how many unique classes can we train?

[stephanecharette]

The most classes I've ever had to use so far in a neural network is 22. How many classes can we have with Darknet's .cfg files? Dozens? Hundreds? Thousands?

@AlexeyAB do you know if there is a limit, or is it only constrained by the quality of the images, markup, training iterations, etc?

If there is a limit, is it the same for each type of .cfg files?

Do you know, or can you give a ballpark idea of what kinds of class limits we'd have with these .cfg files?

• csresnext50-panet-spp-original-optimal.cfg
• yolov3.cfg
• yolov3_5l.cfg
• yolov3-spp.cfg
• yolov3-tiny.cfg
• yolov3-tiny_3l.cfg

[AlexeyAB]

There is no strict limit of number of classes for models. The more classes - the worse accuracy AP.

For example:

• yolov3.cfg/weights - is trained for 80 classes (MS COCO)
• yolov3-openimages.cfg/weights - is trained for 600 classes (OpenImages)
• yolo9000.cfg/weights - is trained for 9000 classes (MS COCO + ImageNet)