Files
619f984c36 Option to provide seed to random generators to ensure reproducibility (#1572)
* Option to provide seed to random generators to ensure reproducibility

I added small function in utilities which imports torch, numpy, python
random and sets seed for all of the libraries to ensure reproducibility
of results.

* Apply recommendations from core contributors on seeding

1. Moved the seeding code to another file
2. Make deterministic as a parameter for trainer class
3. Add assertions for seeding numpy
4. Added warnings
5. torch.manual_seed should be enough for seeding torch

* Revert "Apply recommendations from core contributors on seeding"

This reverts commit a213c8e6882eec8a9e7408b9418926d2db7c5461.

* Revert "Revert "Apply recommendations from core contributors on seeding""

This reverts commit 59b2da53c62878de7aab0aa3feb3115e105eea06.

* Change in test, for correct seeding

* Allow seed equal to 0

* Allow seed to be uint32.max

* Added deterministic to benchmarks

* Cuda manual seed as in benchmark seeding

* Seeding should be done before model initialization

* cuda manual_seed is not necessary

* Fixing seed test_cpu_lbfgs

On some seeds seems like lbfgs doesn't converge.
So I fixed the seed during testing.

* rebasing issue with old reproducibility.py

* Improved documentation and ability to seed before initializing Train
class

* Change in docs

* Removed seed from trainer, update for documentation

* Typo in the docs

* Added seed_everything to _all_

* Fixing old changes

* Model initialization should be earlier then Trainer

* Update pytorch_lightning/trainer/__init__.py

From Example to testcode

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

* Fixing according to the contributors suggestions

* Moving horovod deterministic to Trainer class

* deterministic flag affects horovod docs update

* Improved static typing

* Added deterministic to test runners of horovod

It is failing on some versions, not very predictable

* static seeds for horovod tests

* Change for reset_seed function in tests

* Seeding horovod using reset_seed from tutils

* Update pytorch_lightning/trainer/__init__.py

* chlog

* Update trainer.py

* change "testcode" to "Example" in trainer init documentation

* Update pytorch_lightning/trainer/seed.py, first line in comment

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: Jirka <jirka.borovec@seznam.cz>
Co-authored-by: William Falcon <waf2107@columbia.edu>
2020-05-12 07:53:20 -04:00

43 lines
1.2 KiB
Python

"""Helper functions to help with reproducibility of models. """
import os
from typing import Optional
import numpy as np
import random
import torch
from pytorch_lightning import _logger as log
def seed_everything(seed: Optional[int] = None) -> int:
"""Function that sets seed for pseudo-random number generators in:
pytorch, numpy, python.random and sets PYTHONHASHSEED environment variable.
"""
max_seed_value = np.iinfo(np.uint32).max
min_seed_value = np.iinfo(np.uint32).min
try:
seed = int(seed)
except (TypeError, ValueError):
seed = _select_seed_randomly(min_seed_value, max_seed_value)
if (seed > max_seed_value) or (seed < min_seed_value):
log.warning(
f"{seed} is not in bounds, \
numpy accepts from {min_seed_value} to {max_seed_value}"
)
seed = _select_seed_randomly(min_seed_value, max_seed_value)
os.environ["PYTHONHASHSEED"] = str(seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
return seed
def _select_seed_randomly(min_seed_value: int = 0, max_seed_value: int = 255) -> int:
seed = random.randint(min_seed_value, max_seed_value)
log.warning(f"No correct seed found, seed set to {seed}")
return seed