Skip to main content

​
Overview

The reproducibility module provides utilities for ensuring deterministic execution across different runs by controlling random seeds, threading behavior, and environment variables.

​
Functions

​
set_global_seed

Sets random seeds globally for Python, NumPy, and configures threading for deterministic behavior. 
def set_global_seed(seed: int) -> None
​
seed
int
required
Random seed value to use across all random number generators
Effects:
  • Sets random.seed(seed) for Python’s built-in random module
  • Sets np.random.seed(seed) for NumPy
  • Sets PYTHONHASHSEED environment variable to ensure hash randomization is deterministic
  • Configures threading environment variables (see below)
Threading Configuration: Sets the following environment variables to "1" if not already set:
  • OMP_NUM_THREADS: OpenMP thread count
  • MKL_NUM_THREADS: Intel MKL thread count
  • OPENBLAS_NUM_THREADS: OpenBLAS thread count
This ensures deterministic behavior in numerical libraries by limiting parallel execution. Example: 
from utils.reproducibility import set_global_seed

# Ensure reproducible results
set_global_seed(42)

# Now all random operations will be deterministic
import random
import numpy as np

print(random.random())  # Same value every run
print(np.random.rand()) # Same value every run

​
reproducibility_context

Captures the current reproducibility context including platform information, seed, and environment variables. 
def reproducibility_context(config: object) -> dict
​
config
object
required
Configuration object (typically a dataclass or dict-like) containing a random_seed field
​
return
dict
Dictionary containing reproducibility metadata:python_version (str): Python version string (e.g., “3.10.5”)platform (str): Platform identifier (e.g., “Linux-5.15.0-x86_64”)seed (int): Random seed extracted from config or 0 if not foundthread_env (dict): Current threading environment variable values:
  • OMP_NUM_THREADS: OpenMP thread count setting
  • MKL_NUM_THREADS: Intel MKL thread count setting
  • OPENBLAS_NUM_THREADS: OpenBLAS thread count setting
  • PYTHONHASHSEED: Python hash seed value
Example: 
from dataclasses import dataclass
from utils.reproducibility import set_global_seed, reproducibility_context

@dataclass
class Config:
    random_seed: int = 42
    model_name: str = "detector"

config = Config()
set_global_seed(config.random_seed)

context = reproducibility_context(config)
print(f"Python version: {context['python_version']}")
print(f"Platform: {context['platform']}")
print(f"Seed: {context['seed']}")
print(f"Thread settings: {context['thread_env']}")

# Save context with experiment results
import json
with open("experiment_context.json", "w") as f:
    json.dump(context, f, indent=2)

​
Usage Patterns

​
Standard Setup

from config import CONFIG
from utils.reproducibility import set_global_seed, reproducibility_context

# Set seed at the start of your script
set_global_seed(CONFIG.random_seed)

# Capture context for logging
context = reproducibility_context(CONFIG)

​
With Custom Configuration

from utils.reproducibility import set_global_seed, reproducibility_context

class MyConfig:
    def __init__(self):
        self.random_seed = 123
        self.experiment_name = "test"

config = MyConfig()
set_global_seed(config.random_seed)
context = reproducibility_context(config)

​
Saving Reproducibility Information

import json
from pathlib import Path
from config import CONFIG
from utils.reproducibility import set_global_seed, reproducibility_context

set_global_seed(CONFIG.random_seed)
context = reproducibility_context(CONFIG)

# Save with experimental results
output_path = CONFIG.output_dir / "reproducibility.json"
with open(output_path, "w") as f:
    json.dump(context, f, indent=2)

print(f"Reproducibility context saved to {output_path}")

​
Notes

  • Always call set_global_seed() at the beginning of your script before any random operations
  • The threading configuration limits parallelism to ensure determinism but may impact performance
  • reproducibility_context() works with both dataclasses and dictionary-like configuration objects
  • Environment variables are only set if not already configured, allowing manual overrides

Build docs developers (and LLMs) love

Get started for freeTalk to us