Agent Skills by xsooi · Find it. Install it. Get to work.
← all skills

Plot Training Log

plot-training-log

Parse a machine-learning training log (.out/.log/.err) and plot epoch-vs-metric line charts (折線圖) with a Python script -- accuracy curves, loss curves, learning-rate schedules, train-vs-validation overfitting checks. Handles `rich`-formatted logs whose metric records are soft-wrapped across many lines with timestamp and `trainer.py:233` columns. Times New Roman by default. Triggers on: "plot the training log", "chart the loss", "graph epoch vs accuracy", "折線圖", "plot loss curve", "visualize training metrics", "make a chart from this .out", "did it overfit", "training curve", "plot val accuracy", "/plot-training-log".

file-parsing #python#parsing#machine-learning#charts 623 words
Includes instructions and supporting files Download skill ZIP View on GitHub ↗ Unzip it into your agent's skills folder.

plot-training-log: training log → line chart

scripts/plot_training_log.py does the whole job. Read the log’s shape first, then drive the script. Don’t hand-write a new parser per log — extend this one.

Workflow

1. Look at the log before parsing. head -40 and tail -40. You are checking two things: which metric keys exist, and whether records are soft-wrapped by rich.

2. Discover the metric keys. Never guess them:

python scripts/plot_training_log.py train.out --list

Prints every key=value metric found and how many records carry it. If it reports no records with 'epoch=' found, retry with --x global_step or --x step. If it still finds nothing, the file probably isn’t a training log (serving/data-prep logs in the same directory often aren’t).

3. Plot. Auto-panels (Accuracy / Loss / Learning rate) need no flags:

python scripts/plot_training_log.py train.out --x1 -o chart.png

Explicit panels when you want control — --panel "TITLE:REGEX", repeatable, one panel per flag:

python scripts/plot_training_log.py train.out --x1 \
  --panel "Validation accuracy:^val/.*acc" \
  --panel "Loss:loss_\d" \
  --title "EAGLE-3 draft model" -o chart.png

4. Read the chart back with the Read tool and check it before reporting. Then state what the curves actually show (best epoch, overfitting gap), not just “chart saved”.

Flags

FlagPurpose
--listprint discovered metric keys, exit
--x KEYx-axis key: epoch (default), global_step, step
--x1shift a 0-based axis to start at 1 (trainers log epoch=0 for epoch 1)
--panel "T:RE"one panel; repeat for more; default auto-detects
--aggfold per-step metrics into one x point: mean (default), last, max, min
--fontdefault Times New Roman, falls back through Times clones then serif
--title, -o, --showfigure title, output path, interactive window

Encoding conventions

The script styles curves so a reader can decode them without reading labels:

  • color = head/index parsed off the key tail (loss_0, loss_1, loss_2)
  • dashed + faded = train, solid + markers = val/eval/test
  • marker shape = metric family (full_acc vs cond_acc) within a panel
  • accuracy panels clamp to ylim 0-1; learning-rate panels go log-scale
  • integer x ticks when all x values are integral (no epoch=2.5)

Gotchas that cost real time

rich soft-wrap. One metrics record spans ~10 physical lines, with a timestamp column left and trainer.py:233 right. Line-by-line parsing gets you fragments. The script strips both columns, drops progress-bar lines, and joins the file into one stream before matching.

epoch= matches inside key names. Validation keys look like val/full_acc_0_epoch=0.537. A naive epoch=\d+ record boundary matches the epoch=0 inside that key, truncating every val record — you get zero val data and a chart that silently shows train only. The (?<!\w) lookbehind is load-bearing. If val curves are missing, check this first.

Aggregation asymmetry. Train metrics are logged per step (thousands of records), val metrics per epoch (one record). --agg mean folds train down to one point per epoch so the two overlay honestly. Don’t compare a per-step train curve against a per-epoch val curve on the same axis without folding.

matplotlib install. System Python on macOS is PEP 668 externally managed — pip install matplotlib fails. Make a venv next to the logs:

python3 -m venv .venv && .venv/bin/pip install matplotlib
.venv/bin/python scripts/plot_training_log.py train.out

The download can take several minutes on a slow link; run it in the background and poll rather than assuming it hung.

Extending

New log format: adjust PREFIX / SUFFIX / PROGRESS in the flattening section. New metric layout: KEY and KV regexes. New default panel family: append to AUTO_PANELS. Keep the script one file with no deps beyond matplotlib.