page_03_plot_curve.py

import argparse
import json
from pathlib import Path

import matplotlib.pyplot as plt

from trainer_utils import get_best_checkpoint


def parse_args():
    parser = argparse.ArgumentParser(description="Plot learning_curve.jsonl")
    parser.add_argument("--run-dir", required=True)
    return parser.parse_args()


def load_curve(path):
    rows = []
    with Path(path).open(encoding="utf-8") as file:
        for line in file:
            rows.append(json.loads(line))

    by_step = {}
    for row in rows:
        step = row.get("step")
        if step is not None:
            by_step[step] = {**by_step.get(step, {}), **row}
    return [by_step[step] for step in sorted(by_step)]


def plot_metric(axis, rows, key, label, color):
    selected = [row for row in rows if key in row and row.get("step") is not None]
    if selected:
        axis.plot(
            [row["step"] for row in selected],
            [row[key] for row in selected],
            label=label,
            color=color,
        )
    axis.set_ylabel(label)
    axis.legend()
    axis.grid(True)


def main():
    args = parse_args()
    run_dir = Path(args.run_dir)
    rows = load_curve(run_dir / "learning_curve.jsonl")
    _, best_step = get_best_checkpoint(run_dir)

    fig, axes = plt.subplots(4, 1, figsize=(10, 12), sharex=True)
    plot_metric(axes[0], rows, "loss", "train_loss", "tab:blue")
    plot_metric(axes[1], rows, "eval_accuracy", "eval_accuracy", "tab:orange")
    plot_metric(axes[2], rows, "eval_precision", "eval_precision", "tab:green")
    plot_metric(axes[3], rows, "eval_f1", "eval_f1", "tab:red")

    if best_step is not None:
        for axis in axes:
            axis.axvline(best_step, linestyle="--", color="black", label="best")

    axes[-1].set_xlabel("step")
    fig.tight_layout()
    out_path = run_dir / "learning_curve.png"
    fig.savefig(out_path, dpi=150)
    print("saved:", out_path)


if __name__ == "__main__":
    main()