Replace mapslices with selectdim

[glennmoy]

This PR replaces the call to mapslices with a call to selectdim instead.

We change this now because we no longer need to take slices in the parent apply method(s) (#42).
So we can replace mapslices with selectdim instead, which is simpler and much more performant (see below), yet keeps the desired dims convention.

background

This is a follow up to #42, which removed the need for looping over the slices in apply, and #25, which was aimed at solving our dims convention #18.

We originally liked mapslices because we wanted to adopt its dims convention, which was more intuitive to our users.
However, it has some considerable downsides that make it a poor choice in the long-term, namely:

• Greater attention and long-term development going into eachslice
  • add Slices array type for eachslice/eachrow/eachcol JuliaLang/julia#32310 introducing SlicedArrays and abstracting the dims argument.
• Much poorer performance compared to eachslice
  • mapslices performance JuliaLang/julia#26868
  • mapslices errors with type unstable function (problematic when working with Missings) JuliaLang/julia#26232
• Cannot return a view Equivalent of mapslices with views? JuliaLang/julia#29146
• No support in AxisArrays (but given the above there won't be a hurry to Implement mapslices for AxisArrays JuliaArrays/AxisArrays.jl#195)

Its one technical advantage was allowing multiple dims, but this isn't a necessary feature right now.

Comparing to previous benchmarks this significantly reduces allocations (x100) for element-wise transforms (although MeanStdScaling is a bit of an unfair comparison since it has been simplified in the meantime). It might be worth looking to make the rest more efficient also.

# Periodic - this branch
# apply
0.333721 seconds (1.46 M allocations: 70.581 MiB, 7.08% gc time)
0.000225 seconds (11 allocations: 234.734 KiB)
# apply!
0.014992 seconds (31.75 k allocations: 1.924 MiB)
0.000282 seconds (10 allocations: 234.703 KiB)

# Periodic - main
apply
0.780758 seconds (2.84 M allocations: 146.639 MiB, 5.11% gc time)
0.000626 seconds (1.64 k allocations: 380.922 KiB)
apply!
0.027169 seconds (71.10 k allocations: 4.024 MiB)
0.000667 seconds (1.64 k allocations: 380.922 KiB)

# Power - this branch
# apply
0.080944 seconds (282.02 k allocations: 14.327 MiB)
0.000037 seconds (10 allocations: 78.516 KiB)
# apply!
0.031575 seconds (50.51 k allocations: 2.755 MiB, 39.06% gc time)
0.000034 seconds (9 allocations: 78.500 KiB)

# Power - main
# apply
0.160279 seconds (347.52 k allocations: 18.862 MiB)
0.000445 seconds (1.74 k allocations: 221.500 KiB)
# apply!
0.007062 seconds (5.40 k allocations: 417.305 KiB)
0.000387 seconds (1.74 k allocations: 221.500 KiB)

# MeanStdScaling - this branch
# apply
0.168360 seconds (606.91 k allocations: 31.994 MiB)
0.000052 seconds (13 allocations: 78.734 KiB)
# apply!
0.005720 seconds (2.72 k allocations: 229.367 KiB)
0.000019 seconds (13 allocations: 78.734 KiB)

# MeanStdScaling - main
# apply
0.184251 seconds (419.53 k allocations: 22.622 MiB, 5.67% gc time)
0.000619 seconds (2.04 k allocations: 234.000 KiB)
# apply!
0.007365 seconds (5.70 k allocations: 430.008 KiB)
0.000334 seconds (2.04 k allocations: 234.000 KiB)

# LinearCombination - this branch
# apply
0.342954 seconds (1.24 M allocations: 63.726 MiB, 6.75% gc time)
0.001046 seconds (30.61 k allocations: 812.000 KiB)

# LinearCombination - main
# apply
0.329633 seconds (1.26 M allocations: 64.869 MiB, 3.28% gc time)
0.000814 seconds (30.61 k allocations: 812.000 KiB)

# OneHotEncoding - this branch
# apply
0.057715 seconds (240.52 k allocations: 11.567 MiB)
0.000384 seconds (10.01 k allocations: 254.109 KiB)

# OneHotEncoding - main
# apply
0.021401 seconds (31.42 k allocations: 1.939 MiB)
0.000913 seconds (11.98 k allocations: 968.531 KiB)
"""

[codecov]

Codecov Report

Merging #45 (d27a016) into main (f6bb5e4) will decrease coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main      #45      +/-   ##
==========================================
- Coverage   94.84%   94.79%   -0.06%     
==========================================
  Files           9        9              
  Lines          97       96       -1     
==========================================
- Hits           92       91       -1     
  Misses          5        5              

Impacted Files	Coverage Δ
src/transformers.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f6bb5e4...99efde3. Read the comment docs.

[glennmoy]

dims no longer takes slices in apply so it makes no difference to how OneHotEncoding works unless used with inds

[bencottier]

Looks good to me. Are we leaving LinearCombination for a follow-up?

[glennmoy]

Looks good to me. Are we leaving LinearCombination for a follow-up?

yep, I'm working on it now