Remove usage of global variables in linear and logistic regression tutorial training functions

[christiangnrd]

Based off Slack "thread".

I think we should bikeshed the following:

1. Do we rename all local function variables to be different from a global variable? If so, what should they be.
2. Add return statements to the functions to make it clear what is meant to be returned. Or just a return to make it clear that no output is expected from the train_<something>!() functions.

Personally, I think we should definitely do 1., and probably do 2.

[mcabbott]

Thanks!

I'm in favour of 1, as I remember this being confusing. What the names should be I don't know, as having too many names for the same thing also gets messy. Probably _W, _b, _x, _y are too far from plain text? Are WW, bb, xx, yy too ugly? W_local, b_local, ... too long although maybe the first time?

Not sure about 2. It's a bit weird to put return in when it's optional. Maybe my vote would be none (as it is now) but perhaps a # this value is returned on the first function whose return value we do want?

[christiangnrd]

What about weights, biases, data, targets? Long but for a tutorial I feel like clarity is more important.

[mcabbott]

I'm not so worried about length, as about obscuring the relation between the locals and the globals.

One idea is to call globals full words like weights, biases, data, targets, and corresponding locals shortened versions like wei, bias, dat, targ? Too ugly? (Some word choices work better for this than others.) Maybe w, b, d, t as locals still preserves the link?

[christiangnrd]

One idea is to call globals full words like weights, biases, data, targets, and corresponding locals shortened versions like wei, bias, dat, targ? Too ugly? (Some word choices work better for this than others.) Maybe w, b, d, t as locals still preserves the link?

I agreed with you until I started making the changes and realized that the globals are named that way to match with the mathematical definitions from the tutorial, so I think they should remain as-is and we change the custom training function arguments to be a slightly more verbose term that aligns with their definition.

Coming from a CS background, seeing that W is passing into a function as the weights argument would help me remember what W represents.

See my latest commit and let me know what you think.

[mcabbott]

Ok, good point about matching the equations. Then this is fine, I think.

[mcabbott]

Documentation CI was failing here for other reasons, but once fixed, this tutorial now produces a failure: