Celyrix-1.0-Vision / src /train.py
mikexsav's picture
Release CELYRIX v1.0.0
17124c2
Raw History Blame Contribute Delete
3.83 kB
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()