Deterministic ML Experiments: Seeding More Than Just NumPy

Deterministic ML Experiments: Seeding More Than Just NumPy

Setting one random seed is not enough for reproducibility. Different libraries and execution backends each have their own randomness controls.

Step 1: seed Python, NumPy, and framework runtime together

import os, random, numpy as np

SEED = 20260311
os.environ["PYTHONHASHSEED"] = str(SEED)
random.seed(SEED)
np.random.seed(SEED)

Step 2: configure deterministic backend options

import torch

torch.manual_seed(SEED)
torch.use_deterministic_algorithms(True)
torch.backends.cudnn.benchmark = False

Step 3: log seed and environment fingerprint with metrics

meta = {
    "seed": SEED,
    "python": platform.python_version(),
    "torch": torch.__version__,
}

Pitfall

Comparing runs with different library versions while assuming seed parity guarantees identical results.

Verification

• Two runs on same environment produce matching metrics within tolerance.
• Experiment metadata includes seed and package versions.
• Nondeterministic kernels are disabled where required.