tflite quantized model producing different results on CPU / NNAPI+NPU about tflite-support HOT 3 CLOSED

Hi There,

Thanks for reporting the issue. Task library calls TFLite underneath for the invocation on CPU v.s. NNAPI.

If the CPU results are correct (the same as TF in training), then it might be an issue with NNAPI or vendor (NPU chip vendor) implementation.

If the CPU results are incorrect, it might be an issue with TFLite implementation.

In either cases, please raise the issue in https://github.com/tensorflow/tensorflow so that the TFLite team can have a look :-) Please attach the model, input data, expected output and actual output data, together with some specs on the actual devices that you're using.

Thanks,
Tiezhen

from tflite-support.

@FSet89 can you please also provide the code snippet of how your inference is done? I guess the Support Library is used, right? As Tiezhen said, it will be great if you know which inference is correct, CPU or NNAPI.

from tflite-support.

Also discovered CPU / GPU difference in #534 (comment).

from tflite-support.