Group Actions
Engagement
  • Author Author: ralphjy
  • Date Created: Date Created
  • Views 1655 views
  • Likes 18 likes
  • Comments 9 comments
Related
Recommended

Bee Healthy - Blog 5: Edge Impulse Model

ralphjy

I've selected the Buzz3 Audio Dataset to test model development and deployment to the Nicla Vision board using the Edge Impulse toolset.  I discussed this dataset in my previous post - Bee Healthy - Blog 4: Audio Dataset.

The Buzz3 dataset contains 11746 2sec samples split as follows:

train 
class hour min samples
bee 1 36 2880
cricket 2 0 3600
noise 1 24 2520
test
class min sec samples
bee 35 42 1071
cricket 19 14 577
noise 36 36 1098

I may end up rebalancing the training data, but I'll try it out first.  I already have an Edge Impulse account , so I just needed to create a new project that I named Bee_Present:

image

A great feature of the Edge Impulse Studio is the ability to upload existing data files provided they are in an accepted data format or file type.

Data Acquisition Formats:

  • CBOR
  • JSON
  • CSV

File Types:

  • WAV
  • JPG
  • PNG
  • AVI
  • MP4

The audio files in this dataset are all WAV files, so uploading the files with the appropriate labels was easy.

Data Acquisition

image

image

Impulse Design

For non-voice audio data it is generally recommended to use a Spectrogram Processing block - either linear or MFE (Mel-filterbank energy).  The MFE is designed to match human audio perception.  One might expect MFE to perform better because the data was labeled by human listeners.

image

I decided to try it both ways.

Audio (MFE) Impulse

image

Train the Neural Network:

image

The Quantized (int8) model accuracy is reasonable when validated against the training data:

image

But performs less well against the test data as would be expected:

image

Linear Spectrogram Impulse

image

Train the Neural Network:

image

Again, reasonable accuracy with the training data

image

But worse performance against the test data

image

Summary of Project versions:

image

The inference time of the Mel Spectrogram Impulse (35ms) is half of the Linear Spectrogram Impulse (71ms).  And it also has better memory efficiency and accuracy (with this dataset), so that's the model that I will use.

My first test when I get the Nicla Vision hardware will be to use the microphone to verify that I can do "Live classification" before I try to deploy the model.  

Parents Comment Children