NuViz

Terminal-native ML training visualization. Monitor experiments, compare runs, and generate paper-ready tables — all without leaving the terminal.

Why NuViz?

Existing tools (TensorBoard, W&B, Aim) all require a browser. NuViz keeps everything in the terminal where your training is already running — offline, zero network dependency, SSH-friendly out of the box.

Feature	TensorBoard	W&B	Aim	NuViz
Terminal native				Y
Two-line integration	~	~	~	Y
Fully offline	Y		Y	Y
Real-time monitoring	~	Y	Y	Y
Ablation matrix		~		Y
Per-scene breakdown				Y
LaTeX table export				Y
PLY point cloud stats		~		Y

Installation

Python Logger

pip install nuviz

# Optional extras
pip install nuviz[images]    # Pillow for image logging
pip install nuviz[yaml]      # PyYAML for ablation export

Rust CLI

# From crates.io
cargo install nuviz-cli

# From Homebrew (macOS)
brew install LordKnows/tap/nuviz

# Pre-built binaries: see GitHub Releases

Quick Start

1. Log metrics from your training script

from nuviz import Logger

log = Logger("exp-001", project="gaussian_splatting")

for step in range(30000):
    loss = train_step()
    psnr, ssim, lpips = evaluate()
    log.step(step, loss=loss, psnr=psnr, ssim=ssim, lpips=lpips)

    if step % 500 == 0:
        log.image("render", pred_image)

log.finish()

2. Visualize in the terminal

nuviz watch exp-001                       # Live TUI dashboard
nuviz ls --project gaussian_splatting     # List experiments
nuviz leaderboard --sort psnr --top 10    # Ranked table
nuviz compare exp-001 exp-002 --metric psnr  # Curve overlay

3. Generate paper-ready outputs

nuviz leaderboard --format latex --aggregate  # LaTeX table with mean +/- std
nuviz matrix --rows lr --cols depth --metric psnr --format latex
nuviz breakdown exp-001 --latex               # Per-scene table
nuviz export exp-001 --format csv             # Raw data export

Features

Logging (Python)

Metrics — log.step(step, loss=loss, psnr=psnr) with background JSONL writing
Images — log.image("render", tensor) with auto numpy/torch conversion
Point clouds — log.pointcloud("gaussians", xyz, colors) as binary PLY
Per-scene — log.scene("bicycle", psnr=28.4, ssim=0.92)
Ablation configs — Ablation().vary("lr", [1e-3, 1e-4]).generate() with YAML/JSON export
GPU metrics — Automatic nvidia-smi polling (utilization, memory, temperature)
Environment snapshots — Git hash, pip freeze, GPU info, Python version
Anomaly detection — NaN/Inf detection + 3-sigma spike alerts
JSONL rotation — Automatic file rotation at 50MB/500k lines

Visualization (Rust CLI)

Command	Description
`nuviz watch`	Real-time TUI dashboard with braille charts
`nuviz ls`	List all experiments with status
`nuviz leaderboard`	Metric ranking table (table/latex/markdown/csv)
`nuviz compare`	Multi-experiment curve overlay (TUI)
`nuviz matrix`	Ablation hyperparameter matrix
`nuviz breakdown`	Per-scene metric decomposition
`nuviz export`	Export raw data as CSV/JSON
`nuviz image`	Browse experiment images in terminal
`nuviz diff`	Visual diff between experiments
`nuviz view`	Point cloud (PLY) statistics
`nuviz tag`	Tag experiments for organization
`nuviz cleanup`	Remove low-value experiments
`nuviz reproduce`	Print reproduction guide from metadata

Configuration

Env Variable	Default	Description
`NUVIZ_DIR`	`~/.nuviz/experiments`	Base directory for all experiments
`NUVIZ_FLUSH_INTERVAL`	`2.0`	Seconds between background flushes
`NUVIZ_FLUSH_COUNT`	`50`	Records before triggering flush
`NUVIZ_ALERTS`	`1`	Set `0` to disable anomaly alerts
`NUVIZ_SNAPSHOT`	`1`	Set `0` to disable environment capture
`NUVIZ_GPU`	`1`	Set `0` to disable GPU polling
`NUVIZ_GPU_POLL`	`5.0`	GPU polling interval in seconds
`NUVIZ_ROTATE_SIZE`	`50000000`	JSONL rotation size threshold (bytes)
`NUVIZ_ROTATE_LINES`	`500000`	JSONL rotation line threshold

Architecture

Python (training)              Rust (visualization)
┌──────────────┐              ┌──────────────────┐
│   Logger     │──JSONL──────>│  watch / ls /    │
│   + Writer   │  metrics     │  leaderboard ... │
│   + Images   │──PNG────────>│  image / diff    │
│   + PLY      │──PLY────────>│  view            │
│   + Snapshot  │──meta.json──>│  reproduce       │
└──────────────┘              └──────────────────┘

Data flows one way: Python writes structured files, Rust reads them. No server, no network, no database.

Development

See CONTRIBUTING.md for setup instructions.

# Python
cd python && pip install -e ".[dev,images,yaml]"
pytest tests/ --cov=nuviz --cov-report=term-missing

# Rust
cd cli && cargo test --lib
cargo clippy -- -D warnings

Documentation

Usage Guide — Full command reference and API docs
Design Document — Architecture and roadmap

License

MIT

nuviz-cli 0.1.0