Getting Started

Building Models

Assemble readable neural architectures with NeuraLib’s fluent Sequential API, then build deterministic weights or compile the model for training.

Updated July 12, 2026Publisher Alien_AlgorithmsPine Script v6

Sequential API

Each chained method adds one block, so the code reads from input to output like an architecture diagram.

A small regression modelpine

var nl.Sequential model = nl.sequential("my_model")

if barstate.isfirst
    model := model
      .input(array.from(4), "features")
      .dense(8, nl.ActivationKind.relu, "hidden")
      .dropout(0.10, "regularization")
      .dense(1, nl.ActivationKind.linear, "output")
      .compile(cfg)

Initialize once

Create architecture and training configuration inside barstate.isfirst. Persistent model variables should use var so their learned state survives across bars.

Core layers

Method	Purpose
`.input(dims, name)`	Defines the model input shape. It must be the first architecture call.
`.dense(units, activation, name)`	Adds a fully connected trainable layer.
`.dropout(rate, name)`	Randomly masks activations during training to reduce co-adaptation.
`.layerNorm(name)`	Normalizes layer features to stabilize deeper paths.
`.activation(kind, alpha, name)`	Adds a standalone activation transform.
`.qHead(actionCount, activation, name)`	Adds a core action-value output head.
`.flatten(name) / .reshape(dims, name)`	Changes shape metadata between compatible blocks.
`.block(graphBlock)`	Inserts a custom or advanced GraphBlock.

Build vs. compile

Compile for training

.compile(cfg) builds missing weights and attaches loss, optimizer, metric, schedule, clipping, and execution-gate behavior. Use it when the model will train.

Build for deterministic inference

.build(nl.rng(seed)) initializes weights with a reproducible random stream without requiring a training configuration. It is useful for inference-only demonstrations, graph probes, and architecture testing.

A model can be built and then compiled when you need explicit control over initialization followed by training configuration.

Initialization and persistence

Use stable, unique names for models and blocks when debugging graph structure.
Use nl.rng(seed) for reproducible initialization.
Changing architecture or hyperparameters causes Pine to rebuild script state.
Export weights with .getWeightsArray() and restore them with .setWeightsArray().

Shape discipline

Most architecture errors are shape errors. Keep a written contract for input features, ordering, dimensions, output targets, and batch layout. NeuraLib validates calls before expensive graph work, but a mathematically valid shape can still represent the wrong semantic ordering.

Names do not reorder data

Feature labels improve readability; they do not align inputs automatically. Training and live inference must push values in exactly the same order.