STM32N6 target build is no longer generating network_data.hex file

Help
#1

Question/Issue:
STM32N6 target build is no longer generating network_data.hex file

Project ID:
[849259]

Context/Use case:
[During normal Object detection project build]

Steps Taken:

  1. [Step 1]
  2. [Step 2]
  3. [Step 3]

Expected Outcome:
[All required files to be created and added the project.zip, to then be flashed into STM32N6 dev-kit using flash.sh]

Actual Outcome:
[flash.sh fails reporting missing file]

Reproducibility:

  • [* ] Always
  • [ ] Sometimes
  • [ ] Rarely

[Any other information that might be relevant]

#2

Hi @DJE666 ,

if you are building for CPU, that’s correct, there’s no need to flash the weights as with this build, the inference runs on the MCU, we recently cleaned the output for this build.

image
image1558×440 65 KB

If you want to run inference on the NPU, you need to use this deployment option:

image
image1560×224 47 KB

regards,
fv

#3

Hello FV,

Thanks for the prompt reply.

If I am building for the ST STM32N6 CPU target (which does not need the network_weights.hex file), then why does the auto-generated flash.sh file contain a command to program the network_data.hex file when this file is not going to be created?

Should I manually edit the flash.sh file (if so how), or run individual commands to program the dev-kit when the network_data.hex file is missing (not generated nor required).

Regards,

David

#4

Because I forgot to update the flash.sh … thank you for reporting! gonna clean it

you can use the firmware arg ie
flash.sh firwmare
it will just flash the firmware.

fv

#5

Hello FV,

Thanks for the update. I re-read the STM32N6 instructions and successfully flashed the board with just the FW file late last night, but the system still fails to detect any objects. Not even the briefest flash of a bounding box. :frowning:

Got to think my project is fundamentally broken somewhere, (bad data?), but what’s most perplexing is that I ran this same tools-flow with the same STM32N6 dev-kit back in July, and it all worked 1st time!

image

This is the result of the model running in the browser before being ported to the STM32N6 hardware.

FYI this is from project iD 849259.

Regards,

David

#6

I’ll check the project

fv

#7

Hi @DJE666

i checked your project, first I would remove the unknown label and the sample associated with it.
But even after that, seems the model is not performing well on device. I tested it with other devices and I can’t get any object detected :frowning: I’ll keep checking; I’m testing pointing the camera to my monitor with one of the sample.

I get the right detection on static inference, so you can try by removing the unknown label, retrain and check on real object, let me know!

regards,
fv

PS: FOMO is not available on N6, buyt you can try running Yolo Pro model.