How to decide on the right target size of an input image for inception_v3? (predictions differ)

Question

I am running an Inception V3 model to make predictions. I am finding out that different target sizes for the input image lead to different predictions.

For the same image