Basic Usage

Installation

Install torch_deterministic using pip:

$ pip install torch_deterministic

• Requires python≥3.8

• Semantic versioning.

Example Dataloader

Make a dataset that uses a pseudorandom number generator (PRNG) seeded on the index number to generate augmentations:

import numpy as np

class MyDataset:

    def __getitem__(self, i):
        rng = np.random.default_rng(i)
        return rng, rng.uniform()

Configure a data loader to sample from this dataset:

import torch_deterministic as td
from torch.utils.data import DataLoader

dataset = MyDataset()
sampler = td.InfiniteSampler(
    epoch_size=1000,
    shuffle=True,
    increment_across_epochs=True,
)
dataloader = DataLoader(
    dataset=dataset,
    sampler=sampler,
    collate_fn=td.collate_rngs,
)