What it does

This repo implements a hybrid model for spotting weirdness in time series without labeled anomalies. A VAE compresses short windows of data into low-dimensional embeddings; an LSTM then watches those embeddings unfold over longer horizons. The reconstruction error becomes your anomaly score. It ships with pre-processed Numenta Anomaly Benchmark (NAB) data and a notebook to set detection thresholds by eyeballing histograms.

The interesting bit

The architecture is essentially a division of labor by attention span: the VAE handles local structure, the LSTM handles temporal drift. The authors published this at ICASSP 2020, and the code is pure TensorFlow 1.5 — a version now old enough to vote in some countries.

Key highlights

• Pre-processed NAB datasets included; training set is guaranteed anomaly-free, test set keeps the known anomalies
• Config-driven training via NAB_config.json; model files split across train.py, models.py, trainers.py, etc.
• Jupyter notebook (NAB-anomaly-detection.ipynb) walks through inference and threshold selection
• Reproduces the paper’s office-temperature detection example out of the box

Caveats

• TensorFlow 1.5 dependency means you’ll likely need a dedicated virtualenv or Docker time capsule
• Threshold setting is manual: “observe the histogram and set accordingly” — no automatic calibration provided
• Code structure is somewhat scattered across seven files in codes/

Verdict

Worth a look if you’re researching hybrid VAE-RNN architectures or need a reproducible baseline from the anomaly-detection literature. Skip it if you want production-ready code with modern frameworks or automatic threshold tuning.