- 
- 
- 
Bring licensed data
Datasets are not baked into packages or images. Download them from their official custodians, follow their licenses, and mount local NILMTK HDF5 conversions read-only.
+
+
+ From household power
+
+
+ Open infrastructure for energy disaggregation
+From household power
to reproducible insight.
+ NILMTK is a research ecosystem for loading real energy datasets, resolving their meaning, implementing disaggregation models, and comparing them under frozen protocols.
+ +-
+
- Python 3.11supported research path +
- Real dataREDD · UK-DALE · REFIT +
- Open sourceApache-2.0 ecosystem +
+
+
+ aggregate mains
+ power / time
+
+
+
+ mains
+ fridge
+ washer
+ kettle
+
+
+ One measured channel
+
+ Appliance estimates
+
+
+
+
+
+
+
+ Four repositories.
+
+ Choose by job
+Four repositories.
Four distinct responsibilities.
+ Start with the question you are trying to answer. The packages work together, but each has one canonical job. This separation keeps data semantics, model code, and scientific claims reviewable.
+-
+
-
+ 01+++
Data layer
+NILMTK
++ Load, convert, inspect, and score energy data. ++Use core for dataset converters, buildings and meters, preprocessing, metrics, and classical baselines.
++ Repository + API docs ++
+
+ -
+ 03+++
Model layer
+nilmtk-contrib
++ Use or contribute a disaggregation model. ++Use contrib for maintained classical, TensorFlow, and PyTorch implementations behind shared NILMTK contracts.
++ Repository + Paper ++
+ -
+ 04+++
Evidence layer
+NILMbench
++ Reproduce protocols and compare real runs. ++Use the benchmark for T1/T2/T3 tasks, provenance, CPU smoke tests, CUDA runs, immutable results, and the living leaderboard.
++ Repository + Leaderboard ++
+
+
+
+
+ The system
+One path from files to claims.
+NILM Metadata supplies shared meaning across the entire path. Core turns physical files into consistent meters and windows. Contrib supplies models. NILMbench freezes the comparison and publishes only provenance-complete results.
+ Explore the organization +
+
+
+ 01Real datasetsREDD · UK-DALE · REFIT
+
+ 02NILMTK coremeters · windows · metrics
+
+ 03Contrib modelsshared model contracts
+
+ 04NILMbenchprotocols · results · rank
+
+
+
+
+
+
+ A supported route,
+
+ Start clean
+A supported route,
not another dated tutorial.
+ Use Python 3.11 and uv. Pick the smallest installation that matches your job. The model install brings core and metadata with it.
+
+
+
+
+
+
+
+
+
+
+
+
+
+ PyTorch model environment
+ # Python 3.11 · CPU-safe install
+uv venv --python 3.11
+source .venv/bin/activate
+UV_TORCH_BACKEND=cpu uv pip install \
+ "nilmtk-contrib[torch] @ git+https://github.com/nilmtk/nilmtk-contrib.git"
+python -c "from nilmtk_contrib.torch import Seq2PointTorch; print('ready')"
+
+ NILMTK data environment
+ # Converters, meters, metadata, and metrics
+uv venv --python 3.11
+source .venv/bin/activate
+uv pip install \
+ "nilmtk @ git+https://github.com/nilmtk/nilmtk.git"
+nilmtk-convert --help
+
+ General NILMTK environment
+ # One onboarding Dockerfile; no image per model
+git clone https://github.com/nilmtk/nilmtk-contrib.git
+cd nilmtk-contrib
+docker build -t nilmtk-contrib:latest .
+docker run --rm -it nilmtk-contrib:latest bash
+
+ Reproducible benchmark runner
+ git clone https://github.com/nilmtk/nilmbench.git
+cd nilmbench
+uv venv --python 3.11
+source .venv/bin/activate
+UV_TORCH_BACKEND=cpu uv pip install ".[benchmark]"
+nilmbench list
+
+ A
+
+
+ B
+
+
+
+ C
+
+
+
+
+
Keep one general image
The contrib Dockerfile defines the common starting environment. Model contributions do not create a new container tag or image family.
Separate benchmark runtimes
NILMbench owns pinned CPU-smoke and CUDA-benchmark runtimes. Those images certify results; they are not per-algorithm development images.
+
+
+ Credit the layer
+
+ Cite precisely
+Credit the layer
your work relies on.
+ There is no single citation that describes every part of the current ecosystem. Cite the infrastructure paper for each layer you use, plus the original model and dataset papers.
+ Copy BibTeX from the repositories +
+
+
+
+ | If your work uses… | Cite | DOI |
|---|---|---|
| Dataset conversion, meters, preprocessing, or metricsNILMTK core | +NILMTK: An Open Source Toolkit for Non-intrusive Load Monitoring | +10.1145/ |
+
| Schema, taxonomy, synonyms, or dataset semanticsNILM Metadata | +Towards Reproducible Research in Energy Disaggregation | +10.1109/ |
+
| The shared model suite or rapid experimentation APInilmtk-contrib | +Towards Reproducible State-of-the-Art Energy Disaggregation | +10.1145/ |
+
| T1/T2/T3 protocols, runner, or leaderboard resultsNILMBench2026 · BuildSys '26 | +NILMBench2026: A Benchmark for Energy Disaggregation | +10.1145/ |
+
Also cite: every evaluated model's original paper and every dataset used. A leaderboard row is evidence, not a substitute for source attribution.
+
+
+
+
+
+
+ One model contribution.
+
+ Extend the ecosystem
+One model contribution.
One auditable route to rank.
+ Algorithm code and benchmark evidence move at different cadences. A model can merge once its implementation contract is sound. It enters the leaderboard only after the pinned benchmark matrix passes on real data.
+-
+
- 01
Implement
Add the model to nilmtk-contrib using shared validation, windowing, checkpoint, and logging utilities.
Model repository ↗
+ - 02
Break it on purpose
Test malformed inputs, short chunks, lazy imports, determinism, serialization, CPU smoke, and the targeted CUDA path.
Model tests ↗
+ - 03
Register
Add only the adapter and search space to NILMbench. Dataset and task logic stays in the benchmark, never in a model notebook.
Benchmark repository ↗
+ - 04
Run real data
Freeze source revisions, image digest, data identity, hardware, seeds, and protocol. Candidate results remain separate.
Run protocol ↗
+ - 05
Publish evidence
Review immutable result bundles; CI regenerates the leaderboard and rejects hand-edited or incomparable ranks.
Living leaderboard ↗
+
+
+
+
+ Choose a repository
+ New here?
+Start with a reproducible failure.
+Open an issue with the repository, exact command, operating system, Python version, dataset identity, and the smallest trace that reproduces the problem.
+