INT8 Two Network Support

[pdewan]

Hi,

I tried quantizing my custom model(which takes two inputs) to int8 precision using calibration tool. But i got error like "what():  Just networks with one input are supported."

According to my understanding model optimizer supports two inputs, but i got this issue with calibration tool.

So calibration tool does not support two input network or something wrong in my understanding.

Thanks,

Puru

[Shubha_R_Intel]

Dear pdewan,

You are definitely correct, Model Optimizer does support multiple inputs. It supports multiple outputs also. Can you kindly give me the exact calibrate.py command-line which is giving this error ?

Also, can you try benchmark app on the un-optimized version of your model - does it work OK ?

Thanks,

Shubha

[pdewan]

Hi,

Thanks for the reply.

I would like to add that i am are already using fp32 precision ir files of my custom network for demo purpose. So benchmark app does run.

The command that i used for quantizing fp32 ir files to int8 precision is: python calibrate.py -sm -m /home/darknet/puru/onnx/action_detection.xml -s /home/darknet/puru/ILSVRC2012_img_val/ -ss 1000 -td CPU -o /home/darknet/puru/onnx/.

I carried out this experiment on squeezenet network. Successfully converted fp32 ir files to int8 precision using similar command. But got issue with my custom network(two inputs).

Thanks,

Puru

[Shubha_R_Intel]

Dear pdewan,

I don't know all the steps you've done before encountering the error.  Please read my answers to this idz post . Did you follow the correct steps before jumping to calibrate.py ?

Thanks,

Shubha

[Shubha_R_Intel]

Dear pdewan,

Also Please also review this IDZ post where I've given extensive help on the OpenVino calibration tools. Since the calibration tools use Model Optimizer under the hood (the same code actually), if you are generating IR and benchmark app works, then maybe you're skipping some steps in your calibration process.

Let me know what you think, please post back here.

Also if you are not using OpenVino 2019R2, please upgrade as it has fixed many issues.

Thanks,

Shubha