Skip to content

OpenNN 26.0.3 - #354

Closed
RubyAM wants to merge 1 commit into
masterfrom
dev
Closed

RubyAM wants to merge 1 commit into
masterfrom
dev

Conversation

@RubyAM

@RubyAM RubyAM commented Oct 20, 2026

Copy link
Copy Markdown
Collaborator

Promotes dev to master, including complete input and output names for trained forecasting networks.

Validation: code-quality and architecture checks passed locally. CI passed on dev at 83d7513: https://github.com/Artelnics/opennn/actions/runs/39416123404. Existing local verification logs report CPU: 1,218 passed / 52 skipped; CUDA: 1,408 passed / 7 skipped. These are prior dev results; checks on this PR remain pending. No package version change or release tag is included.

Optimizer::set_names() gave the network the dataset's Input and Target
variables, one name per variable. A forecasting network has one input per
variable and lag and, with several steps ahead, one output per target and
step. After training it was left with fewer names than neurons. In Neural
Designer, a trained 24-lag, 12-step NO2 example then failed Report neural
network ("Size of rows name(1) is not equal to number of rows in data(12)"),
crashed Plot directional output and stopped both input importance tasks with
"bad allocation"; the untrained model was fine.

set_names() now uses get_model_input_variables(), which already expands the
inputs by lag (as input selection does), and a new
get_model_target_variables(). For a multi-step forecasting dataset it names
each target per step, <target>_ahead_t<step>, in the order fill_targets
writes them and as Neural Designer names them on import. Other datasets keep
their target variables.

The size ratchet also absorbs the two lines of 0b2cf35, which made CI's
"Enforce code size and architecture" step fail on dev and master.

Validation: AdamTest.ForecastingTrainingNamesEveryInputAndOutput failed
without the optimizer change (2 input names for 6 inputs) and passes with it;
TabularForecasting.ModelTargetVariablesNameEveryStepAhead pins the names and
their order. Full CPU suite 1270 tests, 0 failures (52 skipped: CUDA, oneDNN,
PHP, Node); CUDA suite on an RTX 4080 1415 tests, 0 failures (7 skipped:
oneDNN, PHP, Node, FlashAttention kernel). Quality, architecture and
dependency relocation checks pass. In Neural Designer the trained NO2 example
keeps 24 lag inputs and 12 step outputs and the four tasks pass; golden 147
compared, 0 FAIL; integration 301 ok, 0 FAIL; oracle, matrix, editor and
viewer suites pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@RubyAM RubyAM closed this Oct 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant