API Reference: data¶

torchlight.data provides datasets, batches, and synthetic problems.

from torchlight.data import Dataset, TensorDataset, DataLoader, make_synthetic

Dataset¶

class Dataset()

Abstract base class: implement __len__ and __getitem__(idx).

TensorDataset¶

class TensorDataset(*tensors)

A fixed-size dataset of parallel array-likes (tensors, numpy arrays, or lists), converted lazily to Tensors per element.

Method	Description
`len(ds)`	number of elements
`ds[i]`	`tuple[Tensor, ...]` — one tensor per field
`to_batch()`	whole dataset as `tuple[Tensor, ...]` shaped `(N, ...)`

ds = TensorDataset(tl.tensor([1., 2.]), tl.tensor([0, 1]))
ds[0]        # (tensor([1.]), tensor([0.]))
ds.to_batch()  # (tensor([1., 2.]), tensor([0., 1.]))

DataLoader¶

class DataLoader(dataset, batch_size=1, shuffle=False, drop_last=False, seed=None)

Iterate over a dataset in batches.

for batch_x, batch_y in DataLoader(ds, batch_size=16, shuffle=True):
    ...

Argument	Meaning
`dataset`	any `Dataset` (e.g. `TensorDataset`)
`batch_size`	elements per batch
`shuffle`	randomize order each epoch
`drop_last`	drop the last incomplete batch
`seed`	RNG seed for shuffling

Samplers¶

from torchlight.data import SequentialSampler, RandomSampler, BatchSampler

Class	Description
`SequentialSampler(n)`	indices `0..n-1` in order
`RandomSampler(n, seed=None)`	indices in random order
`BatchSampler(sampler, batch_size, drop_last=False)`	group indices into batches

Synthetic problems¶

make_synthetic¶

def make_synthetic(name: str, n=100, seed=1) -> TensorDataset

2-D classification problems designed for quick experiments:

Name	Shape
`"simple"`	separable gaussian blobs
`"diag"`	diagonal stripes
`"split"`	split advice region
`"xor"`	XOR corners
`"circle"`	concentric circles
`"spiral"`	two interleaved arms

ds = make_synthetic("spiral", n=300, seed=7)
x, y = ds.to_batch()

synthetic_classification / make_spiral¶

def synthetic_classification(name, n=100, seed=1) -> Graph
def make_spiral(N) -> List[Tuple[float, float]]

Lower-level entry points returning the raw coordinate/label Graph and the spiral point list respectively.