I am working with a custom implementation of MobileNet designed to do binary classification on sparks. After converting to tflite and quantizing it to int8, the model runs on the NPU as expected, but the output does not make sense. The model had good accuracy before quantization, but the int8 outputs seem to be totally random. After testing on some of the training data, frames that should be positive are ranging from -128 to 127, and so are frames that should have been negative. Am I misinterpreting the output, or did something go wrong while converting or quantizing the model?
