torchsig.utils.data_loading.WorkerSeedingDataLoader

class torchsig.utils.data_loading.WorkerSeedingDataLoader(dataset, seed=None, **kwargs)

Bases: DataLoader, Seedable

DataLoader that seeds each worker process differently using a shared seed.

This loader prohibits external worker_init_fn definitions and sets its own init function to ensure reproducible randomness in multi-worker pipelines.

Methods

add_parent	Add parent Seedable object and set up RNGs accordingly.
check_worker_number_rationality
get_distribution	Create distribution function with proper seeding.
get_second_seed	Gets second seed, usually used to seed both torch and numpy generators with slightly different seeds.
init_worker_seed	Set a unique random seed for each worker process.
seed	Set the seed value for both the loader and its dataset.
setup_rngs	Initialize torch and numpy number generators, and update its children.
update_from_parent	Update numpy and torch number generators with parent seed.

Attributes

multiprocessing_context
dataset
batch_size
num_workers
pin_memory
drop_last
timeout
sampler
pin_memory_device
prefetch_factor

__init__(dataset, seed=None, **kwargs)

Initialize DataLoader and Seedable, then assign custom worker init.

Parameters:

• dataset – The dataset to load.

• seed – Optional seed value. If None, a random seed is generated.

• **kwargs – Passed to both DataLoader and Seedable initializers.

Raises:

ValueError – if worker_init_fn is provided in kwargs.

seed(seed_val)

Set the seed value for both the loader and its dataset.

Parameters:

seed_val – The seed value to set.

init_worker_seed(worker_id)

Set a unique random seed for each worker process.

Uses the shared random_generator from the Seedable mixin to derive a new seed per worker_id.

Parameters:

worker_id – The integer ID of the worker process.

__repr__() → str

Printable representation with seed and parent.

Returns:

String representation of the object.

add_parent(parent: Seedable, register: bool = True) → None

Add parent Seedable object and set up RNGs accordingly.

Parameters:

• parent – Parent Seedable object to add.

• register – If True (default), add self to parent.children so that future seed propagation reaches this object. Pass False for transient objects (e.g. per-sample Signal instances) that only need the parent link for metadata/RNG access during their lifetime but must not accumulate in the parent’s child list, which would otherwise cause unbounded memory growth.

get_distribution(params: list | tuple | float, scaling: str = 'linear') → Distribution

Create distribution function with proper seeding.

Parameters:

• params – Parameters for distribution.

• scaling – Scaling param for distribution. Defaults to ‘linear’.

Returns:

Distribution function, seeded.

Return type:

Distribution

get_second_seed(seed: int) → int

Gets second seed, usually used to seed both torch and numpy generators with slightly different seeds.

Parameters:

seed – Seed to use.

Returns:

New seed.

setup_rngs() → None

Initialize torch and numpy number generators, and update its children.

update_from_parent() → None

Update numpy and torch number generators with parent seed.