scripts/hf_timercd_probe.py

#!/usr/bin/env python3
"""Small HF Jobs probe for the TimeRCD reproduction.

Runs on a GPU flavor to satisfy the challenge's cloud-execution expectation,
but keeps the experiment scoped to public-code generator and RCD mechanism
checks rather than expensive checkpoint pretraining.
"""

from __future__ import annotations

import json
import os
import random
import subprocess
import sys
import time
from pathlib import Path

import numpy as np


def sh(cmd: list[str], cwd: Path | None = None) -> str:
    proc = subprocess.run(cmd, cwd=cwd, text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, check=True)
    return proc.stdout


def best_f1(scores: np.ndarray, labels: np.ndarray) -> dict:
    thresholds = np.unique(np.quantile(scores, np.linspace(0, 1, 151)))
    best = {"f1": -1.0, "threshold": 0.0, "precision": 0.0, "recall": 0.0}
    for th in thresholds:
        pred = scores >= th
        tp = int(np.sum(pred & (labels == 1)))
        fp = int(np.sum(pred & (labels == 0)))
        fn = int(np.sum((~pred) & (labels == 1)))
        precision = tp / (tp + fp) if tp + fp else 0.0
        recall = tp / (tp + fn) if tp + fn else 0.0
        f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0
        if f1 > best["f1"]:
            best = {"f1": float(f1), "threshold": float(th), "precision": float(precision), "recall": float(recall)}
    return best


def make_contextual(n: int = 80, length: int = 512) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    rng = np.random.default_rng(20260720)
    xs, labels, groups = [], [], []
    for i in range(n):
        t = np.arange(length)
        period = rng.uniform(40, 80)
        phase = rng.uniform(0, 2 * np.pi)
        y = np.sin(2 * np.pi * t / period + phase) + 0.2 * np.sin(2 * np.pi * t / (period / 3))
        y += rng.normal(0, 0.08, length)
        lab = np.zeros(length, dtype=int)
        if i >= n // 2:
            start = int(rng.integers(160, 330))
            width = int(rng.integers(40, 75))
            end = min(length, start + width)
            y[start:end] = 0.25 * np.sin(2 * np.pi * np.arange(end - start) / (period * 0.45) + phase)
            y[start:end] += rng.normal(0, 0.06, end - start)
            lab[start:end] = 1
        xs.append(y)
        labels.append(lab)
        groups.append(np.full(length, i))
    return np.concatenate(xs), np.concatenate(labels), np.concatenate(groups)


def rcd_score(x: np.ndarray, group: np.ndarray) -> np.ndarray:
    out = np.zeros_like(x, dtype=float)
    for g in np.unique(group):
        idx = np.flatnonzero(group == g)
        y = x[idx]
        s = np.zeros_like(y)
        for i in range(len(y)):
            q = y[max(0, i - 15): min(len(y), i + 16)]
            left = y[max(0, i - 96): max(0, i - 15)]
            right = y[min(len(y), i + 16): min(len(y), i + 97)]
            ctx = np.concatenate([left, right]) if left.size + right.size else q
            s[i] = abs(q.mean() - ctx.mean()) + abs(q.std() - ctx.std())
        out[idx] = s
    return out


def reconstruction_like(x: np.ndarray, group: np.ndarray) -> np.ndarray:
    out = np.zeros_like(x, dtype=float)
    for g in np.unique(group):
        idx = np.flatnonzero(group == g)
        y = x[idx]
        pred = np.r_[y[0], y[:-1]]
        out[idx] = np.abs(y - pred)
    return out


def main() -> None:
    root = Path.cwd()
    print("nvidia-smi:")
    try:
        print(sh(["nvidia-smi"]))
    except Exception as exc:
        print(f"nvidia-smi unavailable: {exc}")

    print(sh(["git", "clone", "--depth", "1", "https://github.com/thu-sail-lab/TSAD_dataset_gen_public", "gen"]))
    print(sh(["git", "clone", "--depth", "1", "https://github.com/thu-sail-lab/Time-RCD", "time_rcd"]))

    sys.path.insert(0, str(root / "gen/src"))
    cwd = Path.cwd()
    os.chdir(root / "gen")
    try:
        from generate_dataset import generate_dataset

        random.seed(11)
        np.random.seed(11)
        t0 = time.perf_counter()
        uni = generate_dataset(num_samples=24, seq_len=768, anomaly_sample_ratio=1.0, is_multivariate=False, use_attribute_set=True, num_workers=1)
        multi = generate_dataset(num_samples=8, seq_len=512, anomaly_sample_ratio=1.0, is_multivariate=True, num_features=4, use_attribute_set=True, num_workers=1)
        elapsed = time.perf_counter() - t0
    finally:
        os.chdir(cwd)

    points = int(sum(np.asarray(d["time_series"]).size for d in uni + multi))
    x, y, g = make_contextual()
    rcd = best_f1(rcd_score(x, g), y)
    rec = best_f1(reconstruction_like(x, g), y)
    wrapper = (root / "time_rcd/model_wrapper.py").read_text(encoding="utf-8")
    pretrain = (root / "time_rcd/models/time_rcd/TimeRCD_pretrain_multi.py").read_text(encoding="utf-8")
    summary = {
        "job_kind": "HF t4-small reduced public-code probe",
        "official_generator_samples": len(uni) + len(multi),
        "official_generator_points": points,
        "official_generator_sec": elapsed,
        "official_generator_points_per_sec": points / elapsed,
        "rcd_proxy_best_f1": rcd,
        "reconstruction_like_best_f1": rec,
        "released_code_has_dual_heads": "self.reconstruction_head" in pretrain and "self.anomaly_head" in pretrain,
        "released_code_uses_zero_shot_logits": "score_list, logit_list = cls.zero_shot(data)" in wrapper,
        "links": [
            "https://github.com/thu-sail-lab/TSAD_dataset_gen_public",
            "https://github.com/thu-sail-lab/Time-RCD",
            "https://huggingface.co/thu-sail-lab/Time-RCD",
        ],
    }
    print("HF_TIMRCD_PROBE_SUMMARY_START")
    print(json.dumps(summary, indent=2))
    print("HF_TIMRCD_PROBE_SUMMARY_END")


if __name__ == "__main__":
    main()