---

name: network-api description: "Build, link, run, inspect, and debug NuPIC legacy Network API pipelines and custom PyRegion regions." disable-model-invocation: true metadata: disco-role: operating license: MIT

NuPIC Legacy Network API

Use this sub-skill when the user asks to build a NuPIC Network, add/link regions, initialize/run a Network API pipeline, read region outputs, debug link names, or create a custom Python region. NuPIC legacy is a Python 2.7-era package; Network API workflows normally require nupic.bindings, numpy 1.12.x-compatible runtime packages, and working Cap'n Proto/pycapnp support for some serialization paths.

Do not use this sub-skill for every prediction task. If the user wants the higher-level OPF model workflow, route to ../opf-prediction/. If the task is about CSV headers, FileRecordStream, or model parameter files, route to ../data-and-configuration/. If the task is about encoder, SP, TM, or classifier parameter semantics outside a Network graph, route to ../htm-algorithms/. For package installation/import failures, use the root troubleshooting guide at ../../references/troubleshooting.md first.

Fast path

1. Confirm that the legacy runtime can import and construct an empty Network:

   python scripts/network_smoke.py
   

2. Build the graph in this order: Network() -> addRegion(...) -> attach RecordSensor.encoder and RecordSensor.dataSource -> set SPRegion inputWidth -> link(...) -> set predictedField/modes -> initialize() -> run(n).
3. Use region type strings exactly: py.RecordSensor, py.SPRegion, py.TMRegion, and py.SDRClassifierRegion unless a custom region has been registered.
4. Use explicit link endpoint names when debugging: srcOutput="dataOut", destInput="bottomUpIn", srcOutput="bottomUpOut", destInput="bottomUpIn", and sensor-to-classifier bucketIdxOut, actValueOut, categoryOut links.
5. Query outputs with network.regions["name"].getOutputData("outputName"); classifier predictions come from actualValues and flattened probabilities arrays.

Bundled runtime resources

• references/network-workflows.md: read this for the complete sensor -> SP -> TM -> classifier construction recipe, link table, output shapes, validation checks, and safe serialization notes.
• references/custom-region.md: read this when the user needs a custom PyRegion skeleton, registration steps, getSpec() contract, or dynamic output width handling.
• references/troubleshooting.md: read this when Network imports, region type strings, link endpoint names, initialization, typed parameters, classifier outputs, or bundle loading/saving fail.
• scripts/network_smoke.py: run this deterministic helper to import nupic.engine.Network, construct an empty network by default, optionally inspect built-in region specs, and print a safe next-step pipeline template without requiring source data files.

Minimal Network API shape

import json
from nupic.engine import Network

network = Network()
network.addRegion("sensor", "py.RecordSensor", "{}")
# Attach a MultiEncoder and FileRecordStream to network.regions["sensor"].getSelf().

network.addRegion("SP", "py.SPRegion", json.dumps(spParams))
network.addRegion("TM", "py.TMRegion", json.dumps(tmParams))
network.addRegion("classifier", "py.SDRClassifierRegion", json.dumps(clParams))

network.link("sensor", "SP", "UniformLink", "", srcOutput="dataOut", destInput="bottomUpIn")
network.link("SP", "TM", "UniformLink", "", srcOutput="bottomUpOut", destInput="bottomUpIn")
network.link("TM", "classifier", "UniformLink", "", srcOutput="bottomUpOut", destInput="bottomUpIn")
network.link("sensor", "classifier", "UniformLink", "", srcOutput="bucketIdxOut", destInput="bucketIdxIn")
network.link("sensor", "classifier", "UniformLink", "", srcOutput="actValueOut", destInput="actValueIn")
network.link("sensor", "classifier", "UniformLink", "", srcOutput="categoryOut", destInput="categoryIn")

network.regions["sensor"].setParameter("predictedField", "consumption")
for name in ("SP", "TM", "classifier"):
  network.regions[name].setParameter("learningMode", 1)
  network.regions[name].setParameter("inferenceMode", 1)

network.initialize()
network.run(1)

Keep generated or copied user examples self-contained. Do not require future agents to open the original repository docs, tests, scripts, examples, notebooks, or local checkout paths.