Latest results from the full-epoch run (20260627023536), evaluated on the full held-out validation set of 322,915 samples spanning the complete loss distribution:

Metric	Value
RMSE	7.36 dB
MAE	4.73 dB
Median Error	2.99 dB
90th Percentile Error	10.71 dB
95th Percentile Error	14.67 dB

A median error of ~3 dB means half of all predictions are within 3 dB of ITM. The gap between the median and the tail percentiles reflects a minority of hard, high-loss cases that dominate the RMSE.

Runtime

Current runtime artifacts are in eval/.

Engine	Time / sample	Throughput
Direct ITM	11.0 us	91,082 pred/s
Transformer	1314.8 us	761 pred/s

The current repo validates the concept that attention can learn the ITM mapping, but it does not yet outperform native ITM in runtime.

Files

File	Description
`model_inference.pth`	Raw `state_dict` for inference (stable canonical name).
`model_weights20260627023536.pth`	Same weights, timestamped (full-epoch run).
`model_weights20260627023536_resume.pth`	Full checkpoint (model + optimizer + scheduler) for resuming training.
`loss_log_20260627023536_.npy.npz`	Training loss history.
`benchmark_20260627023536.json`	Final validation metrics.

Usage

python

import torch
from huggingface_hub import hf_hub_download
from pathloss_transformer import create_model, load_weights  # from the GitHub repo

path = hf_hub_download("alexcpn/elevation_transformer", "model_inference.pth")
model = create_model()
load_weights(model, path)   # torch.load(..., weights_only=True) + load_state_dict
model.eval()

For resuming training, download model_weights20260627023536_resume.pth instead — it is a dict of {model, optimizer, scaler, scheduler, step}, not a raw state_dict.