search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.57k stars 7.95k forks source link

mosaic, cutmix, and mixup #4638

Open stephanecharette opened 4 years ago

stephanecharette commented 4 years ago

In reading through recent issues (was trying to find information on angle=... prior to issue #4626) I see references to 3 new [net] parameters that aren't documented on the wiki and don't show up in the template .cfg files:

mosaic=...
cutmix=...
mixup=...

(When I set cutmix=1, then darknet segfaults, but because I don't know what these do, I've left it at zero.)

Can we get these added to the wiki? When should we use them, or what is a scenario where our network is a good candidate to try changing them from 0 to 1?

AlexeyAB commented 4 years ago

These are new features, which should be tested more, before it will be added to the wiki: https://github.com/AlexeyAB/darknet/projects/1

Use mosaic=1 for training Detector and Classifier.

stephanecharette commented 3 years ago

@AlexeyAB I'd like to update my augmentation page (https://www.ccoderun.ca/darkmark/DataAugmentationMisc.html#Mixup) with new information. Can you provide more information on mosaic, cutmix, and mixup? For example, I don't see mixup listed in the wiki (https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section).

Should we be using these? When do we use them? What are the possible values?

stephanecharette commented 3 years ago

See issue #7114 for additional information.