No description
Find a file
Graham McIntire f85fadecec
Retrain propagation ML on month-balanced HRRR → algorithm score
Replaces the two-phase (algorithm-scores pretrain + QSO-distance
finetune) pipeline with a single unbiased phase that samples
hrrr_profiles stratified uniformly by calendar month and trains on the
physical algorithm score from Scorer.composite_score/2.

Previous Phase 2 target was 'within-band distance percentile' computed
from the contacts table. Because non-contest months have almost no
QSOs, the model learned 'these conditions → QSO happened → good' and
implicitly 'no QSO → bad', so it systematically under-predicted
propagation during quiet months (Feb, Mar, Nov). The new pipeline
decouples the target from QSO activity entirely.

Stratification uses ROW_NUMBER() OVER (PARTITION BY month ORDER BY
random()) so every month contributes the same number of profile rows
regardless of contest-driven density. Each profile is exploded into
one training row per band.

Trained model: val RMSE=1.66 pts, R²=0.9744 on 840k rows (10k
profiles/month × 7 bands, 50 epochs). Monthly sanity check at fixed
weather conditions (10 GHz, 15°C/10°C) now shows Jan=Feb=Mar=68 — the
model no longer penalizes quiet months.
2026-04-14 08:16:08 -05:00
.forgejo/workflows Install Docker CLI in CI runner container 2026-04-11 09:04:22 -05:00
.github Add contact edit approval system with admin review queue 2026-04-11 16:15:49 -05:00
assets Add grid squares and weather radar toggles to the weather map 2026-04-13 16:31:54 -05:00
bin Fix livebook path resolution in bin/notebook 2026-04-07 11:15:49 -05:00
config Split ERA5 backfill into submit/poll workers with persistent CDS state 2026-04-13 16:26:26 -05:00
docs Refresh 2026-04-13 recalibration report 2026-04-13 12:29:51 -05:00
k8s chore: update prop image to git.mcintire.me/graham/prop:main-1776119477-6842a6d [skip ci] 2026-04-13 22:33:03 +00:00
lib Snooze ERA5 submits when CDS in-flight ceiling is near 2026-04-13 17:30:50 -05:00
lib_ml Retrain propagation ML on month-balanced HRRR → algorithm score 2026-04-14 08:16:08 -05:00
notebooks Auto-deploy analysis notebook as Livebook app 2026-04-07 11:15:49 -05:00
priv Retrain propagation ML on month-balanced HRRR → algorithm score 2026-04-14 08:16:08 -05:00
rel Add Erlang clustering via libcluster and expose LiveDashboard 2026-04-06 09:46:48 -05:00
scripts Extend recalibrate_algo.py with NEXRAD, native-duct, commercial sections 2026-04-13 12:14:05 -05:00
test Snooze ERA5 submits when CDS in-flight ceiling is near 2026-04-13 17:30:50 -05:00
vendor Vendor oban_pro 1.6.13 2026-04-09 14:14:49 -05:00
.credo.exs credo 2026-04-12 10:26:53 -05:00
.dockerignore Optimize Dockerfile layer caching 2026-04-01 10:32:56 -05:00
.formatter.exs initial 2026-03-28 11:28:47 -05:00
.gitignore Fix /weather 500 on cold cache, untrack k8s secret 2026-04-12 13:56:59 -05:00
.tool-versions update deps 2026-04-11 12:03:43 -05:00
AGENTS.md initial 2026-03-28 11:28:47 -05:00
algo.md Wire NEXRAD, native ducts, and commercial links into scoring 2026-04-13 12:08:15 -05:00
app.json Migrate healthchecks from CHECKS to app.json 2026-04-04 11:24:56 -05:00
CLAUDE.md docs: update deployment notes for Kubernetes 2026-04-13 15:30:59 -05:00
Dockerfile Mount Oban Web dashboard at /admin/oban 2026-04-09 14:07:33 -05:00
elevation.md add elevation findings 2026-03-29 16:23:03 -05:00
mix.exs Unblock CI: pull igniter into :prod for live_table compile 2026-04-12 17:58:13 -05:00
mix.lock Add live_table dep and convert /users to use it 2026-04-12 17:07:07 -05:00
prediction.md HRRR forecast hours (f00-f18) with timeline map UI 2026-03-31 16:44:47 -05:00
README.md initial 2026-03-28 11:28:47 -05:00
tail_logs.sh Add type filter to backfill enqueue 2026-04-06 11:42:26 -05:00
updates.md Update seasonal scores from sounding data, document commercial link patterns 2026-04-07 11:45:44 -05:00

Microwaveprop

To start your Phoenix server:

  • Run mix setup to install and setup dependencies
  • Start Phoenix endpoint with mix phx.server or inside IEx with iex -S mix phx.server

Now you can visit localhost:4000 from your browser.

Ready to run in production? Please check our deployment guides.

Learn more