[Feature introduction] Upload Label Predictions

Apr 3, 2020 | Feature Introduction

New Feature: Upload Label Predictions

 

When talking about machine learning projects it is oftentimes hard to convey the iterative approach that is necessary to achieve good models. The project usually starts by assessing what data is available to you. This data, or parts of it, are then labeled and first networks are trained on it. But the need for new labeled data doesn’t end here, and results from your training already give you clues as to what changes and additions you need in for you data. So equipped with newfound insights your focus shifts back on data collection and labeling.

This cycle of retraining and data collection can repeat itself multiple times within one project. In order to speed up this process,  it is helpful to remember that, through these iterations the Neural Networks are getting increasingly better at the label task itself. So why not use them to pre-label your data and accelerate your label task? Now instead of labeling images from scratch, your labelers have the reduced task of correcting pre-labeled images.

Our new “Import Label” feature supports you with this workflow. Using our API  you can upload your pre-labeled data in no time and save time labeling. See our comprehensive documentation to find out more. This feature is also coming soon within our UI.

Linear way of creating and using training data. Data is uploaded, labeled, exported and used to train the Network.

 

Our pre-label upload unlocks the more iterative approach. Your labeled data helps you train your network which in turn accelerates the label step through the upload of network predictions as pre-label (red).