13. Find a node, parameter, example or solution

Use this page as an index. Start with the user explanation and worked example; open the exact reference when you need spelling, defaults or availability. The node catalogue is the visual entry point for pipeline operations.

Pipeline nodes, one family at a time

What you want to do	User page with code and figure	Exact reference
Read the basic step forms	operator step	Writing a Pipeline in nirs4all
Change spectral/feature values	preprocessing	Transforms Reference
Change y for fitting and restore units	target processing	Pipeline Keywords Reference
Fit a regressor or classifier	model	Built-in Models Reference
Compare alternatives before fitting	generators	Generator keywords: choices, combinations and their exact outputs
Hold out rows/groups for validation	split	Splitters Reference
Duplicate paths or route sources/rows	branch	Pipeline Branching
Join features, predictions or row subsets	merge	Branch Merging
Annotate observations	tag	Pipeline Keywords Reference
Exclude permitted training rows	exclude	Filters Reference
Make derived training measurements	sample augmentation	Augmentation Operators Reference
Retain extra feature views	feature augmentation	Preprocessing Overview
Concatenate transformed columns	concat_transform	Pipeline Keywords Reference
Reshape related scans	repetitions	Heterogeneous Source Repetitions
Add inspection figures	charts	Analyzer Charts Reference
Choose transfer-oriented preprocessing	auto-transfer	Pipeline Keywords Reference
Learn a correction to a base predictor	residual model	Built-in Models Reference

Data and modalities

Question	Worked guide/example	What to inspect
How do I load arrays, files or partitions?	4. Describe your measurements and align your sources, Loading Data	Rows, features, target and partitions
How do I combine spectrum, image, series and metadata?	Native multimodal early fusion	Source encoders, aligned IDs and fusion
What if a modality is missing?	Fusion tardive avec observations partielles	Presence masks and supported fusion policy
What if a target is missing?	Raw multimodal pipelines	Observed-target masks and per-target fitting
What if series have different lengths?	Raw multimodal pipelines	Offsets, time coordinates and explicit encoder
What if I have repeated scans?	Sample Repetition & Aggregation	Specimen groups, origin IDs and scoring unit
What combinations can my runtime execute?	Public execution matrix	Actual source/model/export combination

Executable source companions include U07 multimodal, U08 targets, U09 missing sources, U12 ragged series, and U13 late missing sources.

Search, training and selection

I need…	Start here	Meaning
A small explicit comparison	5. Build a pipeline you can explain	Generate separate candidate recipes
Model-local parameter search	Hyperparameter Tuning	Tune one model in its allowed training scope
Native fold-safe PLS search	Native PLS HPO within each outer fold	Keep fitted preprocessing inside training folds
Structural search	Search source subsets, preprocessing chains and Ridge versus PLS	Compare chains, source subsets and fusion structures
Durable native tuning and resume	Native tuning and conformal calibration	Retain the real checkpoint/search identity
Three training-parameter scopes	Three training-parameter scopes	Trial controls, ordinary fitting and refit overrides
Retraining or weight continuation	Retrain and Transfer Learning	New fitting campaign versus model-specific continuation
Exact tuning keyword spelling	Pipeline Keywords Reference	Nested fields, defaults and runtime availability

Search changes the comparison, not the amount of independent evidence. Keep the final test cohort outside candidate selection. Resume the same search only with compatible data, recipe, objective and checkpoint.

Scores, uncertainty and stored results

I need…	Open…	Check
RMSE, R² or classification metric definition	6. Measure whether your model answers the question, Metrics Reference	Unit, direction, valid count and aggregation
CV selection versus refit test score	Understanding Scores, Evaluation, and Refit	Which observations produced each score
Diagnostic plots	7. Read the result and keep a useful experiment record, Analyzer Charts Reference	Partition, target names and readable values
Conformal intervals	Native tuning and conformal calibration	Separate calibration/test data and achieved coverage
A fixed-model robustness audit	Robustness/generalization reports	Scenario, severity, reference data and metric changes
Stored prediction queries	Predictions API Reference	Run, candidate, fold, partition and sample IDs
Workspace/session inspection	7. Read the result and keep a useful experiment record, Move a native model or a complete experiment	Stored state versus open runtime resources
Workspace/storage schema	Workspace Architecture, Storage API Reference	Actual producer format and selected-model links

Deployment and language help

10. Run the same recipe in your language: one shared native recipe with actual Python/R/Octave/WASM calls.
8. Save a predictor and use it on new observations: export, cold reload, raw-input prediction and diagnosis.
Move a native model or a complete experiment: moving a supported native predictor or full workspace.
Export and Deployment: complete fitted bundle.
Public Interfaces and Runtime: exact full Python SDK surface.
CLI reference: command groups, arguments and output files.

A recipe, a fitted model, an optimizer checkpoint and a workspace contain different things. Use the loader corresponding to the producing API. Keep exact native package text when large signed integers are present.

Diagnose a problem by what you see

Symptom	Inspect first	Concrete next action
Missing runtime or method	Language package, CLI/library path and versions	Follow 0. Choose the tools for your language and the installed runtime setup
Wrong array dimensions	Observation/feature axes and smallest fold size	Print shapes before fitting
Source schema mismatch	Source IDs, feature order, units and coordinates	Compare with the saved model declaration
Group/fold refusal	Which scans belong to one specimen	Keep the independent unit together; use explicit supported folds
Resume mismatch	Data, space, metric and actual checkpoint	Start a new study if the experiment changed
Model export cannot load	Original producer, bytes and supported consumer	Use the matching loader/profile
Correct values on wrong rows	Sample IDs across sources and output	Join by IDs and retain ordering
Unexpected fit during prediction	Whether fitted state was loaded	Call the prediction API on the saved predictor

Further help: Frequently Asked Questions, Dataset Configuration Troubleshooting Guide, Migration Guide.