Question/Issue: Is it possible to deploy an Edge Impulse project in Akida format for testing on the Brainchip MetaTF environment without having the PCI board? I have previously done a couple of the Brainchip tutorials using their MetaTF text environment which is setup on my PC
Project ID: To be developed, check before starting.
**Context/Use case:**Wanting to develop Object Detection model for Uni Project. I don’t have a Akida 1000 PCI board, but do have access to Jetson Nano if required.
Hi @Bourko1
If you are looking for the trained model in the fbz format, then just use MetaTF deployment in the studio:
In the downloaded archive, you can find akida_model.fbz file that you can use with the Akida Python library.
If you want to build EIM deployment with the model in simulation mode, then also use the MetaTF deployment and follow this instruction. To use the software simulation, you need to make a global define EI_CLASSIFIER_USE_AKIDA_SOFTWARE.
Hi there, thanks for the response. Im not sure what I did during deployment, it shows the following. I just used the tutorial model for voice.
Creating job… OK (ID: 7615524)
Scheduling job in cluster…
Job started
Scheduling job in cluster…
Container image pulled!
Job started
Removing clutter…
Removing clutter OK
Copying output…
Copying output OK
Scheduling job in cluster…
Container image pulled!
Job started
Application exited with code 1
ERROR: No Akida compatible model file. Did you use BrainChip Akida learning blocks?
Hi @Bourko1
If you used a tutorial project without changes, then it won’t work.
You need to go to Create impulse and replace the Classification block with the one for Akida:
The model came together ok, however it was not very successful in detecting my objects. I only used about 30 images and 2 objects, which I am sure is not enough to have an accurate model. However, when training the model, as I only have the basic membership, I can only train my model for 20 minutes.
When you select the retrain, does the model start from the last checkpoint in training ? It looks like this is the case.
My loss values are quite high given my learning rate and the minimal number of epochs I can have given the time constrains, thus the above question.
The public project is below: Brainchip test - Dashboard - Edge Impulse
I only have like 10% precision on one of the objects, which im ok with given my limited dataset and training. My plan is to increase the training and add some more images.
Hi there, as above, I am working on a project for university to the fully paid prescription is out of limits for me. Is there any way to get better access to training, or to allow the model to pause and retrain? My results aren’t very realistic with the limited training window.
Thanks
Allan