Expand description
Validated tensor workflows built from native LibTorch operations.
Helpers allocate outputs and preserve native automatic differentiation; none
mutate an input. flatten_features is the explicit exception to independent
storage: its copy argument selects a view or an independent allocation.
COO coordinate construction and quantization are boundaries without gradients. For individual
operations, use the existing Tensor f_* methods directly. See the
complete recipes
for linear algebra, broadcasting, complex FFTs, special functions and CSR.
Structs§
- Least
Squares - Fit diagnostics from
least_squares, matching the nativegelsdoutputs.
Functions§
- flatten_
features - Flatten
[batch, ...features]into[batch, features]with explicit storage. - gather_
rows - Select one class score per row of a
[batch, classes]tensor. - least_
squares - Fit a possibly rank-deficient CPU system using the SVD-based
gelsddriver. - quantize_
per_ tensor - Quantize finite CPU float32 data with explicit per-tensor affine parameters.
- replace_
masked - Replace masked values without changing the original tensor.
- scatter_
add_ rows - Sum
[items, ...]contributions intorowsoutput groups. - select_
rows - Select samples from
[samples, ...]in the order of a one-dimensional ID tensor. - sparse_
coo_ matrix - Construct a coalesced sparse COO matrix from checked row/column coordinates.
- sparse_
matmul - Multiply a two-dimensional sparse COO/CSR matrix by dense feature columns.
- spectral_
filter - Filter real signals along their last axis with real frequency-bin gains.
- standardize
- Standardize float32/float64 values along one axis using population variance.