Initial commit: astrape-oracle split out of the combined Astrape tree
The forecasting runtime gibil pilots: weather ingest (Open-Meteo forecasts and the Ecowitt station), the oracle and weather stores, the snapshot daemon, the forecast evaluator, and the TCN training scheduler. classes/datasets.py holds the shapes the stores return for display. They sit beside the stores deliberately: a store's return type belongs to the store, and keeping them with the renderers would mean this repo importing the dashboard, which is backwards. Extracted from a single repository where the forecasting engine, the database plumbing, the hardware client, the ingest daemons and the web UI all imported each other freely. Verified that every module imports with gibil resolved from its own installed checkout. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
rpotter6298
@@ -0,0 +1,158 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
from os import environ
|
||||
from pathlib import Path
|
||||
|
||||
from astrape_ems.classes.env import EnvLoader
|
||||
from astrape_ems.classes.store import SigenStore
|
||||
from astrape_oracle.classes.weather.store import WeatherStore
|
||||
from gibil.classes.usage_sequence_dataset import (
|
||||
UsageSequenceDatasetBuilder,
|
||||
UsageSequenceDatasetConfig,
|
||||
)
|
||||
from gibil.classes.usage_training import train_model
|
||||
|
||||
|
||||
def main() -> None:
|
||||
EnvLoader().load()
|
||||
environ.setdefault("TORCH_CPP_LOG_LEVEL", "ERROR")
|
||||
args = parse_args()
|
||||
config = UsageSequenceDatasetConfig.from_env()
|
||||
builder = UsageSequenceDatasetBuilder(
|
||||
config=config,
|
||||
plant_history=SigenStore.from_env(),
|
||||
weather_history=WeatherStore.from_env(),
|
||||
)
|
||||
examples = builder.build(limit=args.limit)
|
||||
|
||||
print(f"usage_sequence_examples={len(examples)}")
|
||||
print(
|
||||
"minimum_history_hours="
|
||||
f"{builder.max_past_hours + config.future_hours}"
|
||||
)
|
||||
print(f"past_features={len(builder.past_feature_names)}")
|
||||
for scale in config.past_scales:
|
||||
print(
|
||||
f"past_scale={scale.name} "
|
||||
f"hours={scale.hours} "
|
||||
f"step_seconds={scale.step_seconds} "
|
||||
f"steps={builder.past_steps(scale)}"
|
||||
)
|
||||
print(f"future_steps={builder.future_steps}")
|
||||
print(f"future_features={len(builder.future_feature_names)}")
|
||||
|
||||
if examples:
|
||||
first = examples[0]
|
||||
last = examples[-1]
|
||||
print(f"first_issued_at={first.issued_at.isoformat()}")
|
||||
print(f"last_issued_at={last.issued_at.isoformat()}")
|
||||
for name, rows in first.past_by_scale.items():
|
||||
print(f"first_past_{name}_shape={len(rows)}x{len(rows[0])}")
|
||||
token_count = sum(
|
||||
len(tokens)
|
||||
for tokens in first.past_tokens_by_scale[name]
|
||||
)
|
||||
print(f"first_past_{name}_tokens={token_count}")
|
||||
print(
|
||||
"first_future_feature_shape="
|
||||
f"{len(first.future_features)}x{len(first.future_features[0])}"
|
||||
)
|
||||
print(
|
||||
"first_future_tokens="
|
||||
f"{sum(len(tokens) for tokens in first.future_tokens)}"
|
||||
)
|
||||
print(f"first_targets={len(first.targets)}")
|
||||
print(
|
||||
"first_target_preview="
|
||||
+ ",".join(f"{value:.0f}" for value in first.targets[:8])
|
||||
)
|
||||
|
||||
if args.dry_run:
|
||||
return
|
||||
|
||||
if not examples:
|
||||
raise SystemExit("No usage sequence examples available for training yet.")
|
||||
|
||||
result = train_model(
|
||||
examples=examples,
|
||||
builder=builder,
|
||||
epochs=args.epochs,
|
||||
batch_size=args.batch_size,
|
||||
learning_rate=args.learning_rate,
|
||||
artifact_path=args.artifact_path,
|
||||
seed=args.seed,
|
||||
base_artifact_path=args.base_artifact_path,
|
||||
mode=args.mode,
|
||||
)
|
||||
print(f"training_mode={result.mode}")
|
||||
print(f"training_examples={result.training_examples}")
|
||||
print(f"validation_examples={result.validation_examples}")
|
||||
if result.final_train_loss is not None:
|
||||
print(f"final_train_pinball_loss={result.final_train_loss:.4f}")
|
||||
if result.final_validation_loss is not None:
|
||||
print(f"final_validation_pinball_loss={result.final_validation_loss:.4f}")
|
||||
|
||||
|
||||
def parse_args() -> argparse.Namespace:
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Build training windows for the sequence usage oracle."
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dry-run",
|
||||
action="store_true",
|
||||
help="Only build examples and print dataset shape.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--limit",
|
||||
type=int,
|
||||
default=None,
|
||||
help="Optional maximum number of examples to build.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--epochs",
|
||||
type=int,
|
||||
default=20,
|
||||
help="Training epochs.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--batch-size",
|
||||
type=int,
|
||||
default=32,
|
||||
help="Training batch size.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--learning-rate",
|
||||
type=float,
|
||||
default=0.001,
|
||||
help="Adam learning rate.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--artifact-path",
|
||||
type=Path,
|
||||
default=Path("models/usage_sequence_tcn_v1.pt"),
|
||||
help="Where to save the trained TCN artifact.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--base-artifact-path",
|
||||
type=Path,
|
||||
default=None,
|
||||
help="Existing artifact to continue training from.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--mode",
|
||||
choices=("full", "finetune"),
|
||||
default="full",
|
||||
help="Training mode recorded in the saved artifact.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--seed",
|
||||
type=int,
|
||||
default=7,
|
||||
help="Deterministic shuffle seed.",
|
||||
)
|
||||
return parser.parse_args()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user