Currently micro prefers, what in DSP would be referred to as, fixed point representations of all it’s integer tensors. This is very acceptable for situations where the expected dynamic range is low, i.e. NNs with batch norm, etc. I’m interested in using micro for general audio/other DSP where I’m used to using block floating point.
What are my options here? Any plans to support it/is it supported?
My ultimate goal is to move audio frontend DSP code of an audio pipeline into tf so as to consolidate our DSP(block floating point) and ML(fixed/floating point) frameworks.
Thanks
1 Like
I just saw ’ * Dynamic quantized models support’ on TensorFlow Lite Roadmap
Is this what I am asking for?
1 Like
My ultimate goal is to move audio frontend DSP code of an audio pipeline into tf so as to consolidate our DSP(block floating point) and ML(fixed/floating point) frameworks.
Hi @asj! Looping in @Advait_Jain 
1 Like