Hi @davwang this is a great project! I am getting much better accuracy after manipulating the window size in the impulse design and then changing the neural network architecture to use a 1d-convolutional layers. You can see the modified design here: https://studio.edgeimpulse.com/public/49291/latest
Hope this helps.
Although I have used your neural network to improve the accuracy for trainning my dataset , these result is worser than your ones. Maybe the dataset should be handled? Can you give some advice for handing different sample?
Ok, I know what happened, I did “rebalance dataset” from the Dashboard page so that there is some test data in the project. Your original project only had training data. Usually you want to have about 20% test data which is data the model has never seen before. Using your original data amount in the training bucket, if you set the window size to 2000 ms instead of 3000 ms, this will give you some better accuracy, arriving at the following result.
The generic formula is (number of axes * window size * sample rate), so in this case it would be:
(1 * 2 * 5388 ) = 10776 features. If window size is 3 seconds, then it would be 16164 features.
Do you mind explaining us how did you build your logger?
If you do a blog post (hackster, medium or whatsoever) about your project, we’d be happy to share it too (and I’d be happy to try your tutorial on my side ) !
@louis Thanks for your affirmation about this project. Maybe sometime later, I will try to create a blog post. I have an account from hackster. I havenot known the medium or whatsoever.
) !