February 17, 2023

NEW FEATURES 🚀

Addition of three new optimizers for classification and tagging trainings : Nadam, Rectified Nadam and Yogi

These optimizers have been benchmarked and the default values for the learning rate correspond to the values resulting from this benchmark work.

When the optimizer is changed, the value of the learning rate changes as well.

Addition of AutoAugment policies for detection trainings

• We now use different algorithms of AutoAugment tailored for detection or for classification/tagging tasks.

• Furthermore, for detection, you can choose two different policies which apply different operations (cf screenshot)

Custom number selection in Search

When using any search interfaces (work orders or assets), it is now possible to specify the number of items to be selected.

New design for homepage

The mobile application homepage has been redesign to offer a smoother experience to users.

IMPROVEMENTS 👍

• Automatic best model selection for detection trainings

• Reduce internal file updates to reduce I/O errors at training time

• Conformity of the project's export format with the new import format

• The ability to export images in their full resolution by posting the "original_images" key

• After he has completed a specific task group, the user is now automatically redirected to this specific task group on the work order page so he does not have to scroll down again

• When a work order is marked as processed, the user is now redirected to the next work order to review

FIXES 🔧

• AutoAugment issue has been fixed

• The issue with review annotations galleries has been resolved

• The issue regarding photos that were not displayed in the Asset management whereas the photo was present in the associated work order is now fixed

• When using the arrows to navigate between assets, the information is now automatically refreshed without the need to do it manually

• The issue on the required photos KPI where the denominator was sometimes higher than the numerator is now fixed

• The fact that the assigned_user_name was not updated in the Search after a review has been corrected

• The issue regarding the association of one task to two task groups when updating the order of tasks is now fixed.