Graham McIntire
e7a7ae073d
Task 9.3 - Weight recalibration via gradient descent:
- Recalibrator module fits logistic regression weights using Nx
- Trains on QSO positives vs random baseline negatives
- Cross-validates by month, normalizes weights to sum to 1.0
- Mix task: mix recalibrate_scorer --sample 5000 --epochs 2000
Task 9.4 - Side-by-side scorer comparison:
- ScorerDiff.compare/3 re-scores grid with old vs new weights
- Reports mean diff, regressions, improvements, per-band breakdown
- Mix task: mix scorer_diff --new-weights '{...}'
Phase 3 - NEXRAD ingestion pipeline:
- NexradClient fetches IEM n0q composite PNGs, extracts per-point
box statistics (mean/max dBZ, texture variance)
- NexradObservation schema with unique (lat, lon, observed_at)
- NexradWorker on :nexrad queue for background processing
- nexrad_texture backtest feature in Features module
- mix nexrad_backfill --limit 200
All tasks added to AdminTaskWorker and Release for production use.
1116 tests, 0 failures.
92 lines
2.9 KiB
Elixir
92 lines
2.9 KiB
Elixir
defmodule Mix.Tasks.RecalibrateScorer do
|
|
@shortdoc "Recalibrate propagation scorer weights via gradient descent"
|
|
@moduledoc """
|
|
Fits new scoring weights using logistic regression on the historical
|
|
QSO corpus. Compares QSO conditions (positive examples) against
|
|
random-time baselines (negative examples) to learn which factors
|
|
best discriminate actual propagation events.
|
|
|
|
## Usage
|
|
|
|
mix recalibrate_scorer
|
|
mix recalibrate_scorer --sample 5000 --epochs 2000 --lr 0.01
|
|
|
|
## Options
|
|
|
|
* `--sample` - max contacts to load (default: 5000)
|
|
* `--epochs` - training iterations (default: 2000)
|
|
* `--lr` - learning rate (default: 0.01)
|
|
"""
|
|
use Mix.Task
|
|
|
|
alias Microwaveprop.Propagation.BandConfig
|
|
alias Microwaveprop.Propagation.Recalibrator
|
|
|
|
@impl Mix.Task
|
|
def run(argv) do
|
|
Mix.Task.run("app.start")
|
|
Oban.pause_all_queues(Oban)
|
|
|
|
{opts, _, _} =
|
|
OptionParser.parse(argv,
|
|
switches: [sample: :integer, epochs: :integer, lr: :float]
|
|
)
|
|
|
|
sample_size = Keyword.get(opts, :sample, 5000)
|
|
epochs = Keyword.get(opts, :epochs, 2000)
|
|
learning_rate = Keyword.get(opts, :lr, 0.01)
|
|
|
|
Mix.shell().info("Recalibrating scorer weights...")
|
|
Mix.shell().info(" sample_size: #{sample_size}")
|
|
Mix.shell().info(" epochs: #{epochs}")
|
|
Mix.shell().info(" learning_rate: #{learning_rate}")
|
|
Mix.shell().info("")
|
|
|
|
result =
|
|
Recalibrator.fit(
|
|
sample_size: sample_size,
|
|
epochs: epochs,
|
|
learning_rate: learning_rate
|
|
)
|
|
|
|
current_weights = BandConfig.weights()
|
|
|
|
Mix.shell().info("=== Results ===")
|
|
Mix.shell().info("")
|
|
Mix.shell().info("Train loss: #{format_float(result.train_loss)}")
|
|
Mix.shell().info("Val loss: #{format_float(result.val_loss)}")
|
|
Mix.shell().info("Initial loss: #{format_float(result.initial_loss)}")
|
|
Mix.shell().info("")
|
|
|
|
Mix.shell().info("=== Weight Comparison ===")
|
|
Mix.shell().info("")
|
|
|
|
header =
|
|
String.pad_trailing("Factor", 18) <> String.pad_trailing("Current", 10) <> String.pad_trailing("New", 10) <> "Delta"
|
|
|
|
Mix.shell().info(header)
|
|
Mix.shell().info(String.duplicate("-", 48))
|
|
|
|
for key <- result.weights |> Map.keys() |> Enum.sort() do
|
|
current = Map.get(current_weights, key, 0.0)
|
|
new_val = result.weights[key]
|
|
delta = new_val - current
|
|
|
|
sign = if delta >= 0, do: "+", else: ""
|
|
|
|
line =
|
|
String.pad_trailing(to_string(key), 18) <>
|
|
String.pad_trailing(format_float(current), 10) <>
|
|
String.pad_trailing(format_float(new_val), 10) <>
|
|
"#{sign}#{format_float(delta)}"
|
|
|
|
Mix.shell().info(line)
|
|
end
|
|
|
|
Mix.shell().info("")
|
|
Mix.shell().info("To apply these weights, update @weights in lib/microwaveprop/propagation/band_config.ex")
|
|
end
|
|
|
|
defp format_float(f) when is_float(f), do: :erlang.float_to_binary(f, decimals: 4)
|
|
defp format_float(f), do: to_string(f)
|
|
end
|