Spectrogram Processing block

Hi @shawn_edgeimpulse

I am Training a model by recognising gestures made with the smartphone - using the phone`s inbuilt accelerometer for motions like idle, leftright and updown . while designing the impulse i am adding processing block as spectral analysis and learning block as classification(keras) with window size 2000ms ,window increase 80ms , no. of epochs 100 and learning rate 0.0005 and getting the model accuracy as 100% , when i will deploy to the device and run the impulse continuously i am getting the error as " ERR: Unknown extract function, only MFCC, MFE and spectrogram supported
ERR: Failed to run classifier (-5)" , To resolve this error i am training model by stacking 3 individual spectrogram blocks and associate them with one axis each. For two types of motion(idle,lefrright) model is giving 100% accuracy , when i add one more type of motion(updown) model accuracy is decreasing below 72% .To increase the accuracy , i have increased the no. of training cycles upto 300 but there was no change in accuracy.
why we are not getting accuracy (for three motion types) for spectrogram unlike spectral analysis ?
Can we change something in parameters to get more accuracy for spectrogram?

Thanks
Mahadevi

Hi @Mahadevi,

Which device are you trying to deploy to? Recognizing gestures would work better with the spectral block. If you need to use the Spectrogram, you most likely need to use a larger network as the default Dense Network is very small vs the number of features out of the spectrogram (Conv Networks could also be a good alternative).

Aurelien

Hi @aurel

I am trying to deploy it on nrf52840 devkit.

Thanks
Mahadevi

Hi @Mahadevi,

Can you share your project ID so we can see how you set up your project?

As @aurel mentioned, the “spectral analysis” block will likely perform a lot better (and be more efficient) than spectrogram blocks for things like gesture recognition.

1 Like

Hi @shawn_edgeimpulse

Thanks for the reply , i have trained the model with spectral analysis i am able to proceed.

Thanks
Mahadevi

2 Likes