Hello Edge pulse team, can you tell me where the parameters in the Spectrum Analysis are not shown
Hi,
@norik.badalyan your existing projects and deployments can and will continue to use v1 of this block, and will have the same parameters.
However, for new projects (or if you delete and re add the spectral analysis block), you will get v2. It’s the same power and concept as v1, just much easier to use (hence less parameters). This video explains: Loom | Free Screen & Video Recording Software | Loom
1 Like
PS let me know how v2 goes for you, especially if you can compare to v1!
1 Like
hi yes i already used v2 in v1 i got better results because i could change the fft parameters.I took version 1 from git and added custom block after deployed to microcontroller, will it work or do i have to implement this block myself (v1)?
Thnaks Norik
Nope, no need to implement anything, when you deploy you’ll still get v1 cpp and header files generated to run v1.
I’m curious about the performance difference, do you mind if I peek at your project? If not, can you share your project Id or a url?
1 Like
yes of course I can add, I tried to train the model using v2, it turned out a little worse and then I tried to add a bone block and got better results (v1)
Project ID 127696 , now i used v2 but yesterday using v1 i got 99% accuracy
Hi @norik.badalyan , thanks for sharing your work. I achieved 100% accuracy (0.03 loss) with v2 by applying a low pass filter. There’s double goodness in v2 for filtering…in addition to applying a Butterworth filter to the input signal, we also achieve dimensionality reduction by removing FFT bins above/below the cutoff frequency. Less dimensions means less data needed for good results!
I didn’t overwrite your DSP block settings or your trained NN. I added my work as a new version. You can see the results here: Login - Edge Impulse. Note, if you want to to deploy that model, you have to select “Set version as primary” in the version control for the NN version I trained.
Out of curiosity, what is this a model of?
1 Like
hello, yes, I will definitely try this recognition of some commands using an accelerometer on a low energy microcontroller
Hi @AlexE , please look with v1, my result is better than in v2.
First I tried it on v2, I got 98.21 maximum in training and 92 on test, and on v1 100 in training and 97 on test
Hmm, did you run that with the settings I applied to v2? (You would have had to go into the versions of both the block and the NN to see them…I didn’t want to stomp on what you already had). Since what I noted was 100% accuracy, I think you’re still running v2 with default settings perhaps, which I agree, did not hit 100%
1 Like