import argparse import importlib.metadata import json import time from pathlib import Path import numpy as np import yaml from src.common import CLASSES, configure from src.dataset import load_data, pipeline from src.model import build_model def main(): p = argparse.ArgumentParser( description="Train CELYRIX from scratch on real EuroSAT RGB." ) p.add_argument("--config", default="configs/training.yaml") for key, typ in [ ("batch-size", int), ("epochs", int), ("learning-rate", float), ("seed", int), ("sample-limit", int), ]: p.add_argument("--" + key, type=typ) p.add_argument("--output-dir") args = vars(p.parse_args()) cfg = yaml.safe_load(Path(args.pop("config")).read_text()) cfg.update({k: v for k, v in args.items() if v is not None}) if cfg["sample_limit"] and cfg["sample_limit"] < 5: p.error("sample limit must be at least five") device = configure(cfg["seed"]) import tensorflow as tf out = Path(cfg["output_dir"]) out.mkdir(parents=True, exist_ok=True) start = time.monotonic() splits, provenance = load_data(cfg["data_dir"], cfg["seed"]) x, y = splits["train"] if cfg["sample_limit"] and cfg["sample_limit"] < len(y): from sklearn.model_selection import train_test_split selected, _ = train_test_split( np.arange(len(y)), train_size=cfg["sample_limit"], stratify=y, random_state=cfg["seed"], ) x, y = x[selected], y[selected] provenance["training_subset"] = selected.tolist() model = build_model(cfg["seed"]) model.compile( optimizer=tf.keras.optimizers.Adam(cfg["learning_rate"]), loss="sparse_categorical_crossentropy", metrics=["accuracy"], ) model.summary() history = model.fit( pipeline(x, y, cfg["batch_size"], True, cfg["seed"]), validation_data=pipeline(*splits["validation"], cfg["batch_size"]), epochs=cfg["epochs"], callbacks=[ tf.keras.callbacks.ModelCheckpoint( str(out / "model.keras"), monitor="val_loss", save_best_only=True ), tf.keras.callbacks.EarlyStopping( monitor="val_loss", patience=cfg["patience"], restore_best_weights=True ), tf.keras.callbacks.ReduceLROnPlateau( monitor="val_loss", factor=0.5, patience=3, min_lr=1e-5 ), tf.keras.callbacks.CSVLogger(str(out / "training.csv")), ], verbose=2, shuffle=False, ) (out / "history.json").write_text(json.dumps(history.history, indent=2)) (out / "class_mapping.json").write_text( json.dumps(dict(enumerate(CLASSES)), indent=2) ) (out / "provenance.json").write_text(json.dumps(provenance, indent=2)) (out / "training_config.json").write_text(json.dumps(cfg, indent=2)) versions = { d.metadata["Name"]: d.version for d in importlib.metadata.distributions() } (out / "environment.json").write_text( json.dumps({"packages": versions, **device}, indent=2) ) (out / "training_summary.json").write_text( json.dumps( { "duration_seconds": time.monotonic() - start, "epochs_completed": len(history.epoch), "best_epoch": int(np.argmin(history.history["val_loss"])) + 1, "training_samples": len(y), "parameters": model.count_params(), "model_bytes": (out / "model.keras").stat().st_size, "initialization": "random; no pretrained weights", }, indent=2, ) ) print("Saved real trained model:", out / "model.keras") if __name__ == "__main__": main()