Single Class detection

[CloudRider-pixel]

Hi,

I successfully train detr_50 on my custom dataset. Nevertheless, it's a single-class one. What I have to change in order to train this model on a single class problem ? For example in projects/detr_r50 I saw that the number of classes is harcoded to 80 (coco), also loss is CE which normally is not suited.

In advance thanks,

Max.

[rentainhe]

You can train your own custom dataset with one class as follows:

1. Convert your dataset annotation to COCO format
2. Register your own dataset follow: custom.py
3. Modify the model config to model.num_classes=1
4. Modify the criterion config to num_classes=1

[wwhhh2000]

how can I modify model.num_classes and num_classes? I have not found them. In advance thanks.

[rentainhe]

how can I modify model.num_classes and num_classes? I have not found them. In advance thanks.

They have been defined in the model config, you can find them here: https://github.com/IDEA-Research/detrex/blob/f34380d3652c0dbc8f9de911626117943a8208c4/projects/dino_eva/configs/models/dino_r50.py#L76

@wwhhh2000

[wwhhh2000]

how can I modify model.num_classes and num_classes? I have not found them. In advance thanks.

They have been defined in the model config, you can find them here:

https://github.com/IDEA-Research/detrex/blob/f34380d3652c0dbc8f9de911626117943a8208c4/projects/dino_eva/configs/models/dino_r50.py#L76

@wwhhh2000

Thank you so much!