skimage / sklearn / dask blogpost on neurofinder data

[TomAugspurger]

@jakirkham I think things are (finally) ready on the Dask side to start on this blogpost.

My basic plan was

1. Load neurofinder into a Dask array
2. Use skimage to do the pre-processing (help?)
3. Fit an Incremental(MiniBatchDictionaryLearning) model

While doing 3, only one core on the cluster will be active. We would ideally wrap that Incremental in a GridSearchCV. However, MiniBatchDictionaryLearning doesn't have a score method. Is something like the following reasonable?

from sklearn.decomposition import MiniBatchDictionaryLearning


class ScorableDictionaryLearning(MiniBatchDictionaryLearning):
    
    def predict(self, X):
        code = self.transform(X)
        return code.dot(self.components_)
    
    def score(self, X, y=None):
        if isinstance(X, np.ndarray):
            scoring = np.linalg.norm
        else:
            scoring = da.linalg.norm
        if y is None:
            y = X
            
        return dask.compute(scoring(self.predict(X) - y))[0]