The overall mAP is inconsistent with the mAP of each class for custom dataset.

[puffdrum]

Hi, I try to train EfficientDet-d7x on my custom dataset.

Here is my config.yaml:

Here is my command:

Here is my evaluation result during training:

The overall mAP is not low while the AP of each class is 0. How to explain this result, and how to improve the AP of each class?