Regarding tnc and combine_classes and detection head

[AkshatDogra]

Hi,
a.Suppose in my model.train my args are like this:
device=[1,2],
classes=[0, 2, 3, 4, 7, 8, 10, 11],
combine_class=[2, 3, 4],
single_cls=False,

Can you please tell me what should be my:
1.tnc
2.nc_list
3. Value of detection head in model.yaml file.
be?

b. Also if I want to merge multiple classes into 2 subclasses like [2,3,4] into 1 and [7,8] into another, should my val for combine_class be
[[2,3,4],[7,8]] ?

[JiayuanWang-JW]

1. 8
2. I do not know how many detection and segmentation tasks you have and how many heads you used. So please provide more information about what I mentioned.
3. I think you only have one detection head and classes 2, 3, 4 for the detection task, right? If so, you should be set 3.

b. You need to modify the code to implement the function you mentioned. This is an advanced work.

[AkshatDogra]

Hi,
Sorry for the late reply.
I only have 1 detection head and 2 segmentation heads and I want to use [0, 2, 3, 4, 7, 8] for detection. In which I want to merge [2,3,4] into a single class.

[JiayuanWang-JW]

1. 8
2. [6,1,1]
3. 6

combine_class do not effect the training stage. Only effect in val stage. That means, during training, the model will learn 0 2 3 4 7 8 separately, only in val stage, the model will combine 2 3 4 into class 2 to calculate the metrics.

If you set single_cls as True, that is another story.

[AkshatDogra]

Hi,
I tried doing this but I am getting this error.
This is happening whenever I set single_cls to False.
I tried a fresh git clone and created a new env with the steps that you mentioned, double checked the python and the torch version. And tried this on the default ultralytics/train.py script but changed single_cls to false from the default True.