nuviz 0.1.0

Terminal-native ML training visualization

NuViz

Terminal-native ML training visualization — the Python logging library.

Install

pip install nuviz

# Optional: image logging (Pillow) and ablation YAML export
pip install nuviz[images,yaml]

Usage

from nuviz import Logger

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

for step in range(10000):
    loss = train_step()
    log.step(step, loss=loss, psnr=psnr)

log.finish()

Images and Point Clouds

log.image("render", predicted_image)         # numpy/torch tensor -> PNG
log.pointcloud("gaussians", xyz, colors)     # -> binary PLY
log.scene("bicycle", psnr=28.4, ssim=0.92)  # per-scene metrics

Ablation Experiments

from nuviz import Ablation

sweep = Ablation("lr_sweep")
sweep.vary("lr", [1e-3, 1e-4, 1e-5])
sweep.toggle("use_augmentation")

for config in sweep.generate():
    log = Logger(config.name, seed=42, config_hash=config.hash)
    # ... train with config.params ...

GPU Metrics

GPU utilization, memory, and temperature are collected automatically via nvidia-smi polling. Disable with NUVIZ_GPU=0.

Visualize

Install the Rust CLI to visualize logged data:

cargo install nuviz-cli

nuviz watch exp-001           # Live dashboard
nuviz leaderboard --sort psnr # Ranked table
nuviz compare exp-001 exp-002 # Curve overlay

Configuration

All settings can be overridden via environment variables:

Variable	Default	Description
NUVIZ_DIR	~/.nuviz/experiments	Data directory
NUVIZ_GPU	1	Set 0 to disable GPU collection
NUVIZ_GPU_POLL	5.0	GPU poll interval (seconds)
NUVIZ_ALERTS	1	Set 0 to disable anomaly alerts

Requirements

• Python 3.10+
• No required dependencies (numpy, Pillow, PyYAML are optional)

License

MIT