| """Getting real data into lattice layout. |
| |
| Scoped by one distinction: **building a lattice from data is lattice |
| construction, which this library already owns. Running a training loop is |
| not.** Everything here is the former. There is no trainer, no optimizer, no |
| normalization policy, and no dataset downloads — and there never will be. |
| |
| There is also no DataLoader. ``torch.utils.data.DataLoader`` is fine; this |
| module supplies the three things it needs:: |
| |
| import torch_dimensions as td |
| from torch.utils.data import DataLoader |
| |
| table = td.data.from_table(coords, times, values, names=("state", "sku")) |
| windows = td.data.LatticeWindow(len(table), input_len=36, horizon=1) |
| train, test = windows.split_at_time(table.times, "2025-01") |
| |
| ds = td.data.LatticeDataset(td.data.TensorSource(table.series, table.lattice), train) |
| dl = DataLoader(ds, batch_size=8, shuffle=True, collate_fn=td.data.collate_lattice) |
| |
| model = td.LSTM(d_model=64, n_layers=6, lattice=table.lattice, d_input=table.n_features) |
| for batch in dl: |
| model(batch.x).pow(2).mean().backward() |
| """ |
|
|
| from torch_dimensions.data.collate import Batch, collate_lattice |
| from torch_dimensions.data.coords import CoordMap, from_coords |
| from torch_dimensions.data.memmap import MemmapSource, Normalizer, masked_stats |
| from torch_dimensions.data.source import LatticeDataset, LatticeSource, Sample, TensorSource |
| from torch_dimensions.data.sparsity import SparsityReport, sparsity |
| from torch_dimensions.data.table import LatticeTable, from_table |
| from torch_dimensions.data.window import LatticeWindow, Window |
|
|
| __all__ = [ |
| "Batch", |
| "SparsityReport", |
| "CoordMap", |
| "LatticeDataset", |
| "LatticeSource", |
| "LatticeTable", |
| "LatticeWindow", |
| "MemmapSource", |
| "Normalizer", |
| "Sample", |
| "TensorSource", |
| "Window", |
| "collate_lattice", |
| "sparsity", |
| "from_coords", |
| "from_table", |
| "masked_stats", |
| ] |
|
|
