max n_nonzero_coefs

[Nikolai M Chapochnikov]

Hello again,

As far as I understand from the error messages I am getting, n_nonzero_coefs should be smaller or equal than n_kernels. What is the reason for this? For example, if one wants to find similarly looking patters in a dataset.

Thanks!

[Sylvain Chevallier]

Hi, this is mandatory as the sparse decomposition needs to find a decomposition on a subset of the frame elements (here kernels).

If you want to find similar looking patterns in a dataset, you could increase the number of kernels n_kernels (think of kernels as possible patterns stored in a large collection). As this collection is very large, there are multiple solution to reconstruct any input signal by combining these patterns. To make the problem tractable, it is possible to use a sparse regularization. The n_nonzero_coefs restricts the number of kernels that you could use to reconstruct an input signal.

I hope this answer your question.

[Sylvain Chevallier]

I uploaded a new version on pypi (1.0.2), that correct the error you reported in #26

[Nikolai M Chapochnikov]

Thank you for clarifying! I will try that!