NanoDet-Plus-m-1.5x number of parameters

[onchiptech]

Can someone please explain what the number of parameters used by the NanoDet-Plus-m-1.5x model.
NanoDet-Plus-m-1.5x | 416x416 | 2.44M | 4.7MB(FP16) | 2.3MB(INT8)
The NanoDet-Plus-m-1.5x model benchmark indicates 2.44M parameters but actually, when verified the parameters are more than 2.44M

Here is the verified parameter breakdown estimate

Backbone (ShuffleNetV2 1.5x): ~2.0M
FPN (GhostPAN): ~1.5M
Main Head: ~1.5M
Auxiliary Head: ~3.0M ← Training only
Total: ~8.0M
Deployment model (without aux_head): ~5.0M

So, ~5.0M is still larger than 2.44M benchmark!

Am I missing something here?

Thanks,
Surendra