Not to sure what this is, maybe it’s edge impulse guys trying to make implementing akida easier by giving examples?
Contribute to edgeimpulse/example-akida-qat-deploy-block development by creating an account on GitHub.
github.com
edgeimpulse/example-akida-qat-deploy-block
init
Go to file
Latest commit
dansitu
…
16 hours ago
Git stats
Files
View code
Akida Custom Deploy Block
This repository provides a Custom Deploy Block that will convert and quantize deep learning models trained with Edge Impulse for use with the
Akida inferencing library, which supports Akida PCIe and RPi development kits, as well as software simulation of spiking neural nets.
This repostitory offers an early look at our partnership work with Brainchip - with sensor input, preprocessing blocks, and full integration with Edge Impulse tools coming soon!
Using this repository
- Chooose your favorite Edge Impulse image classification project. If you don't already have one, learn how to build your first ML model with Edge Impulse here
- Delete any existing learn block, create a new Classification (Keras)block. Within the block, use Expert Mode to define a model architecture that is compatible with Akida. For guidance on model compatibility, see the MetaTF documentation.
- Add the custom deploy blocks present in this repository to your account. To do this just install the Edge Impulse CLI and run the following commands in the meta_tf_deploy directory.
edge-impulse-blocks init
edge-impulse-blocks push
Follow the prompts and select Deployment Block when prompted.
After pushing the deployment block, navigate to
Uploading your block to edge impulse and follow the instructions to enable permissions for the deployment.
IMPORTANT Enable mount training block under /data, as this allows direct access to the training data, which is used to perform quantization-aware fine tuning of the model.
- Deploy your project. During deployment the model will be quantized for Akida inferencing, and then fine tuned in order to retain model performance. Observe the trained.fbz model artifact. This is your serialized akida model. The training and testing datasets are additionally provided as a test data source. Refer to MetaTF documentation here for information on how to run your newly trained spiking NN.
- If you wish to experiment with different values for learning rate and the number of epochs for fine tuning you can configure the --learning_rate and --fine_tune_epochs flags in the block configuration in your Edge Impulse Organization.
if you do return Factfinder we will be ecstatic and receive you with open arms.
......................................... 
You are to me the essence of this forum - your wisdom, knowledge and eloquence will be sorely missed. I agree and have felt that after the Megachips IP deal that BrainChip were on their way to greatness. After hearing and reading the transcript of the last AGM and then most recently Jerome’s podcast I have no doubts about where things will go from here. It’s just a matter of time now. I look forward to meeting you FF at the 23 AGM, I have not attended any previous AGMs but that one is not to be missed imo. Good luck and good health to you and your family and all the best for the future.
