Sparse v.s. Dense Vector Discussion on Linear Classifier via @mathieuen @ogrisel @atpassos_ml

Olivier Grisel @ogrisel

@atpassos_ml @mathieuen and then contribute sparse autoencoders to scikit-learn :)

Alexandre Passos @atpassos_ml

@ogrisel @mathieuen I'll put it on my queue; the lack of a decent sparse vector (not matrix) type in scipy always bugged me.

Olivier Grisel @ogrisel

@atpassos_ml you can use csr_matrix with shape (1, vector_dim). It's a bit of a cognitive overload but you can get use to it quickly

Olivier Grisel @ogrisel

@atpassos_ml @mathieuen furthermore when you treat one vector at a time in python level code you have a design / perf issue anyway

Alexandre Passos @atpassos_ml

@ogrisel csr_matrix is slow to change sparsity (although that's what I use)

Olivier Grisel @ogrisel

@atpassos_ml yes I agree your approach looks better, once at the cython level it's trivial to init your datastruct from a single csr vector.

Mathieu Blondel @mblondel_ml

@ogrisel @atpassos_ml I've started a github project. Nothing much yet but this is to get started. Feel free to fork! http://bit.ly/deK80D