How tflite calculates exp under int8 quantization?

TensorFlow General Discussion
1

I print the result between exp(-128) ~exp(127) in tflite and find that the result is monotonically increasing which makes sense.
I’m curious about how tflite achieves that.
If tflite implementes exp as I find in the repo, it will overflow and cannot be monotonically increasing which I have tested.

3

Hi @Manuzhu,

Apologies for the late reply. The overflow you are observing with exp.h might be due to using extremely large input values. As tflite model is optimized for specific edge devices based on the optimization, the exp.h will receive the quantized input. Now TFLite is moved to google-ai-edge repo as LiteRT. For further assistance, please share the working code or can raise an issue on LiteRT repo.

Thank You