Luca Palmieri

No need to justify: all of us are doing OSS out of our free time and sometimes life just crashes that free time (or the will to do anything with...

I looked at https://github.com/LukeMathWalker/ndarray/pull/3 and it seems good to me - I'll merge it and in the meanwhile you can polish the edits required to revert the performance regression you...

Re-running benchmarks: **Desktop (AMD):** ```text name master ns/iter branch ns/iter diff ns/iter diff % speedup clip 1,730 1,658 -72 -4.16% x 1.04 contiguous_sum_1e2 16 14 -2 -12.50% x 1.14 contiguous_sum_1e4...

The new formulation for `cap` is much clearer, thanks! I have reran the benchmarks with your latest changes - the inner discontinuous case is the only one that is clearly...

I agree in principle @bplevin36, but I disagree in practice - if nothing else, because I tried to do this already in https://github.com/rust-ml/classical-ml-discussion/pull/2 and I quickly realised that it's difficult...

My current heuristics (which might or might not survive the test of time as we start adding more models to `linfa`): - `fit` should be used as constructor for the...

> > given a trained model, it should be possible to know which hyperparameters were used to train it (how does this work when a model can be partially/incrementally trained...

> But if it's edgy enough, users who need it could just store it manually. I don't think I've seen that feature anywhere in SKLearn. Let's give our minds permission...

Related to @TommyCpp - have you considered splitting out the `sdk` module in its own crate? It would be ideal to be able to isolate the instrumentation side (e.g. most...

If they are all part of the same crate and you break the public API of the `sdk` sub-module you are forced to publish a new major (if you are...