Integrate VNR model

[mbanth]

Add the VNR model into its destination repository. (AV2b)

[shuchitak]

@andrewdewhurst, could you explain this a bit more. Currently the tensorflow model is in py_vnr repo. The corresponding tflite model and the xcore optimised version of it is in the sw_avona repo as part of the sw_avona lib_vnr module. This makes the most sense to me. Could you let me know, which, if any of the models you want to move to xcore sdk repo? Thanks.

[mbanth]

We will discuss this issue at the 27 April 2022 Voice Pipeline Performance meeting to gain clarity about its scope and necessity.

[shuchitak]

As discussed in the 27 April 2022 Voice Pipeline Performance, py_vnr repo will contain the tf and tflite models, documentation regarding training, tf -> tflite conversion etc. accessible to anyone within XMOS.

[mbanth]

@andrewdewhurst, is this issue still unresolved in your mind?

[mbanth]

@shuchitak, can you please provide information about the state of the documentation in this repository that covers the use of the xformer?

[shuchitak]

By way of documentation, we currently have this python script to which you provide a tflite model as argument and it will convert it to tflite micro using the xformer, generate the tensor arena scratch size needed and model quant spec related defines and optionally copy it to the VNR module so it can run with the new model.

I tried documenting this but it makes much more sense to have this in a script. What's pending is:

• Add explanatory comments to the script.
• Add a readme documenting how to run the script and expected output.
• Move it to a different directory (possible sw_avona/tools, I haven't decided) from test_wav_vnr where it currently is.

[shuchitak]

VNR model is integrated to lib_vnr and documented. All changes merged to develop.