ltui 0.1.1

Terminal UI for PyTorch Lightning CSVLogger metrics.

ltui

TensorBoard-like live monitoring for PyTorch Lightning CSVLogger metrics, directly in the terminal.

ltui turns Lightning metrics.csv files into a focused terminal plot for live monitoring, inspection, and run comparison, similar in spirit to tools like TensorBoard. Because it runs entirely in the terminal, it is especially useful for headless monitoring on remote machines over SSH, inside tmux, and without a browser or server process.

Showcase

Live loss monitoring: train and validation loss in one terminal plot, with log scaling enabled.	Run comparison: multiple Lightning versions overlaid on an epoch axis with separate run and train/val legends.
Metric selection: fuzzy, keyboard-first selection over grouped train/validation metric families.	Config inspection: inspect the YAML config associated with a selected Lightning run without leaving the terminal.

Highlights

• Live in-terminal visualization of PyTorch Lightning training metrics
• Automatic train/validation metric grouping in one plot
• Multi-run comparison with readable run legends
• Step and epoch x-axis modes with Lightning-friendly alignment
• Log scaling and EMA smoothing for noisy or wide-range metrics
• YAML config inspection for runs with associated model/training configs
• Keyboard-first selectors with fuzzy search

Installation

pip install ltui

Quick Start

Point ltui at any directory containing Lightning CSV logs:

ltui /path/to/log/root

Examples:

ltui ./lightning_logs
ltui /data/experiments/my_model
ltui ~/runs/my_experiment

On startup, ltui recursively discovers metrics.csv files, selects the latest modified run, and chooses a loss metric when one is available.

What It Reads

ltui targets PyTorch Lightning CSVLogger output. Each discovered metrics.csv is treated as one selectable run/version.

Supported layouts include:

lightning_logs/version_0/metrics.csv
run_a/version_0/metrics.csv
run_a/lightning_logs/version_0/metrics.csv
experiments/group_1/run_a/lightning_logs/version_3/metrics.csv

Display names are derived from paths relative to the scanned root.

Metric Grouping

Train/validation grouping is enabled by default. Metrics are grouped when they use the standard prefixes:

train_
val_

For example:

train_loss + val_loss -> loss
train_recon_loss + val_recon_loss -> recon_loss
train_kl + val_kl -> kl

Lightning step/epoch suffixes are handled as part of the same family, so train_loss_step, train_loss_epoch, and val_loss appear as loss.

When the x-axis is step, validation epoch metrics use the step value from their CSV row. This places validation points at the training step where validation was logged.

Controls

Key	Action
m	Open metric selector
r	Open run/version selector
c	Open config viewer for runs with a unique YAML config
/	Fuzzy search inside selector
arrow keys	Navigate selector
space	Toggle selection in selector, pause/resume on main screen
enter	Apply selector
n	Next selected metric/family
p	Previous selected metric/family
a	Toggle x-axis between step and epoch
d	Toggle dark/light plot theme
s	Toggle smoothing
x	Toggle log-x
y	Toggle log-y
R	Force rescan
q	Quit

Plot Behavior

• Preferred x-axis: step
• Fallback x-axis: epoch
• Final fallback: row index
• Step and epoch axes start at 0
• Numeric metric columns are plotted, excluding step and epoch
• Rows with NaN for the selected metric are dropped
• Points are sorted by x-axis before plotting
• Smoothing is EMA with alpha 0.2
• Log scaling drops nonpositive points and reports the count
• UI state is stored per root under ~/.local/state/ltui/