Thomas Moreau

ENH add optinal torch-based solver

1

In our [recent paper](https://openreview.net/forum?id=rI0LYgGeYaw&noteId=3E0u6Xyua1F) with @malezieux, we achieved faster performances with almost no performance drop with a torch-based algorithm. It would be nice to implement such algorithm in `alphcsc`, with...

API add post-processing functions for activations

1

A particularly tricky part to look at the results in neuroscience in `alphcsc` is how to process the resulting activations. In particular, we could add: - a helper to shift...

I don't remember why we did not used numba to accelerate the coordinate descent for loop, which is often critical for algorithms performances but this should be investigated. This issue...

The benchmark script [`benchmarks/bench_method_run.py`](https://github.com/alphacsc/alphacsc/blob/master/benchmarks/bench_methods_run.py) does not run anymore due to some changes in the API. It should be fixed, and the solver `dicodile` should be added too. Maybe this benchmark...

DOC example for population study

1

We should provide an example where we learn atoms from multiple subjects at once. The script should leverage the `OnlineCDL` class and potentially split each subject into sub-windows. A potential...

We should implement a MNE compatible object with similar API to the [mne.preprocessing.ICA](https://mne.tools/stable/generated/mne.preprocessing.ICA.html?highlight=ica#mne.preprocessing.ICA) object. This would help a lot the neuroscience users to try this on their data.

ENH improve integration with encoder API

2

Things that we can do/discuss for `alphacsc/dicodile` integration. Let's discuss those on monday. - [x] Add tests for the `z_encoder` behavior and the different backends. - [x] Implement a first...

When using `window=True`, the selection of the max error patch should be changed to take into account a weighted norm for the error computation, [like it is done in dicodile](https://github.com/tomMoral/dicodile/blob/main/dicodile/utils/dictionary.py#L66)....

Discussing in the CIRM, some people told me that having a summary of results would be good. maybe some ranking with a plot similar to this: 

A complicated thing is often to understand what is the API of one solver and it is not super easy to discover this everywhere. Having a template for the solver/dataset...