot_contacts}
)
posed_joints = out77["posed_joints"]
foot_contacts = out77["foot_contacts"]
return posed_joints, foot_contacts
def _run_eval_on_group(
group: list[tuple[Path, Path]],
skeleton: torch.nn.Module,
metrics_list: list,
device: str,
group_name: str = "",
soma30_skel: SOMASkeleton30 | None = None,
) -> tuple[
list[dict[str, float]],
list[dict[str, float]],
dict[str, float],
dict[str, float],
dict[str, float],
list[dict[str, Any]],
]:
"""Run two passes: gen (motion.npz + embeddings) and GT (gt_motion.npz only). Return
per_sample_gen, per_sample_gt, aggregated_gen, aggregated_gt, tmr_metrics, tmr_per_sample.
"""
n = len(group)
sample_ids: list[str] = []
texts: list[str] = []
motion_embs: list[np.ndarray] = []
text_embs: list[np.ndarray] = []
# ----- Pass 1: generation (motion.npz + all embeddings) -----
clear_metrics(metrics_list)
desc = f"Samples ({group_name})" if group_name else "Samples"
for sample_dir, rel_path in tqdm(group, desc=desc, unit="motion"):
stem = rel_path.name
sample_ids.append(stem)
meta_path = sample_dir / "meta.json"
meta = load_json(meta_path)
texts_parsed, _ = parse_prompts_from_meta(meta)
texts.append(texts_parsed[0] if texts_parsed else "")
posed_joints, foot_contacts = _load_npz_motion(sample_dir / "motion.npz", device, soma30_skel)
nframes = posed_joints.shape[0]
lengths = torch.tensor(nframes, dtype=torch.long, device=device)
constraints_path = sample_dir / "constraints.json"
constraints_lst = (
load_constraints_lst(str(constraints_path), skeleton=skeleton) if constraints_path.is_file() else []
)
metrics_in: dict[str, Any] = {
"posed_joints": posed_joints,
"foot_contacts": foot_contacts,
"lengths": lengths,
"constraints_lst": constraints_lst,
}
text_this = texts_parsed[0] if texts_parsed else ""
embs = _load_pair_embeddings(sample_dir)
if (text_this or "").strip() and embs is not None:
motion_emb, text_emb, gt_motion_emb = embs
metrics_in["motion_emb"] = motion_emb
metrics_in["text_emb"] = text_emb
if gt_motion_emb is not None:
metrics_in["gt_motion_emb"] = gt_motion_emb
motion_embs.append(motion_emb)
text_embs.append(text_emb)
compute_metrics(metrics_list, metrics_in)
per_sample_gen = _per_sample_metrics_from_saved(metrics_list, n)
raw_aggregated_gen = aggregate_metrics(metrics_list)
aggregated_gen = {}
tmr_metrics: dict[str, float] = {}
has_text = len(motion_embs) == n and len(text_embs) == n
for key, v in raw_aggregated_gen.items():
val = _to_scalar(v)
if key.startswith("TMR/"):
if has_text:
tmr_metrics[key] = val
else:
aggregated_gen[key] = val
if "constraint_root2d_err" in raw_aggregated_gen:
aggregated_gen["constraint_root2d_err_p95"] = _to_p95(raw_aggregated_gen["constraint_root2d_err"])
tmr_per_sample: list[dict[str, Any]] = []
if has_text and motion_embs and text_embs and len(motion_embs) == n and len(text_embs) == n:
motion_emb_stack = np.stack(motion_embs, axis=0)
text_emb_stack = np.stack(text_embs, axis=0)
tmr_per_sample = compute_tmr_per_sample_retrieval(motion_emb_stack, text_emb_stack, sample_ids, texts, top_k=5)
# ----- Pass 2: GT (gt_motion.npz only, no embeddings) -----
clear_metrics(metrics_list)
for sample_dir, rel_path in tqdm(group, desc=f"GT ({group_name})" if group_name else "GT", unit="motion"):
posed_joints, foot_contacts = _load_npz_motion(sample_dir / "gt_motion.npz", device, soma30_skel)
nframes = posed_joints.shape[0]
lengths = torch.tensor(nframes, dtype=torch.long, device=device)
constraints_path = sample_dir / "constraints.json"
constraints_lst = (
load_constraints_lst(str(constraints_path), skeleton=skeleton) if constraints_path.is_file() else []
)
metrics_in = {
"posed_joints": posed_joints,
"foot_contacts": foot_contacts,
"lengths": lengths,
"constraints_lst": constraints_lst,
}
compute_metrics(metrics_list, metrics_in)
