First time posting here!
When using the spectrogram to pre-process accelerometer data, the number of inputs that are created from that step does not correspond with the explanation of the spectrogram in the documentation.
3 axis accelerometer data for 1 second (fs = 100 Hz)
Frame length = 0.02 seconds
Frame stride = 0.01 seconds
FFT length = 128
Number of NN inputs
According to documentation: 99 frames * (128/2+1) * 3 (axis) = 19305 inputs
According to edge impulse: 9750 inputs (which is 50 frames * (128/2+1) * 3)
By calculating (and counting on the spectrogram previews) I found out that it used 50 frames instead of 99. It also uses FFT/2+1 number of frequency bins, which does not correspond with the documentation (there it says FFT/2).
So how should I actually calculate the number of frames that are used?
Hi @JLannoo With a one second sample at 100Hz and 0.02 frame length / 0.01 frame stride this yields me a 99x65 matrix per axis, and lists 19,305 input layer as well:
What’s your project ID? I’ll take a look.
FFT/2+1 is indeed a documentation bug, have fixed this.
Hmm, that’s how it should be!
My project ID is: 27208
Thanks for the quick response!
Very interesting, your data is at 101Hz, and indeed only generates a 50x65 spectrogram… Let me dive into that!
Indeed, sorry for that detail!
My data is at 101Hz (although I configured a datarate of 100 Hz in the forwarding microcrontroller).
Looking forward what the influence of that 1 Hz does!
OK, so this is what happens: To deal with frame stride / length that does not align exactly to your frequency, we normalize the values via
ceil(sampling_freq * frame_length). For 2ms @ 100Hz => 2 values per slice, but for 2ms @ 101Hz => 2.02 => 3 values per slice. The same goes for the frame stride, leading to taking 3 values per frame, and a stride of 2 values => leading to 50 frames in total.
Quick workaround: set frame length to
0.019 and frame stride to
Note: if the data forwarder sees an incorrect frequency, you can override via
edge-impulse-data-forwarder --frequency 100 (but you shouldn’t mix multiple frequencies in one project, so I’d keep it at 101 for now).