mini-batch api coverage

[ColdTeapot273K]

Hey, I've made a draft implementations of mini-batch api (like learn_many) for several river abstractions like column selector, transformer union and iterating on one for one hot encoder. Together with already covered abstractions (linear models for regression/classification, standard scaler and pipeline) it should be an essential set of mini-batch tools ready. I can show here/submit a PR later if it's interesting

Mini-batch api has given me logarithmic speedup in training, from 30+ hours to 3 minutes, on some quite large dataset.

[MaxHalford]

Hey there @ColdTeapot273K! That's music to my ears. As you say, there are some essential pieces of logic that indeed should support mini-batches. Feel welcome to open a PR, we can iron out the details together.

@raphaelsty did quite a bit of work in this direction too. At some point it might be worth considering properly benchmarking these mini-batch implementations to compare them with their *_one alternatives. For example with a tool like pytest-benchmark.

[ColdTeapot273K]

@MaxHalford hey Max, a quick question:
From what I see in the source, the gradient updates related to learn_many methods function correctly if some columns are removed/reordered arbitrarily, correct? I know that learn_one is robust to this.

Examples:

Fitting on all these three samples in any order should produce the same coefficients, if my assumption is true:

prefix_h	prefix_a
0	0
0	1
0	0
1	0

prefix_ff	prefix_h	prefix_a
0	0	0
0	0	1
0	0	0
0	1	0

prefix_h	prefix_a	prefix_ff
0	0	0
0	1	0
0	0	0
1	0	0

---

I'm trying to make sure that my implementation of sparse modality for mini-batch one-hot encoder does not violate any guarantees expected by the glm.py

[MaxHalford]

You're correct @ColdTeapot273K, processing these 3 dataframes in any order should produce the same model state. You could even add a unit test to make sure of this.

[ColdTeapot273K]

@MaxHalford noice 👌

Funnily enough, when writing that message i've figured it's trivial to test, so implemented a test and verified

[ColdTeapot273K]

Finally:

#827