Skip to contents

Getting Started

vistool.Rmd

What is vistool?

vistool helps visualize optimization concepts and model predictions: - Loss functions (regression/classification) - Objective functions (1D/2D test functions and custom objectives) - Optimization traces (GD, Momentum, Nesterov) - Model prediction surfaces and decision boundaries

You can render either with ggplot2 (2D) or interactively with plotly (3D surfaces).

Installation

Install the development version from GitHub:

# install.packages("pak")
pak::pak("slds-lmu/vistool")

Plotly backend and static export

  • Interactive surfaces use plotly.
  • Saving plotly surfaces as static images uses plotly::save_image(), which relies on the Python package kaleido (a separate Chrome/Chromium installation is required).
  • On first use of save/export, vistool triggers reticulate::py_require("kaleido") to provision a cached Python environment automatically.

If you prefer to manage Python yourself, ensure kaleido is installed and that reticulate points to your Python:

# optional manual setup
install.packages("reticulate")
# Sys.setenv(RETICULATE_PYTHON = "/path/to/python")
reticulate::py_install("kaleido")

Quick tour

The workflow is consistent across plot types:

  1. Create a visualizer with as_visualizer(...) (computational setup)
  2. Add optional layers via add_*()
  3. Render with plot(...) (and optionally save(...))

1) Visualizers 

# Loss functions
vis_loss = as_visualizer(lss("regr.mse"), y_pred = seq(-3, 3), y_true = 0)

# Objectives (1D/2D)
vis_obj2d = as_visualizer(obj("TF_branin"))                 # ggplot2
vis_surf  = as_visualizer(obj("TF_branin"), type = "surface")  # plotly

# Models (mlr3 Task + Learner)
library(mlr3); library(mlr3learners)
task = tsk("pima")$select(c("insulin", "mass"))
learner = lrn("classif.svm", predict_type = "prob")
vis_model = as_visualizer(task, learner = learner)

2) Layers 

vis_model$add_boundary(values = c(0.3, 0.5, 0.7))
vis_model$add_training_data()
vis_surf$add_contours()

3) Plot (and Save) 

# ggplot2 (2D)
vis_model$plot(show_title = FALSE)

# plotly surface (3D or flattened)
vis_surf$plot()                    # interactive surface
vis_surf$plot(flatten = TRUE)      # 2D contour

# Theme overrides per render
vis_model$plot(theme = list(palette = "plasma", text_size = 13))

# Optional: save the last rendered plot
vis_model$save("model.png", width = 8, height = 6)

Explore specific topics:

Styling and composition: - Customization Guide - Advanced Visualization