tacco.utils.projection

projection(A, B, metric='bc', deconvolution=False, parallel=True)

Calculate pairwise normalized projections of sparse and dense inputs. Depending on the parameters this can be an asymmetric operation: the argument A here is projected on the argument B.

Parameters:

• A – A 2d ndarray or a scipy sparse matrix.

• B – A 2d ndarray or a scipy sparse matrix with the same second dimension as A; gives the basis to project on.

• metric –

  A string specifying the metric to use. Available are

  • ’cosine’: Euclidean projection of normalized vectors

  • ’naive’: Euclidean projection of vectors normalized as probability distributions

  • ’bc’: Bhattacharyya coefficient, which is the overlap of two probability distributions.

  • ’bc2’: squared Bhattacharyya coefficient.

• deconvolution –

  Which method to use for deconvolution of the projections based on the cross-projections of B. If False, no deconvolution is done. Available methods are:

  • ’nnls’: solves nnls to get only non-negative deconvolved projections

  • ’linear’: solves a linear system to disentangle contributions; can result in negative values which makes sense for general vectors and amplitudes, i.e.

• parallel – Whether to run the operation in parallel.

Returns:

A ndarray containing the distances.