Partial Usage of Edge Impulse

Help
#1

Hi Edge Impulse Team,

I have a model that I have trained not in the Edge Impulse ecosystem (it is a keras / tensorflow model). How can I use the following Edge Impulse functionality with my model:

  • i) Optimization (e. g. quantization)
  • ii) Automated C++ Linux library generation

Regards

Lukas

#2

Hi @Lukas we don’t support importing pretrained models, but you can replicate the neural network architecture in a Classification block (click the three dots => show expert mode) and retrain from there to get all the features.

#3

@janjongboom thanks for your fast response. Is it planned to offer such a feature? Is it technically not possible?

#4

@Lukas Probably not. There’s a number of things that we need to give a proper experience for which we need raw data:

  1. Suggesting better networks which combine DSP + NN, either manual or through the EON Tuner.
  2. Determining right quantization technique (post-training or quantization-aware training? Per layer? Int8, int4, float16?)
  3. Application testing (see https://www.youtube.com/watch?v=3ls3hJcw4Ug&t=2615) to get confidence in your model.

Hope that helps!

2 Likes