how can I change FP32 unet model to 8-bit integer model in semantic segmentation case

[liang__heng]

I had a try  according to this “https://docs.openvinotoolkit.org/latest/_inference_engine_tools_calibration_tool_README.html”

 and had changed the fp32 model to int8 model  by the calibration tool in the simplified mode. but the accuracy loss is too large.

So I start to test the model calibration flow, but do not know how to  Convert data annotation files. the example  is  a  classification

case,the labels are words, and in the semantic segmentation the labels are pictures,  i can not  passing parameters in the same way as the example

[liang__heng]

Can anyone tell me how to solve this?

[HemanthKum_G_Intel]

Hi Heng,

You can try the solution provided in the following discussion - https://software.intel.com/en-us/forums/computer-vision/topic/810212

[liang__heng]

Hemanth Kumar G. (Intel) wrote:

Hi Heng,

You can try the solution provided in the following discussion - https://software.intel.com/en-us/forums/computer-vision/topic/810212

Many thanks！

[liang__heng]

Hemanth Kumar G. (Intel) wrote:

Hi Heng,

You can try the solution provided in the following discussion - https://software.intel.com/en-us/forums/computer-vision/topic/810212

I had a try and find that it is so difficult to convert a PF32 model to INT8，because  the command-lines have many difference  between  semantic segmentation case and  classification case .Could you send me some examples or any detailed parameter descriptions？thank you

[HemanthKum_G_Intel]

Hi heng,

You can also try DLWorkbench

[liang__heng]

Hemanth Kumar G. (Intel) wrote:

Hi heng,

You can also try DLWorkbench

I will try it right away.thank you ~~~

[Shubha_R_Intel]

Dear liang, heng,

DL Workbench is an early tool (in preview model) and still doesn't have full functionality. So you can always "go stick shift" (meaning use the manual method) using our Calibration tools. This github issue 171 post which I replied to is slightly old but it's still valid. If DL Workbench doesn't work for you the manual tools will.

Hope it helps,

Thanks,

Shubha