1.How much data/samples should be considered for edge impulse data forwarder to build a model.
2.what will be the model size , what is the frequency and baudrate to consider.
3. If i consider only 1 sample, can build a model from it.
You can start building a model with very few samples, what kind of project you are looking for, I’d love to here more about your project. I cannot give an estimation as it really depends on the use case. Also, keep in mind that any machine learning project is very iterative and gets better over time.
Again, that depends on your project type. Usually, accelerometer-based projects are pretty lightweight, audio can be either small or big depending on your NN architecture and your pre-processing method.
Only one sample won’t work as you need to have a least a couple of samples in your training set (e.g. 80% will be used for training and 20% will be used for validating), if there’s not enough data to perform validation then you can’t evaluate your model. Also, your model learns on your data so it’s important to add more data over time so your model can get better.
For one of our requirement, I was training a model for my nrf52840 dev kit with static IMU data. I have created a buffer in main.cpp with static data, and through “./edge-impulse-data-forwarder” we are sending to edge impulse studio. Since, our nrf52840 dev kit doesn’t have IMU sensor. Here, I’m able to make array of 19 samples only, if I add more than 19 samples in array means I’m getting the below error.
Error: “Failed to get information off device No valid sensor readings received from device. Note that the data forwarder can only auto-detect frequencies >2Hz. You can override the frequency via --frequency.”.
I was trying to set frequency with command below, even though I was getting the same error.
“./edge-impulse-data-forwarder --frequency 102”.
If we train our model with different kinds of IMU datasets(for updown motion, for waveform motion, for IDLE). We are getting training model accuracy as 23% only. We are suspecting that we provide only 19 samples. So, we are trying to give >19samples to obtain more accuracy (near to 100%). But, I was facing the above error.
Can you please provide your inputs on this error??