the _preprocess function resizes all inputs to be 640 by 640. if you give it something larger, it’ll take it. the output might be that size, but you can resize that to fit your input. yes, that means it won’t be “pixel-accurate”. it probably isn’t even if the input is sized exactly as the model requires.
try changing the occurences of 640. try doubling that, or halving. not all arbitrary numbers might work, only some specific sizes. if nothing works, the model is probably not fully convolutional.