plot_loss_from_trainer_state.py

# -*- coding: utf-8 -*-
"""
Usage:
  python plot_loss_from_trainer_state.py --input trainer_state.json --outdir ./plots \
         --checkpoint_steps 263,526,789,1052

功能:
- Curve: 黃橘色實線
- Grid: x,y 虛線
- Epoch markers: 藍色虛線 + EpochN 標籤（含最後一個 epoch）
- Checkpoints: 藍色小圓點（線性插值；超出範圍時使用端點值，並自動擴張 x 軸確保能看見）
"""
import json, argparse
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np

YELLOW_ORANGE = "#d58f00"
BLUE = "#1f77b4"

def find_epoch_boundaries(log_items):
    """找到每個 epoch 邊界 (包含最後一個)"""
    boundaries = []
    prev_epoch_int = None
    seen = set()
    last_step, last_epoch = None, None
    for it in log_items:
        step = it.get("step")
        ep = it.get("epoch")
        if step is None or ep is None:
            continue
        last_step, last_epoch = step, ep
        ep_int = int(ep)
        if prev_epoch_int is None:
            prev_epoch_int = ep_int
            continue
        if ep_int != prev_epoch_int:
            if (step, ep_int) not in seen and ep_int >= 1:
                boundaries.append((step, ep_int))
                seen.add((step, ep_int))
            prev_epoch_int = ep_int
    # 最後一個 epoch 也補上
    if last_step is not None and last_epoch is not None:
        ep_final = int(float(last_epoch)) + 1
        if (last_step, ep_final) not in seen:
            boundaries.append((last_step, ep_final))
    boundaries.sort(key=lambda x: x[0])
    return boundaries

def plot_series(x, y, xlabel, ylabel, title, outpath,
                epoch_marks=None, checkpoint_steps=None,
                color=YELLOW_ORANGE, linestyle='-'):
    fig = plt.figure(figsize=(10,6))
    ax = fig.add_subplot(111)
    ax.plot(x, y, color=color, linestyle=linestyle, linewidth=2)

    # 標記 checkpoint 藍點（線性插值；邊界外使用端點值）
    extra_x = []
    if checkpoint_steps:
        for s in checkpoint_steps:
            y_interp = np.interp(s, x, y, left=y[0], right=y[-1])
            ax.plot(s, y_interp, marker='o', color=BLUE, markersize=6)
            extra_x.append(s)

    # === 計算 x 範圍時把 epoch 標線也納入，並加右側 padding ===
    xmin = 0
    all_x_candidates = [max(x)]
    if extra_x:
        all_x_candidates.append(max(extra_x))
    if epoch_marks:
        # 把所有 epoch 標線的 step 納入考量
        ep_steps = [s for (s, _) in epoch_marks]
        if ep_steps:
            all_x_candidates.append(max(ep_steps))

    xmax_base = max(all_x_candidates) if all_x_candidates else x[-1]

    # 右邊加一點 margin，避免剛好貼齊看不到線
    span = max(xmax_base - xmin, 1.0)
    right_pad = max(1.0, 0.02 * span)   # 至少 +1 step 或 2% 寬度
    ax.set_xlim(left=xmin, right=xmax_base + right_pad)

    # y 仍從 0 起
    ax.set_ylim(bottom=0)

    # 虛線格線
    ax.grid(True, which='major', axis='both', linestyle='--', linewidth=0.8, alpha=0.6)

    # epoch 標記 (藍色虛線)
    if epoch_marks:
        for step, ep in epoch_marks:
            ax.axvline(x=step, color=BLUE, linestyle='--', linewidth=1.2)
            ymax = ax.get_ylim()[1]
            ax.text(step, ymax*0.98, f'Epoch{ep}', rotation=90,
                    va='top', ha='right', fontsize=8, color=BLUE)

    # label & look（放到最後避免被 set_xlim/set_ylim 影響）
    ax.set_xlabel(xlabel); ax.set_ylabel(ylabel); ax.set_title(title)
    ax.spines['left'].set_linewidth(2); ax.spines['bottom'].set_linewidth(2)
    ax.spines['right'].set_visible(False); ax.spines['top'].set_visible(False)

    fig.savefig(outpath, bbox_inches="tight")
    plt.close(fig)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--input", required=True, help="Path to trainer_state.json")
    ap.add_argument("--outdir", default="./plots", help="Directory to save PNGs")
    ap.add_argument("--no_epoch_marks", action="store_true", help="Disable vertical epoch markers")
    ap.add_argument("--checkpoint_steps", default="", help="Comma-separated steps (e.g., 100,200,500)")
    args = ap.parse_args()

    src = Path(args.input)
    with open(src, "r", encoding="utf-8") as f:
        state = json.load(f)

    log = state.get("log_history", state.get("logs", []))

    steps, train_losses = [], []
    eval_steps, eval_losses = [], []
    lr_steps, lrs = [], []

    for item in log:
        step = item.get("step")
        if step is None: 
            continue
        if "loss" in item:
            steps.append(step); train_losses.append(item["loss"])
        if "eval_loss" in item:
            eval_steps.append(step); eval_losses.append(item["eval_loss"])
        if "learning_rate" in item:
            lr_steps.append(step); lrs.append(item["learning_rate"])

    outdir = Path(args.outdir); outdir.mkdir(parents=True, exist_ok=True)

    epoch_marks = None if args.no_epoch_marks else find_epoch_boundaries(log)
    # 允許空白與混合格式
    raw = [s.strip() for s in args.checkpoint_steps.replace("，", ",").split(",") if s.strip()]
    checkpoint_steps = []
    for s in raw:
        try:
            checkpoint_steps.append(int(float(s)))
        except:
            pass

    if steps and train_losses:
        plot_series(steps, train_losses, "Step", "Training Loss", "Training Loss vs Step",
                    outdir / "loss_curve.png", epoch_marks=epoch_marks, checkpoint_steps=checkpoint_steps)
    if eval_steps and eval_losses:
        plot_series(eval_steps, eval_losses, "Step", "Eval Loss", "Eval Loss vs Step",
                    outdir / "eval_loss_curve.png", epoch_marks=epoch_marks, checkpoint_steps=checkpoint_steps)
    if lr_steps and lrs:
        plot_series(lr_steps, lrs, "Step", "Learning Rate", "Learning Rate vs Step",
                    outdir / "lr_curve.png", epoch_marks=epoch_marks, checkpoint_steps=checkpoint_steps)

    print(f"Saved plots to: {outdir.resolve()}")

if __name__ == "__main__":
    main()