optiviz 0.1.3

Educational tool for visualizing PyTorch-compatible optimisers on any differentiable 1D or 2D function.

OptiViz

OptiViz enables effortless visualisation of the optimisation sequence of any PyTorch optimiser on any differentiable function in one or two variables. OptiViz might find educational use in an introductory nonlinear optimisation or deep learning class.

Installation

To install OptiViz, please use:

pip install optiviz

Usage

All functionality of OptiViz is exposed through the optiviz.optimise function.

import torch
from optiviz import optimise

Any optimisation problem has an objective function. OptiViz works with differentiable, real-valued objective functions in one or two variables.

f : \mathbb{R} \rightarrow \mathbb{R}

g : \mathbb{R}^2 \rightarrow \mathbb{R}

In code, every input and output to the objective function must be a torch.Tensor of shape (1,)

def f(x: torch.Tensor) -> torch.Tensor:
    """
    Example of an objective function in one variable.
    """
    return x ** 2
def g(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    """
    Example of an objective function in two variables.
    """
    return x ** 2 + y ** 2 + x.sin() * y.sin()

The optiviz.optimise function (please see docstring for complete usage) is used to visualise the optimisation sequence of the objective function using a PyTorch optimiser.

arg_g_min = optimise(
        g, # objective function
        (12.5, 12.5), # initial values of the parameters being adjusted
        plot_centre=(0, 0),
        plot_boundary=25,
        iters=100,
        optimiser=torch.optim.Adam, # PyTorch-compatible optimiser
        lr=5e-1 # any keyword arguments for the optimiser
    )