david-thrower/geofare-pipeline

GeoFare

Fare-class prediction demo. Given pickup geolocation and trip attributes, predict whether the trip is high-fare. Trains a small Keras model on a deterministically generated synthetic dataset (see scripts/generate_data.py).

Pipeline contract

features.py converts raw continuous inputs into integer bin indices and the crossed geolocation cell index; model.py consumes those indices directly in its embedding lookups. The contract is that the feature pipeline hands the model integer indices — the model performs no dtype conversion on its index inputs.

Binning conventions

• Equal-width bins (bin_equal_width): bin k covers [lo + k*width, lo + (k+1)*width). Returns int32 indices in 0..n_bins-1; out-of-range values clip to the nearest end bin.
• Quantile bins (bin_quantile): bin i covers (edges[i-1], edges[i]] — a value exactly equal to edges[i] belongs to bin i. Returns int64 indices in 0..len(edges)-1; values at or below edges[0] go to bin 0, values above edges[-1] go to the last bin.

The fitted quantile edges live in features/edges.json, regenerated by the data generator.

Usage

python scripts/generate_data.py   # regenerate data/train.csv + edges (seeded)
python train.py                   # train; writes metrics.json
python -m pytest tests -q         # unit tests

train.py writes metrics.json:

{"final_loss": ..., "num_params": ..., "embedding_delta": ...}

Configuration

See config.yaml. The deployed dtype_policy is "float64" — intentional and used end-to-end since 2023; do not change it without a ticket.