qstack.regression.local_kernels

Local (atomic) kernel implementations.

Provides:

local_kernels_dict: Dictionary mapping kernel names to their implementations. RAM_BATCHING_SIZE: Max. RAM (in bytes) that can be used for batched compuation

of Manhattan distance matrix for L_custom_py kernel. Can be modified before call using local_kernels.RAM_BATCHING_SIZE = ….

qstack.regression.local_kernels.compute_distance_matrix(R1, R2)[source]

Compute the Manhattan-distance matrix.

This computes (||r_1 - r_2||_1) between the samples of R1 and R2, using a batched python/numpy implementation, designed to be more memory-efficient than a single numpy call and faster than a simple python for loop.

This function is a batched-over-R1 implementation of the following code: return np.sum( (R1[:,None, …]-R2[None,:, …])**2, axis=tuple(range(2, R1.ndim)))

Parameters:

  • R1 (numpy ndarray) – First set of samples (can be multi-dimensional).

  • R2 (numpy ndarray) – Second set of samples.

Returns:

squared-distance matrix of shape (len(R1), len(R2)).

Return type:

numpy ndarray

Raises:

RuntimeError – If X and Y have incompatible shapes.

qstack.regression.local_kernels.cosine_similarity_wrapper(x, y, *_kargs, **_kwargs)[source]

Compute cosine similarity kernel.

Parameters:

  • x (numpy ndarray) – First set of samples.

  • y (numpy ndarray) – Second set of samples.

  • *_kargs – Unused positional arguments (for compatibility).

  • **_kwargs – Unused keyword arguments (for compatibility).

Returns:

Cosine similarity matrix.

Return type:

numpy ndarray

qstack.regression.local_kernels.custom_C_kernels(kernel_function, return_distance_function=False)[source]

Create kernel function wrappers using C implementation for speed.

Parameters:

  • kernel_function (str) – Kernel type (‘L’ for Laplacian, ‘G’ for Gaussian).

  • return_distance_function (bool) – If True, returns distance function instead of kernel. Defaults to False.

Returns:

Kernel or distance function, or None if C library cannot be loaded.

Return type:

callable or None

qstack.regression.local_kernels.custom_laplacian_kernel(X, Y, gamma)[source]

Compute Laplacian kernel between X and Y using Python implementation.

K(x, y) = exp(-gamma * ||x - y||_1)

Parameters:

  • X (numpy ndarray) – First set of samples (can be multi-dimensional).

  • Y (numpy ndarray) – Second set of samples.

  • gamma (float) – Kernel width parameter.

Returns:

Laplacian kernel matrix of shape (len(X), len(Y)).

Return type:

numpy ndarray

Raises:

RuntimeError – If X and Y have incompatible shapes.

qstack.regression.local_kernels.dot_kernel_wrapper(x, y, *_kargs, **_kwargs)[source]

Compute linear (dot product) kernel.

Parameters:

  • x (numpy ndarray) – First set of samples.

  • y (numpy ndarray) – Second set of samples.

  • *_kargs – Unused positional arguments (for compatibility).

  • **_kwargs – Unused keyword arguments (for compatibility).

Returns:

Linear kernel matrix.

Return type:

numpy ndarray

qstack.regression.local_kernels.local_laplacian_kernel_wrapper(X, Y, gamma)[source]

Decide which kernel implementation to call.

Wrapper that acts as a generic Laplacian kernel function.

Parameters:

  • X (numpy ndarray) – First set of samples (can be multi-dimensional).

  • Y (numpy ndarray) – Second set of samples.

  • gamma (float) – Kernel width parameter.

Returns:

Laplacian kernel matrix of shape (len(X), len(Y)).

Return type:

numpy ndarray

Raises:

RuntimeError – If X and Y have incompatible shapes.