Hi,
Since Nano only has 4G memory, this will limit you from using too complicated model.
Swap memory can only increase CPU based memory, but in general inference requires GPU-accessible memory.
A possible alternative is to use other frameworks that requires less memory usage, ex. TensorRT.
Thanks.