v0.1.2

🐛 Fixes:

1. SparseMultivariateNormal Initialisation Warning: Addressed a user warning emerging from the SparseMultivariateNormal class initialisation due to arg_constraints not being defined.
2. Enhancements to PairwiseVoxelEncoder:
   • Module Inheritance: Adjusted PairwiseVoxelEncoder to inherit from torch.nn.Module, aligning it more closely with PyTorch's expected behavior for neural network modules.
   • Device Management: Implemented the _apply function in PairwiseVoxelEncoder. This change facilitates the movement of indices created during class initialization to a designated device using methods such as .to(device), .cpu(), or .cuda().
   • Device Property: Introduced the .device attribute for PairwiseVoxelEncoder. This property provides users with insights regarding the device to which the encoder's indices are assigned.

Note: v.0.1.1 was abandoned due to a code typo error (arg_contraints instead of arg_constraints)

v0.1.0

Release Notes:

We are excited to announce the first release of TorchSparseGradUtils, a suite of efficient utilities that extend the functionality of PyTorch sparse tensor operations to support sparse gradient back-propagation from sparse input tensors.

Here are the key features included in this release:

PyTorch Matrix operations with sparse gradients support:

1. Sparse-Dense matrix multiplication with batch support sparse_matmul.
2. Sparse-Dense triangular linear solver with batch support sparse_triangular_solve
3. Sparse-Dense generic linear solver sparse_generic_solve

Sparse Gaussian Distribution:

1. SparseMultivariateNormal Distribution parameterised by either a sparse lower triangular covariance or precision matrix with reparameterised sampling

Sparse Encoder:

1. PairwiseVoxelEncoder to encode sparse matrices with relationships between pairs of voxels in local neighbourshoods of 3D volumetric images.

Sparse utilities:

1. Convert COO indices and values to CSR indices and values with convert_coo_to_csr_indices_values, with batch support.
2. Convert COO sparse tensors to CSR sparse tensors with batched support with convert_coo_to_csr
3. Equivalent of torch.block_diag() for sparse COO and CSR matrices sparse_block_diag and a function to perform the reverse sparse_block_diag_split
4. Equivalent of torch.eye for sparse COO and CSR matrices sparse_eye
5. Equivalent of torch.stack() for sparse CSR tensors stack_csr

Generating Random Sparse Matrices:

1. Equivalent of torch.rand() for sparse COO and CSR matrices rand_sparse
2. rand_sparse_tri used for generating random strictly triangular sparse matrices in either COO or CSR format.

Additional backbone solvers implemented in pure PyTorch:

1. BICGSTAB (ported from pykrylov) bicstab
2. CG (ported from cornellius-gp/linear_operator) linear_cg
3. LSMR (ported from pytorch-minimize) lsmr
4. MINRES (ported from cornellius-gp/linear_operator) minres

CuPy and JAX solvers:

We also provide wrappers around cupy sparse solvers and jax sparse solvers. Allowing linear systems of PyTorch sparse matrices to be solved using a CuPy or JAX back-end:

1. Sparse-Dense linear solver with CuPy back-end sparse_solve_c4t
2. Sparse-Dense linear solver with JAX back-end sparse_solve_j4t

Installation:

This version can be installed using:

pip install torchsparsegradutils==0.1.0

We welcome any feedback, suggestions, and contributions via our issues page.

For more details about this release, you can refer to the Full Changelog.