Training via CLI script yields only ~35% accuracy, while I previously observed ~99% accuracy using the GUI with the same parameters

Project name: nRF52840DK-continous-motion-recognition
Project ID: 728239

I was using the REST API to post neural network settings to the Edge Impulse cloud. When I train the model using this CLI script, I consistently get around 35% accuracy. However, when I train the same neural network settings via the GUI, I get 99% accuracy.

I’ve verified that the parameters — including training cycles, learning rate, batch size, and optimizer settings — are identical between the CLI and GUI.

Can someone help me understand why the CLI training yields significantly lower accuracy despite matching the GUI configuration?

The below are my screenshots for both via GUI and CLI script:
a) via script

b) via GUI


Best Regards,
Venkatesh K.

Hi Venkatesh,
It may be easier to see side-by-side results if you create a second Experiment in your Impulse Experiments and train one with your API calls, then the other via the UI.

The main things I would check here once you’ve done that are:

  • Are the input features into your learning block identical between the two models? If you head to “Live classification” and open the same sample in both experiments you can see the raw features and processed features. It’s possible your API calls are affecting the processing block as well (hard to tell without seeing the calls themselves)
  • It is unusual to see a neural network with 1 neuron, it is possible this is causing unusual behaviour
  • If you’re trying to use the API in order to speed up experimentation, I reccomend giving the EON tuner a go first- it is set up to run large sweeps of different parameters easily EON Tuner | Edge Impulse Documentation