I want to change the input size of mobilenet v2 as the backbone of object detector, like 320x320, 448 x448, and fine tune it, I wonder if it is possible? Will it hugely affect the performance?
It it totally possible to use those new input shapes, because the forward method calls a adaptive_average_pool function before the Linear layer is called and therefor there should be no shape issues!
The performance will be a bit slower, because there are more dot products being calculated, but it should not be magnificent!
Thank you for your answer!