Hi ST community,
I’m investigating an unexpected behavior regarding model weights generated by the STM32N6 CPU backend vs Neural-ART backend.
Model tested:
- Visual Wake Word (MobileNet-like INT8 model)
Observations:
CPU backend
Report shows:
- params # ≈ 210,850
- weights(ro) ≈ 42 KB
Neural-ART backend
Report shows:
- model weights ≈ 219 KB
- octoFlash weights ≈ 206 KB
This ~200 KB size is coherent with the real INT8 parameter count (~210k params).
What is surprising is that only the ST CPU backend reports ~42 KB.
I also tested:
--compression none
but the generated CPU backend still reports ~42 KB.
So this does NOT appear to be related to the documented CLI compression options (lossless, medium, high, etc.).
Questions:
- What exactly does weights(ro) represent in the CPU backend report?
- Is the CPU backend internally using another packed/compressed representation even when --compression none is selected?
- Is there any undocumented optimization specific to MobileNet/1x1 convolution-heavy architecturesfor the cpu ? ( I found the one specifiying special cases for 1x1 convolution for neural art st)?
- Why do the Neural-ART remain close to the raw parameter size (~200 KB), while the ST CPU backend drops to ~42 KB?
I could not find documentation explaining this behavior, so any clarification would be very appreciated.
i will attach the three generated reports .
