I am trying to use an audio classification model trained in Pytorch with the Neural Compute Stick 2. The model is a ResNet18. While comparing the accuracy of the network after the model optimizer step (classifying on the stick) with the original model, I noticed a very large performance drop (in the order of 30% accuracy loss). Preprocessing is exactly the same for the two models. Is this expected? Is there any way to mitigate this phenomenon?
Thank you
