first class scalar

[finiteprods]

The purpose of this PR is two-fold:

• Adds a new safe API for working with model scalars.
• Fixes the "floating point bug" affecting the test_masking_and_aggregation! tests, discovered in fix extended coverage of masking tests #542.

Currently, scalars are represented as f64 floats. But since scalars are (since #496) an integral part of the (un)masking process, they should be represented similarly to model weights, namely in "exact" form as Ratios of big integers.

Indeed, the bug mentioned above arises from the inexact nature of scalars: in the test_masking_and_aggregation! tests, we take scalars to be a fraction (morally 1 / model_count) which as per the usual limitations of floating point, cannot be represented exactly as f64 in many cases. Subsequently, the error this introduces propagates through the (un)masking process, and leads to failing closeness checks. #542 introduced a workaround by essentially adjusting model_count to a known "good" value.

OTOH by working with f64 scalars, it is easy for clients to construct imprecise (or worse, invalid e.g. negative, NaN) scalars, whether accidentally or maliciously. This could be prevented by insisting that scalars be constructed (or converted) via a safe API that statically guarantees their validity, similar to that of Model.

Summary of changes:

• Introduces an API around a new "first class" Scalar type:

pub struct Scalar(Ratio<BigUint>);

• Adjusts Masker::mask to take a scalar: Scalar (instead of scalar: f64).
• Adjusts downstream clients of masking accordingly, using the Scalar API:
  • e.g. Scalar::new(1, model_count) creates an exact rational scalar 1 / model_count.
  • e.g. Scalar::unit() instead of 1.0.
• Adjusts masking tests, in particular:
  • Fixes the floating point bug in test_masking_and_aggregation! tests by using Scalar, and removes workaround.
  • Adjusts test_xyz_scalar! families of randomized scalar tests (introduced in randomized scalar tests #589).

[codecov]

Codecov Report

Merging #712 (1c9dfb6) into master (dbaa0e5) will increase coverage by 0.47%.
The diff coverage is 72.83%.

@@            Coverage Diff             @@
##           master     #712      +/-   ##
==========================================
+ Coverage   61.51%   61.99%   +0.47%     
==========================================
  Files         102      104       +2     
  Lines        4563     4581      +18     
==========================================
+ Hits         2807     2840      +33     
+ Misses       1756     1741      -15     

Impacted Files	Coverage Δ
rust/xaynet-core/src/mask/model.rs	63.72% <ø> (ø)
rust/xaynet-mobile/src/ffi/mod.rs	0.00% <ø> (ø)
rust/xaynet-mobile/src/ffi/settings.rs	0.00% <0.00%> (ø)
rust/xaynet-mobile/src/settings.rs	0.00% <0.00%> (ø)
rust/xaynet-sdk/src/settings/mod.rs	0.00% <0.00%> (ø)
rust/xaynet-sdk/src/state_machine/phase.rs	16.66% <ø> (ø)
...ust/xaynet-server/src/state_machine/coordinator.rs	96.42% <ø> (ø)
rust/xaynet-server/src/state_machine/mod.rs	75.00% <ø> (ø)
rust/xaynet-server/src/state_machine/phases/mod.rs	73.33% <53.84%> (+3.94%)	⬆️
.../xaynet-server/src/state_machine/phases/handler.rs	71.87% <71.87%> (ø)
... and 8 more

---

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 ae6a723...1c9dfb6. Read the comment docs.

[janpetschexain]

looks good, i'm glad that this fixes the precision issues 👍

i left some comments mostly regarding trait streamlining.

[Robert-Steiner]

A great improvement👍 The code looks much cleaner.
I just added a small note on the Settings.

[Robert-Steiner]

🚀