Reproducible ML

Overview

Reproducibility means: same code + same data + same config → same result. It is a prerequisite for trusting comparisons, debugging regressions, and shipping. Three pillars: seed everything, pin everything, version the data.

When to use

• Results vary run-to-run.
• Setting up a new project or research codebase.
• Preparing work others must reproduce (papers, audits, reviews).

Pillar 1 — Seed everything

import os, random, numpy as np, torch

def seed_everything(seed: int = 42):
    os.environ["PYTHONHASHSEED"] = str(seed)
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False  # disables nondeterministic autotuner

For DataLoaders, also set worker_init_fn and a generator so workers are deterministic.

Pillar 2 — Pin the environment

• Pin exact versions: requirements.txt with ==, or uv.lock / poetry.lock / conda env export.
• Record Python + CUDA + cuDNN versions in the run metadata.
• Containerize (Docker) for cross-machine reproducibility.

Pillar 3 — Version the data

• Hash datasets (sha256) and log the hash with every run.
• Use DVC or dataset snapshots; never overwrite data.csv in place.
• Treat data as immutable inputs keyed by version (data_v="2026-06-01").

Project layout that supports reproducibility

project/
├── data/            # raw (immutable) + processed, both versioned
├── src/             # importable code, no notebooks doing real work
├── configs/         # YAML/Hydra configs, one per experiment
├── scripts/         # entrypoints: train.py, evaluate.py
├── requirements.txt # or uv.lock / poetry.lock (pinned)
└── README.md        # exact commands to reproduce

Determinism gotchas

• GPU reductions can be nondeterministic even with seeds — use torch.use_deterministic_algorithms(True) and set CUBLAS_WORKSPACE_CONFIG=:4096:8.
• Parallel groupby/apply ordering can vary — sort before reducing.
• set/dict ordering across processes — set PYTHONHASHSEED.
• Non-pinned dependencies silently change behavior between installs.

Pitfalls

• Seeding only NumPy but using PyTorch/CUDA RNGs.
• "Works on my machine" — unpinned env + unversioned data.
• Doing real computation in notebooks with hidden execution order — move logic to src/.

Hand-off

A repo where pip install -r requirements.txt && python scripts/train.py --config configs/exp.yaml recreates the exact result that experiment-tracking logged.