Hi, I am converting an int8 tflite model using tvm. My target device is raspberry pi 4, and I am using llvm 6.0 and python 3.7. My original accuracy( using only tflite model) was 97.5%. But using TVM, my accuracy is dropping to 61.5%, on the same test dataset. I followed this tutorial https://tvm.apache.org/docs/tutorials/frontend/from_tflite.html
I double checked everything: the tflite model, the target device, my tvm version, my llvm version.
I tried to follow what TVM does with prequantised model: it does some requantisation and dequantisation in the conversion of conv2d operator. But ideally it does not affect the accuracy in such a major way. Any suggestions regarding this? Ideally there should be not be such a major accuracy drop.