Remove the recurrent and LSTM architectures

[michakraus]

Removes RecurrentNeuralNetwork and LSTMNeuralNetwork, along with src/architectures/recurrent_neural_network.jl, src/architectures/LSTM_neural_network.jl and the driver scripts under scripts/Script_using_fully_GML/{RNN,LSTM}/.

They were already unusable

Chain(::RecurrentNeuralNetwork) returned an AbstractNeuralNetworks.GridCell rather than a Chain, and the cells it held still defined the pre-0.6 initialparameters(cell, backend, T; init, rng) signature — so constructing a NeuralNetwork from either architecture errored out. Nothing in the test suite referenced either one, which is why nothing noticed.

Why now

AbstractNeuralNetworks 0.7 deleted src/cells/ outright — Recurrent, LSTM, GRU, IdentityCell and GridCell (AbstractNeuralNetworks.jl#34) — so these imports at the top of GeometricMachineLearning.jl have to go regardless of what happens to the architectures that used them:

import AbstractNeuralNetworks: Architecture, Model, AbstractExplicitLayer,
                               AbstractExplicitCell, ...
import AbstractNeuralNetworks: Chain, GridCell
import AbstractNeuralNetworks: Dense, Linear, Recurrent

Doing it here, against main and the current AbstractNeuralNetworks 0.6.4, keeps the removal independent of the 0.7 bump — #246 then only has to be a dependency change, not a dependency change carrying a behavioural one.

Concretely: #246 as it stands still imports all three removed names, and loading it against 0.7.0 emits

WARNING: Imported binding AbstractNeuralNetworks.AbstractExplicitCell was undeclared at import time
WARNING: Imported binding AbstractNeuralNetworks.GridCell was undeclared at import time
WARNING: Imported binding AbstractNeuralNetworks.Recurrent was undeclared at import time

Julia downgrades those to warnings so the module still loads, but they should not survive the merge. Once this lands on main, #246 picks it up on its next merge from main.

Docs

The footnote in docs/src/architectures/neural_network_integrators.md had flagged the LSTM implementation as likely to be deprecated since it was written; it now says it is gone. The transformer-derived architectures are what GML maintains for time series.

Verification

• Pkg.test() on this branch against AbstractNeuralNetworks 0.6.4 — 56 testsets, zero failures, zero errors.
• Loads with no undeclared-binding warnings, and RecurrentNeuralNetwork/LSTMNeuralNetwork are no longer defined.
• grep -rniE "RecurrentNeuralNetwork|LSTMNeuralNetwork|GridCell|AbstractExplicitCell|IdentityCell" over *.jl/*.md/*.toml returns nothing outside the CHANGELOG entry documenting the removal.

Separately confirmed while preparing this: with this removal applied on top of #246's branch, GML's full suite also passes against the released AbstractNeuralNetworks 0.7.0 (with a dev'd SymbolicNeuralNetworks, which needed its own QPTOAT → ArrayOrNamedTuple fix), zero failures.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 65.19%. Comparing base (33d76f2) to head (df89a6e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #247      +/-   ##
==========================================
- Coverage   65.59%   65.19%   -0.40%     
==========================================
  Files         101       99       -2     
  Lines        2985     2931      -54     
==========================================
- Hits         1958     1911      -47     
+ Misses       1027     1020       -7     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[michakraus]

Reviewed at 30f47127. The removal is correct and complete — I could not find a way for it to break anything, and I checked the premise rather than taking it on trust. Two minor cleanups, both now pushed as df89a6e6.

What I verified

The two architectures really were unconstructible. Against main + AbstractNeuralNetworks 0.6.4:

julia> Chain(RecurrentNeuralNetwork(4, 4, (1,1)))
GridCell{1, 1, Matrix{Any}, Vector{Vector{Float64}}}

julia> NeuralNetwork(RecurrentNeuralNetwork(4, 4, (1,1)), CPU(), Float32)
ERROR: initialparameters not implemented for model type GridCell{1, 1, Matrix{Any}, Vector{Vector{Float64}}}

Same error for LSTMNeuralNetwork(4, (1,1)). So the description is accurate: nothing could have been depending on these, and no deprecation shim is owed.

No dangling references. git grep -l -in "RecurrentNeuralNetwork|LSTMNeuralNetwork|GridCell|AbstractExplicitCell|IdentityCell" across the whole tree — every file type, not just .jl/.md/.toml — returns CHANGELOG.md and nothing else: the new removal entry plus the historical 0.3.0 line. src/, test/, docs/, scripts/, legacy/ and .github/ are clean, and docs/make.jl never referenced them.

The import removal is forced, not merely tidy. src/cells/{abstract,grid,gru,identity,lstm,recurrent}.jl are present in AbstractNeuralNetworks 0.6.4 and absent from 0.7.0, so AbstractExplicitCell, GridCell and Recurrent had to go either way. None of the three was ever re-exported by GML, so RecurrentNeuralNetwork and LSTMNeuralNetwork are the only public names this drops.

Findings

1. The IdentityActivation / ZeroVector import is left dead — minor

src/GeometricMachineLearning.jl:71

import AbstractNeuralNetworks: IdentityActivation, ZeroVector

Chain(::RecurrentNeuralNetwork) was the only thing in src/ that named IdentityActivation, so this PR is what makes it dead. ZeroVector was dead already — its last src/ use moved to legacy/ in ec8e8fa5.

Worth saying why this is a nit and not a bug, since it cuts both ways:

• It is not a source of undeclared-binding warnings under 0.7 — both names survive there (ZeroVector in src/utils/zero_vector.jl, IdentityActivation still exported), so leaving it would not have undone what the PR set out to do.
• Removing it is not observable either. IdentityActivation is an AbstractNeuralNetworks export and line 3 is a blanket using AbstractNeuralNetworks, so GeometricMachineLearning.IdentityActivation still resolves without the explicit import. The two scripts/loss/write_loss_*.jl occurrences spell it fully qualified as AbstractNeuralNetworks.IdentityActivation and never depended on the line.

So: redundant, and this PR is the one in a position to notice. Deleted.

2. The reworded footnote repeats the sentence it hangs off — minor

docs/src/architectures/neural_network_integrators.md:42

The body already ends "…we almost always use a transformer-derived architecture when dealing with time series[^2]", and the new footnote then said "It was removed in favour of the transformer-derived architectures, which are the ones maintained for time series here" — the same claim twice, one line apart. Cut to the fact the footnote exists to record:

[^2]: `GeometricMachineLearning` used to ship an LSTM implementation. It has been removed.

Checked and dismissed

• Trailing whitespace on the footnote line — the entire file is written that way; not this PR's business.
• Project.toml still at 0.5.0 — this repo bumps at release, and ## [Unreleased] already carries other breaking entries.
• Open Issues C8 ("scripts/ has been dead since SymbolicNeuralNetworks 0.2") — deleting six of those scripts does not close it; the entry still stands as written.
• C10 (ten undefined exported names) — unaffected; both removed names were defined.

Verification of the follow-up

• Pkg.test() on df89a6e6: 56 testsets, zero failures, zero errors (twice, plus the pre-push hook run).
• Loads with no undeclared-binding warnings; isdefined is false for RecurrentNeuralNetwork, LSTMNeuralNetwork and ZeroVector, and IdentityActivation still resolves through the blanket using, as above.

🤖 Generated with Claude Code