torch_staintools.functional.concentration package

Submodules

torch_staintools.functional.concentration.implementation module

class torch_staintools.functional.concentration.implementation.ConcentCfg(algorithm: Literal['ista', 'cd', 'fista', 'ls'] = 'fista', regularizer: float = 0.01, rng: torch._C.Generator | None = None, maxiter: int = 50, lr: float | None = None, tol: float = 1e-05, positive: bool = True)

Bases: object

algorithm: Literal['ista', 'cd', 'fista', 'ls'] = 'fista'
lr: float | None = None
maxiter: int = 50
positive: bool = True
regularizer: float = 0.01
rng: Generator | None = None
tol: float = 1e-05
class torch_staintools.functional.concentration.implementation.ConcentrationSolver(cfg: ConcentCfg)

Bases: Callable

cfg: ConcentCfg
torch_staintools.functional.concentration.implementation.get_concentration_batch(od_flatten: Tensor, stain_matrix: Tensor, regularizer: float, algorithm: Literal['ista', 'cd', 'fista', 'ls'], lr: float | None, maxiter: int, tol: float, rng: Generator | None, positive: bool)
torch_staintools.functional.concentration.implementation.get_concentration_one_by_one(od_flatten: Tensor, stain_matrix: Tensor, regularizer: float, algorithm: Literal['ista', 'cd', 'fista', 'ls'], lr: float | None, maxiter: int, tol: float, rng: Generator | None, positive: bool)
torch_staintools.functional.concentration.implementation.get_concentrations(image: Tensor, stain_matrix: Tensor, regularizer: float, algorithm: Literal['ista', 'cd', 'fista', 'ls'], lr: float | None, maxiter: int, tol: float, rng: Generator | None, positive: bool)

Estimate concentration matrix given an image and stain matrix.

Warning

algorithm = ‘ls’ May fail on GPU for individual large input (e.g., 1000 x 1000), regardless of batch size. Better for multiple small inputs in terms of H and W.

Parameters:

  • image – batched image(s) in shape of BxCxHxW

  • stain_matrix – B x num_stain x input channel

  • regularizer – regularization term if ISTA algorithm is used

  • algorithm – which method to compute the concentration: Solve min||HExC - OD||p support ‘ista’, ‘cd’, and ‘ls’. ‘ls’ simply solves the least square problem for factorization of min||HExC - OD||F (Frobenius norm) but is faster. ‘ista’/cd enforce the sparse penalty (L1 norm) but slower.

  • lr – learning rate for FISTA and ISTA. If None then computed from 1 / Lipschitz constant.

  • maxiter – max iteration for sparse code optimization.

  • tol – tolerance for early convergence for sparse code optimization.

  • rng – torch.Generator for random initializations

  • positive – whether to enforce positive concentration. Default: True

Returns:

B x num_stains x num_pixel_in_tissue_mask

Return type:

concentration matrix

torch_staintools.functional.concentration.implementation.get_concentrations_single(od_flatten: Tensor, stain_matrix: Tensor, regularizer: float, algorithm: Literal['ista', 'cd', 'fista', 'ls'], lr: float | None, maxiter: int, tol: float, rng: Generator | None, positive: bool)

Helper function to estimate concentration matrix given an image and stain matrix with shape: 2 x (H*W)

For solvers without batch support. Inputs are individual data points from a batch

Parameters:

  • od_flatten – Flattened optical density vectors in shape of (H*W) x C (H and W dimensions flattened).

  • stain_matrix – the computed stain matrices in shape of num_stain x input channel

  • regularizer – regularization term if ISTA algorithm is used

  • algorithm – which method to compute the concentration: coordinate descent (‘cd’) or iterative-shrinkage soft thresholding algorithm (‘ista’)

  • lr – learning rate for ISTA and FISTA. No effect for other methods. If None, will be computed by 1 / Lipschitz constant.

  • maxiter – maximum number of iterations for CD, ISTA, and FISTA

  • tol – tolerance of convergence

  • rng – torch.Generator for random initializations

  • positive – enforce positive concentration

Returns:

num_stains x num_pixel_in_tissue_mask

Return type:

computed concentration

Module contents