I’ve trained a model just now, but I found that the ROM size is 2.2MB, which is much capacity to my Nano 33 BLE. Thus I wanna know the factors that may infect the ROM size.
I’ve collected the voice data about 46 mins , using spectrogram and NN classifier
The ROM size will reduce if I’am using MFCC, but the accuracy will decrease…
Is there any solutions?
Thanks for helping me!
Hi @zhongrui what are the parameters of your DSP block? And do you see the 2.2M at the neural network page or the spectrogram block? 2.2M is a lot more than we usually see, you should see ~10K RAM with the default config for the spectrogram, and a similar number for the neural network.
My project ID is 25647
And here is my configuration
I modified the WindowSize to 1300 and the training cycles to 220
So must I use MFCC ?
You can keep using the Spectrogram but your NN architecture is too large for the Arduino. Have you tried witht the default 1D Conv architecture preset? This will decrease drastically the ROM usage.
I reduced the quantity of neurons and it worked! Very appreciate your help!