Open BaofengZan opened 2 weeks ago
Hi.
You need to understand the continual segmentation settings.
CONT.BASE_CLS
means the number of classes at step 0, termed base classes. The total number of classes for COCO-Panoptic dataset is 133. We split the 133 classes into 83 and 50 classes. CONT.BASE_CLS = 83
means that we set the number of base classes for COCO as 83.
CONT.INC_CLS
means the number of incremented classes. CONT.INC_CLS = 5
means that we repeatedly guide 5 new classes to the model and repeat this process 10 times (until the number of total incremented classes becomes 50).
base_queries
is a hyperparameter, and we set it to the number of base classes: base_queries = CONT.BASE_CLS
.
num_prompts
is also a hyperparameter, num_prompts = CONT.INC_CLS
but the minimum value of num_prompts is set to 10.
BASE_CLS
is set to the number of classes at step 0 and INC_CLS
is set to 10.
Firstly, I would like to thank you for your work. Regarding the code, I have some doubts. What do these variables represent in the training script?
If I have a set of data with 10 categories. I have already trained step 0. When I proceed to train step 1, it's still these 10 categories but with different images. Should I set BASE_CLS equal to INC_CLS, both as 10?