CNN inference behaves differently on android than on pytorch

vision
2

note that even with arrays/tensors full of 1.0f values, both models still produce different outputs. I also tried normalizing input during training (and android inference)