Very low accuracy for tflite models with tvm on raspberry pi

Application
#1

Hi,

I am trying to run a tflite model (my own trained tflite model and not the mobilenet as in the tutorial) by referring to the tflite tutorial(https://tvm.apache.org/docs/tutorials/frontend/from_tflite.html on raspberry pi 3, 32 bit OS, with llvm 6 and python 3.7.

I have just changed the image preprocessing part according to my requirement in the tutorial and the target for tvm(tvm.target.arm_cpu("rasp3b")). With tflite I was able to get 91% accuracy but with the tflite model compiled with tvm I am getting 62.3% accuracy.

Any suggestion as to why this happened? Thanks!

#2

Also, forgot to add that the tflite model was a fully integer quantised int8 model. So to sum up, I was trying to use a pre quantised int8 tflite model on raspberry pi, and compare the accuracy of tvm and tflite.

Any idea, why there is such a huge accuracy drop?