Working on this. This is a problem with resizing. For now, you can resize the image to the onnx trained size before feeding it into the predictor.