Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Beginner
21 Views

Loading network with very long time in NCS2

I test load a model in NCS2 , it speed 10 minutes.

When i load it in CPU, it only speed 1 second.

log.info("Loading IR to the plugin...")
exec_net = ie.load_network(network=net, num_requests=2, device_name=device)
print("load time:",(time.time()-load_begin)*1000)

What 's wrong with NCS2?

My openvino version is 2020.1

0 Kudos
3 Replies
Highlighted
21 Views

Hi Jin,

Thank you for reaching out.

Could you please tell us the following:

  • Which demo or sample are you using?
  • Can you share with us the command you are using for the model optimizer?
  • Do you think you can attach your model and files for us to test it from our end?

Regards,

Randall B.

0 Kudos
Highlighted
Beginner
21 Views

Hi Randall,

        Thanks for your reply. I use the model converted by myself. I use the RetinaNet and it's backbone is ResNet50.Is the backbone is so big?The model is export use torch.onnx.export and the input size is [1,3,800,1088] . This is the model optimizer command.

python mo_onnx.py --input_model onnx_retina_ori_out.onnx -o F:\Intel\FP16 --data_type FP16

       I also test the adivce from you in other post. It works.   The load time is normal, but the forward time become slower.  The OPTIMIZATION may speed long time. Is there any way to save the  optimized model?

ie.set_config({'VPU_HW_STAGES_OPTIMIZATION': 'NO'}, "MYRIAD")

if necessary, i can provide the converted model.

              Jin Yang

Randall B. (Intel) wrote:

Hi Jin,

Thank you for reaching out.

Could you please tell us the following:

  • Which demo or sample are you using?
  • Can you share with us the command you are using for the model optimizer?
  • Do you think you can attach your model and files for us to test it from our end?

Regards,

Randall B.

0 Kudos
Highlighted
21 Views

Hi Jin Yang,

Can you try to input a smaller image?

The default value for VPU_HW_STAGES_OPTIMIZATION works for the majority of the models. Consider that the ONNX is not supported on MYRIAD. Can you try the ResNet50 from Tensorflow* Model?

Regards,

Randall B.

0 Kudos