IceVisionDashboard

IceVisionDashboard is an extension to the IceVision object detection framework. Main goal of the library is to support data scientists with there work on object detection problems. This is done by providing different dashboards to provide different steps of the workflow. The dashboards cover investigating datasets, creating new datasets, comparing datasets and analyzing the results of a training.

Test stuff

Example

The example shows how a set of records (here the training_records) can be visualized with the dashboard lib to get some fast insights into the data. The last 3 lines of code are from the dashboard library all the lines before are usual icevision code.

from icevision_dashboards.data import BboxRecordDataset
from icevision_dashboards.dashboards import ObjectDetectionDatasetOverview
# load some data from the icedata
data_dir = icedata.fridge.load_data()
class_map = icedata.fridge.class_map()
parser = icedata.fridge.parser(data_dir)
train_records, valid_records = test_parser.parse()
# create a dataset that can be consumed by the dashboards
train_dash_ds = BboxRecordDataset(train_records, class_map)
# create a new dashboard instance and display it with the .show() function
overview_dashboard = ObjectDetectionDatasetOverview(train_dash_ds, width=1500, height=900)
overview_dashboard.show()

The output will look like this:

An overview of some descriptive statistics for the dataset, the images and the classes.

Some more indepth information about the classes, how they mix (how often they appear at the same time in an image), distribution of annotations per image and a 2D histogram that can be customized.

A gallery with sorting functionality of have a direct look at the images.

Tabular representation of all annotations.

Install

IceVisionDashboard is available as a pip package via PyPi. To install, simply type:

pip install icevision-dashboards

If you are using JupyterLab to view and use your notebooks, a few extra steps are needed. In a terminal, you should type the following:

jupyter labextension install @pyviz/jupyterlab_pyviz

And then in a new cell inside the notebook in which you want to load IceVision dashboards, you should type and execute the following code:

import panel
panel.extension()

Contributing

If you want to contribute add the following lines to your pre-commit file to ensure the notebook cell output don't get pushed into the repo.

# ensure the oupt of the notebooks is empty
jupyter nbconvert --ClearOutputPreprocessor.enabled=True --inplace nbs/*.ipynb
jupyter nbconvert --ClearOutputPreprocessor.enabled=True --inplace examples/*.ipynb
jupyter nbconvert --ClearOutputPreprocessor.enabled=True --inplace test_data_generation/*.ipynb
nbdev_build_lib
git add .