per_sample_gt = _per_sample_metrics_from_saved(metrics_list, n)
raw_aggregated_gt = aggregate_metrics(metrics_list)
aggregated_gt = {}
for key, v in raw_aggregated_gt.items():
if key.startswith("TMR/"):
continue
aggregated_gt[key] = _to_scalar(v)
if "constraint_root2d_err" in raw_aggregated_gt:
aggregated_gt["constraint_root2d_err_p95"] = _to_p95(raw_aggregated_gt["constraint_root2d_err"])
return (
per_sample_gen,
per_sample_gt,
aggregated_gen,
aggregated_gt,
tmr_metrics,
tmr_per_sample,
)
def _write_json(path: Path, payload: dict[str, Any]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(payload, indent=2) + "\n", encoding="utf-8")
def main():
parser = argparse.ArgumentParser(
description="Recursively evaluate generated motions; write metrics.json per folder and <name>.json per parent.",
)
parser.add_argument(
"folder",
type=Path,
help="Root folder to search recursively for meta.json + motion.npz + gt_motion.npz",
)
parser.add_argument("--device", default=None, help="cuda/cpu. Default: auto")
args = parser.parse_args()
folder = args.folder.resolve()
if not folder.is_dir():
raise SystemExit(f"Folder does not exist: {folder}")
device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")
examples = discover_motion_folders(folder)
if not examples:
raise SystemExit(f"No directories with meta.json, motion.npz, and gt_motion.npz found under {folder}")
print(f"Discovered {len(examples)} motion folders.")
first_posed = np.load(examples[0][0] / "motion.npz")["posed_joints"]
num_joints = first_posed.shape[-2]
# SOMA models could generate 30-joint output; upscale to 77 for evaluation
soma30_skel: SOMASkeleton30 | None = None
if num_joints == 30:
soma30_skel = SOMASkeleton30().to(device)
_ = soma30_skel.somaskel77 # trigger lazy init
soma30_skel.somaskel77.to(device)
skeleton = soma30_skel.somaskel77
print("Detected SOMA30 motions; will upscale to SOMA77 for evaluation.")
else:
skeleton = build_skeleton(num_joints).to(device)
fps = DEFAULT_FPS
kwargs = {"skeleton": skeleton, "fps": fps}
metrics_list = [
FootSkateFromHeight(**kwargs),
FootSkateFromContacts(**kwargs),
FootContactConsistency(**kwargs),
FootSkateRatio(**kwargs),
ContraintFollow(**kwargs),
TMR_EmbeddingMetric(**kwargs),
]
groups = group_by_parent(examples)
for group in tqdm(groups, desc="Evaluating folders"):
sample_dirs = [g[0] for g in group]
folder_for_group = sample_dirs[0].parent
folder_name = folder_for_group.name
(
per_sample_gen,
per_sample_gt,
aggregated_gen,
aggregated_gt,
tmr_metrics,
tmr_per_sample,
) = _run_eval_on_group(group, skeleton, metrics_list, device, group_name=folder_name, soma30_skel=soma30_skel)
texts = []
for sample_dir, _ in group:
meta = load_json(sample_dir / "meta.json")
texts_parsed, _ = parse_prompts_from_meta(meta)
texts.append(texts_parsed[0] if texts_parsed else "")
for i, (sample_dir, _) in enumerate(group):
metrics_path = sample_dir / "metrics.json"
out = {
"num_motions": 1,
"folder": str(sample_dir),
"per_motion_mean_gen": per_sample_gen[i] if i < len(per_sample_gen) else {},
"per_motion_mean_gt": per_sample_gt[i] if i < len(per_sample_gt) else {},
}
if i < len(tmr_per_sample):
out["tmr"] = {
"t2m_rank": tmr_per_sample[i]["rank"],
"text": texts[i] if i < len(texts) else "",
"top5_retrieved": tmr_per_sample[i]["top_k"],
}
_write_json(metrics_path, out)
parent_json_path = folder_for_group.parent / f"{folder_name}.json"
full_metrics = {
"num_motions": len(group),
"folder": str(folder_for_group),
"per_motion_mean_gen": aggregated_gen,
"per_motion_mean_gt": aggregated_gt,
}
if tmr_metrics:
full_metrics["tmr"] = tmr_metrics
_write_json(parent_json_path, full_metrics)
print(f"Wrote metrics.json in each of {len(examples)} folders and folder-level JSONs for {len(groups)} groups.")
if __name__ == "__main__":
main()