SniperGold_ML/ml/calib_time.py

# -*- coding: utf-8 -*-
"""Kalibrasi waktu training LSTM 2-lapis pd subset — ekstrapolasi utk gate/wf."""
import os
import sys
import time
import numpy as np

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)
import train_hybrid as TH


def time_run(n_train, n_ep, tag):
    F, label, close, midx, split_pos, split_bar = TH.load_base()
    X = F
    mean = X[midx[:split_pos]].mean(0)
    std = X[midx[:split_pos]].std(0)
    std[std < 1e-9] = 1.0
    Xs = (X - mean) / std
    n = min(n_train, split_pos)
    va2 = int(0.85 * n)
    Xtr = TH.build_sequences(Xs, midx[:va2])
    Xva = TH.build_sequences(Xs, midx[va2:n])
    yl = (label[midx] == 1).astype(float)
    ys = (label[midx] == -1).astype(float)
    Ytr = np.column_stack([yl[:va2], ys[:va2]])
    Yva = np.column_stack([yl[va2:n], ys[va2:n]])
    t0 = time.time()
    st, va, ep = TH.train_lstm2(Xtr, Ytr, Xva, Yva, seed=42, epoch_max=n_ep,
                                verbose=False)
    dt_ = time.time() - t0
    per_ep = dt_ / (ep + 1)
    print(f"[{tag}] n_train={len(Xtr)} epochs={ep + 1} total={dt_:.0f}s "
          f"({per_ep:.1f}s/ep) val_AUC={va:.4f}")
    return per_ep


if __name__ == "__main__":
    per = time_run(6000, 3, "probe-6k")
    # ekstrapolasi gate: ~39k train, 5 seed, ~15 ep/seed
    est_gate = per * (39000 / 6000) * 15 * 5
    print(f"Estimasi GATE penuh (39k train, 5 seed, ~15 ep): ~{est_gate / 60:.0f} menit")