RANSAC estimator

[santiago-imelio]

Hello! This is work in progress but I would love to get some feedback on how this looks so far.

Specifically on the while code, not sure if I'm passing things correctly (loss_fn and min_samples are being passed as tensor shapes because otherwise they turn into Nx.Defn.Expr 0_o).

Ideally, RANSAC could take any fit-predict model as base estimator, so I'd love to hear suggestions on how to go about implementing that (could be something similar to the loss_fn pattern, but not sure).

[josevalim]

Remember to namespace the module :)

[josevalim]

You want to avoid Nx.to_number because it means fit then cannot be called from another defn. The best way is to write some tests calling this function directly and also from another defn.

[santiago-imelio]

@josevalim Hey! thanks for the suggestions. I've been looking for a way to use n_inliers without converting to number but failed to do so. Since it is a Nx.Defn.Expr, I cannot use it for indexing. Any suggestion on how to approach this? The objective is to extract from x and y the inliers, meaning the indices from inliers_mask such that the value is 1.

[josevalim]

Specifically on the while code, not sure if I'm passing things correctly (loss_fn and min_samples are being passed as tensor shapes because otherwise they turn into Nx.Defn.Expr 0_o).

Everything that is not an option is assumed to be an expression and converted to a tensor or nx.defn.expr. But that's generally good. When it is an expression, it can be cached and optimised.