Trained models and predictions for "Compact sensor-aided millimeter-wave beam prediction runs in real time on low-cost edge hardware"

Abstract

Trained models, per-sample test predictions and analysis outputs for the revised article "Compact sensor-aided millimeter-wave beam prediction runs in real time on low-cost edge hardware" (Scientific Reports, under revision). Code: https://github.com/himansh24dev/compute-efficient-beam-prediction (branch "revision").Data: DeepSense 6G scenarios 31-34 (https://www.deepsense6g.net), not redistributed. Usage: tar -xzf crisp_revision_models_predictions.tar.gz (MD5 c45f568dac17b0504dbb8d3a96745d41) Contents- ckpt/ .pt: validation-selected checkpoint of every trained model (126 runs), with the split's train-only GPS statistics, split seed, initialization seed and model configuration.- runs/ .json: metrics, validation-selected epoch, per-epoch validation history, split description.- runs/ .npz and runs/ __last.npz: per-sample test softmax (float16), target, scenario, row and pass key for the validation-selected and final-epoch models; Q_int8_* are the ONNX INT8 models.- analysis/: analysis.json (validation-selected), analysis_last.json (final epoch), robustness, communication metrics, INT8, delay-aligned and Raspberry Pi summaries. Run names: A_* headline split, five initializations (deployed model and cores/modalities); B_* further ablations; C_* split-protocol leakage (seed drives split and initialization); D_h{0,1,2}_* delay-aligned; E_w{2,4,8,16}_* window sweep. Every table and interval in the paper is recomputed from these files by experiments/revision/analyze.py in the code repository, without retraining.