Plot gallery

A visual overview of what Malet can produce. All figures below were generated from real CIFAR-10 ResNet20 experiment logs. Click any heading to jump to the corresponding guide page.

Curve plots

Training and validation curves with error bands, comparing methods or sweeping hyperparameters.

ADMM vs SAFE across sparsity levels

Training dynamics: masked val accuracy over epochs

Sequential blue colormap for ordered sparsity levels

Log-log curve: penalty vs constraint distance

Scatter plots

Metric-vs-metric relationships with multi-line encoding (color + marker shape).

Color = lambda, marker = schedule (dual multi-line)

Marker = lambda, color = proj dev (continuous colorbar)

Multi-plot grids

Subplot grids split by one or two fields.

4x4 grid by lambda (columns) x sparsity schedule (rows), colored by proj dev

3x4 grid by lambda x lambda schedule

Multi-line

Multiple curves on the same axes. Ordered numeric fields get sequential colormaps automatically.

Tab10 palette: ADMM, SAFE, SAM, SGD at sparsity 0.9

Method comparison across sparsity levels

Animation (GIF)

Animate any plot type over time. The x-axis shows hyperparameters; the animation field (epoch/step) drives the frames.

Scatter animated over steps: color = lambda, marker = lr

Faceted scatter animated over late training

Other visualizations

Loss landscapes and custom analysis plots.

ADMM loss landscape (sharpness: 0.2)

SAFE loss landscape (sharpness: 0.09 — flatter minimum)

Lambda penalty schedules: constant, cosine, linear

Dense/sparse/BNT accuracy vs lambda across sparsities