Styling and Configuration

Styling and Configuration#

Malet provides three levels of styling control: CLI flags for quick adjustments, YAML config files for reusable presets, and the Python API for full matplotlib access.

CLI styling flags#

Flag	Default	Description
`-colors`	`'default'`	Colormap or palette name(s)
`-annotate` / `-noannotate`	`True`	Toggle data point annotations
`-annotate_field`	—	Additional fields to show in annotations
`-fig_size`	—	Figure size: one number (square) or two (width height)
`-style`	`'default'`	Matplotlib style (e.g., `'ggplot'`, `'seaborn'`)
`-xscale`	—	X-axis scale: `linear`, `log`, or `unif`
`-yscale`	—	Y-axis scale: `linear` or `log`
`-zscale`	—	Colorbar scale: `linear` or `log`
`-title`	—	Custom plot title
`-xlabel`, `-ylabel`, `-zlabel`	—	Custom axis labels
`-font_size`	`22`	Font size for title and labels
`-marker_size`	`10`	Marker size

The special x-scale value unif (uniform) converts the x-axis to evenly spaced ticks regardless of actual numeric values. This is useful when x-axis values are irregularly spaced (e.g., [0.001, 0.01, 0.1, 1.0]).

Colors#

The -colors flag accepts palette names, colormap names, or specific hex codes:

-colors default              # seaborn default palette
-colors Blues                 # matplotlib colormap
-colors 'light:#9467bd'      # named palette with specific hue
-colors 'default Blues'       # multiple palettes (cycled)

Fine-tune color sampling with -colors_rep_skip_shift [repeat skip shift]:

repeat — number of times to repeat each color
skip — skip every Nth color from the palette
shift — offset into the palette

YAML plot config#

For reusable styling, create a YAML file and pass it via -plot_config:

malet-plot -exp_folder ./exp -plot_config style.yaml -mode curve-epoch-val_accuracy

Config structure#

# Default style applied to all plots
default_style:
  annotate: false
  std_plot: fill
  line_style:
    linewidth: 4
    marker: 'D'
    markersize: 10
  ax_style:
    frame_width: 2.5
    fig_size: 7
    legend: [{fontsize: 20}]
    grid: [true, {linestyle: '--'}]
    tick_params:
      - axis: both
        which: major
        labelsize: 25
        direction: in
        length: 5

# Per-mode overrides
'curve-epoch-val_accuracy':
  annotate: true
  ax_style:
    title: ['Validation Accuracy', {size: 27}]

`line_style`#

Controls matplotlib line properties:

Key	Type	Description
`linewidth`	float	Line width
`marker`	str	Marker style
`markersize`	float	Marker size
`markevery`	int	Show marker every N points
`color`	str	Line color
`linestyle`	str	Line style (`-`, `--`, `-.`, `:`)

`ax_style`#

Controls figure and axes properties. Most keys map directly to matplotlib.axes.Axes setter methods. The value format is [*positional_args, {**keyword_args}]:

ax_style:
  title: ['My Title', {size: 27}]
  xlabel: ['Epoch', {fontsize: 20}]
  yscale: ['log']
  ylim: [0, 1]

Two special keys are provided for convenience:

Key	Description
`fig_size`	Figure width (and height if square)
`frame_width`	Width of the axes border spines

Mode aliases#

Define reusable plot configurations with custom names:

'sam_comparison':
  mode: curve-rho-val_accuracy
  multi_line_fields: optimizer
  filter: 'optimizer sgd sam'
  annotate: true
  colors: ''
  std_plot: bar
  ax_style:
    title: ['SGD vs SAM', {size: 27}]
    xlabel: ['$\rho$', {size: 30}]
    ylabel: ['Val Accuracy (%)', {size: 30}]

Invoke with:

malet-plot -exp_folder ./exp -plot_config style.yaml -mode sam_comparison

When using a mode alias, any conflicting CLI arguments are overridden by the alias configuration.

Style hierarchy#

When styles conflict, higher-priority sources take precedence:

default_style  <  CLI flags  <  per-mode config  <  mode alias

Field ordering#

Control the order of categorical values in plots with -field_orders:

-field_orders 'optimizer sgd adam / lr 0.1 0.01 0.001'

Values appear in the specified order. Malet validates that all specified values exist in the data.

Best-reference arguments#

By default, each x-axis value gets its own optimal hyperparameter set. These arguments let you fix the reference point for optimization:

-best_ref_x_fields {value}: Choose the best hyperparameters at a specific x value, then apply them to all x values. For step-based x-axes, defaults to last.
-best_ref_ml_fields {value}: Use the same hyperparameters across all multi-line field values, chosen based on a specific value.
-best_ref_metric_field {metric}: Optimize hyperparameters based on one metric, then plot a different metric with those hyperparameters.

Custom preprocessing#

For advanced use cases, pass a custom preprocessing function via the preprcs_df argument in the Python API:

def preprcs_df(p_df, legend, mlvs):
    # Modify the DataFrame, legend string, or multi-line values
    return p_df, legend, mlvs

This runs after avgbest_df and before drawing. Useful for custom legend formatting or additional row filtering.