the result mismatch between running the model passing vela with & without --optimise Size

[kris-himax]

I run the .pte files which pass vela with and without --optimise Size option on ml toolkit.

• mv2_quant_himax_ini_opt_size.pte-> the answer is wrong

• mv2_quant_himax_ini.pte -> the answer is correct

[manuelcandales]

Thank you for reporting this mismatch. Could you provide the commands you used for creating those .ptes.

[kris-himax]

Hi @manuelcandales ,
I add the "--optimise Size" at here

    elif "ethos-u" in target:
        compile_spec = EthosUCompileSpec(
            target,
            system_config=system_config,
            memory_mode=memory_mode,
            extra_flags=["--verbose-operators", "--verbose-cycle-estimate", "--optimise Size"],
            config_ini=config,
        )

And run the python.

python3 -m examples.arm.aot_arm_compiler --model_name=mv2 --target=ethos-u55-64 --delegate --quantize --config=himax_vela.ini --system_config=My_Sys_Cfg --memory_mode=My_Mem_Mode_Parent --output=mv2_quant_himax_ini_opt_size.pte

The himax_vela.ini is download from here.

Thanks,
Kris

[zingo]

Hi @kris-himax just a fast check, a problem in this area was fixed in Vela a short while ago, depending on the version of Vela you use it might already be fixed can you check that you have version 4.5.0 (vela --version) and if not you could retry re-runing the examples/arm/setup.sh so that you get the latest version.

[kris-himax]

Hi @zingo ,
I change the vela version from 4.4.1 to version 4.5.0 (vela --version) , and then it work normally.
The answer is correct by running model "--optimise Size".

Thank a lot.
Kris