File size: 3,929 Bytes
4ec41f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | #!/usr/bin/env python3
"""Compare explicit StatePlay state traces with the selected dataset states."""
from __future__ import annotations
import argparse
import json
import sys
from collections import defaultdict
from pathlib import Path
import numpy as np
import pandas as pd
import torch
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from stateplay.utils.state import ( # noqa: E402
STATE_COLUMNS,
STATE_NORM_MAX,
denormalize_state,
load_true_state,
)
def summarize(distances: np.ndarray) -> dict[str, object]:
per_state = distances.mean(axis=(0, 1))
distance = float(per_state.mean())
return {
"samples": int(distances.shape[0]),
"latent_frames_per_sample": int(distances.shape[1]),
"distance": distance,
"score": 1.0 - distance,
"per_state_distance": dict(zip(STATE_COLUMNS, map(float, per_state))),
"per_state_score": dict(zip(STATE_COLUMNS, map(lambda x: 1.0 - float(x), per_state))),
}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--selection", type=Path, required=True)
parser.add_argument("--predictions", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
samples = json.loads(args.selection.read_text(encoding="utf-8"))["samples"]
norm_max = STATE_NORM_MAX.numpy()
all_distances = []
by_label = defaultdict(list)
max_true_disagreement = 0.0
for sample in samples:
video_name = Path(sample["output_name"])
state_path = args.predictions / f"{video_name.stem}_state.txt"
trace = pd.read_csv(state_path, sep=r"\s+")
frames = len(trace)
if frames != 26 or not np.array_equal(trace["frame"].to_numpy(), np.arange(frames)):
raise ValueError(f"Unexpected latent-frame trace in {state_path}: {frames} rows")
predicted = trace[[f"{name}_pred" for name in STATE_COLUMNS]].to_numpy(dtype=np.float64)
recorded_true = trace[[f"{name}_true" for name in STATE_COLUMNS]].to_numpy(dtype=np.float64)
normalized_true = load_true_state(
sample["actions"], num_frames=101, latent_frames=frames,
sampling="end", device="cpu", dtype=torch.bfloat16,
)
target = denormalize_state(normalized_true.float()).squeeze(0).numpy()
disagreement = float(np.max(np.abs(recorded_true - target)))
max_true_disagreement = max(max_true_disagreement, disagreement)
if disagreement > 0.002:
raise ValueError(f"Saved true trace differs from source parquet: {state_path} ({disagreement})")
if not np.isfinite(predicted).all() or not np.isfinite(target).all():
raise ValueError(f"Non-finite state value in {state_path}")
distance = np.abs(predicted - target) / norm_max
all_distances.append(distance)
by_label[sample["label"]].append(distance)
distances = np.stack(all_distances)
report = {
"selection": str(args.selection.resolve()),
"predictions": str(args.predictions.resolve()),
"normalization_max": dict(zip(STATE_COLUMNS, map(float, norm_max))),
"definition": "1 - mean(abs(predicted_raw - target_raw) / state_max) over samples, latent frames, and five variables",
"state_sampling": "end",
"max_saved_true_vs_parquet_difference": max_true_disagreement,
"all_frames": summarize(distances),
"excluding_pinned_initial_frame": summarize(distances[:, 1:]),
"by_label": {label: summarize(np.stack(values)) for label, values in sorted(by_label.items())},
}
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|