How to set a random seed in PyTorch

Randomized initialization and sampling can hide whether a code change or a different random draw caused an experiment to move. A shared seed gives repeated runs the same starting random streams, which makes small comparisons and regression checks easier to interpret.

The random-number generators in PyTorch, Python, and NumPy keep independent state. The torch.manual_seed() call covers PyTorch random generation on all devices, while random.seed() and np.random.seed() cover Python's standard generator and NumPy's legacy global generator.

A seed does not guarantee identical results across PyTorch releases, platforms, or nondeterministic kernels. Code that creates np.random.default_rng() instances must seed each instance separately, and an operation that needs an isolated PyTorch stream can receive its own seeded torch.Generator.

1. Create seed_check.py with one seed helper for the random generators used by the program.

   seed_check.py

   import random
    
   import numpy as np
   import torch
    
    
   SEED = 20260717
    
    
   def seed_everything(seed):
       random.seed(seed)
       np.random.seed(seed)
       torch.manual_seed(seed)

2. Append a sampling function below the seed helper to reset the generators before randomized work.

   seed_check.py

   def draw_values():
       seed_everything(SEED)
    
       model = torch.nn.Linear(3, 2)
       return {
           "weights": model.weight.detach().clone(),
           "torch": torch.rand(3),
           "numpy": np.random.random(3),
           "python": [random.random() for _ in range(3)],
       }

   Random choices consumed before the helper call are not rewound by later reseeding.

3. Append equality checks below the sampling function to compare two resets of every random stream.

   seed_check.py

   first = draw_values()
   second = draw_values()
    
   print(f"model weights match: {torch.equal(first['weights'], second['weights'])}")
   print(f"PyTorch samples match: {torch.equal(first['torch'], second['torch'])}")
   print(f"NumPy samples match: {np.array_equal(first['numpy'], second['numpy'])}")
   print(f"Python samples match: {first['python'] == second['python']}")

4. Run the completed seed check script.

   $ python3 seed_check.py
   model weights match: True
   PyTorch samples match: True
   NumPy samples match: True
   Python samples match: True

   A False line means the corresponding generator was not reset before that random operation or another source of randomness still affects the path.