simonblanke / progressboard

Add on for Hyperactive package to visualize progress of optimization run.

Progress Board

An addon for the Hyperactive package to visualize the progress of optimization runs.

The Progress Board is a dashboard (opens in webbrowser) that provides visualization of live-updated data from Hyperactive. It integrates seamlessly with Hyperactive (v4) and opens up the optimization run with useful information. It also supports multiprocessing and multiple searches at the same time without any added complexity or work for the user.

The Progress Board should be used for computationally expensive objective functions (like machine-/deep-learning models).

The Progress Board is tested in Ubuntu, but Windows support maybe added in the future.

State of project

This project is in an early development stage. If you encounter a problem it would be very helpful to open an issue and describe it in detail.

Installation

pip install hyperactive-progress-board

Example

from sklearn.model_selection import cross_val_score
from sklearn.tree import DecisionTreeRegressor
from sklearn.datasets import fetch_california_housing

from hyperactive import Hyperactive
from hyperactive_progress_board import ProgressBoard # import progress board

data = fetch_california_housing()
X, y = data.data, data.target

progress = ProgressBoard() # init progress board


@progress.update # add decorator
def dtr_model(opt):
    dtr = DecisionTreeRegressor(
        min_samples_split=opt["min_samples_split"],
    )
    scores = cross_val_score(dtr, X, y, cv=5)
    return scores.mean()


search_space = {
    "max_depth": list(range(2, 50)),
    "min_samples_split": list(range(2, 50)),
    "min_samples_leaf": list(range(1, 50)),
}

progress.open() # open progress board before run begins

hyper = Hyperactive()
hyper.add_search(dtr_model, search_space, n_iter=1000)
hyper.run()

FAQ

Command line opens and closes immediately

This happens because of the command line of a previous run of the progress-board is still running. Close the command-line from the previous run to start a new one.

progressboard's Issues

Current state of progress-board

After coming back to this project after some time I feel, the source-code is still in an experimental phase:

• Most of the code-base is not tested
• The installation via pip seems unreliable. Conda is needed for reliable usage.

I will look into a better way to run the progress-board in parallel to the optimization run and how to test this GUI via selenium. I am also thinking of using plotly-dash instead of panel in the future, because panel seems very tied to conda.