Inference

The inference for the track reconstruction problem is executed with:

python src/trackml_inference.py \
    --param tune-5 \
    --cluster "transformer" \
    --epoch -1 \
    --event_start 8750 --event_end 9999 \
    --read_tag f-0p1-hyper-5 \
    --out_tag  f-0p1-hyper-5 \
    --rfactor 0.1 --noise_ratio 0.05 \
    --node2node hyper --ncell 262144 \
    --fp_dtype "float32"

See more examples under /tests/inference* for different pile-up scenarios.

Main parameters

The inference steering program parameters can be seen with the command --help. The main parameters are:

--edge_threshold 0.55
--cluster "transformer" (cut, dbscan, hdbscan)

The edge_threshold parameter defines the clustering efficiency/purity/latency tradeoff. Small values retain more graph edges for the clustering stage after the GNN and load the transformer, but increase latency.

Similarly with --cluster algorithm, the transformer should provide always the highest efficiency and purity, but at the cost of latency. The clustering algorithm parameters are under model hyperparameters /hypertrack/models/global_<TAG>.py, such as the pivot search for the transformer. More exhaustive search can improve results but at the cost of increased computing time.