Pybunny/nilmbench-faustine

FaustineCNN — NILMbench baseline

FaustineCNN trained on UK-DALE House 1 (sparse 6-second 16 kHz V/I frames) and evaluated on the dense House 2 set of 60,000 frames. This is the best-scoring baseline from the NILMbench paper (MJ_{20W} = 0.504 raw, 0.559 with the recall-constrained cutoffs in cutoffs.json).

Files

File	Purpose
faustine_best.pt	PyTorch state-dict, loadable with nilmbench.models.FaustineCNN.
cutoffs.json	Recall-constrained per-class cutoffs c_k (calibrated on House 1).
classes.json	Ordered category names (length 7).

Quick start

import torch, json
from huggingface_hub import hf_hub_download
from nilmbench.models import FaustineCNN

ckpt = hf_hub_download("Pybunny/nilmbench-faustine", "faustine_best.pt")
classes = json.load(open(hf_hub_download("Pybunny/nilmbench-faustine", "classes.json")))

model = FaustineCNN(n_categories=len(classes))
model.load_state_dict(torch.load(ckpt, map_location="cpu"))
model.eval()

Citation

NILMbench paper (2026). See https://github.com/Saharmgh/NILMbench for the full code and reproducible-figure scripts.

License

MIT.