Model Optimizer Parameters for Accuracy

[Serdjan_R_Intel]

I just trained my own object detection model in Tensorflow using the ssd_mobilenet_v1_coco model pre-trained checkpoint to train my own labeled data and make my own frozen inference graph.  I know in the openVINO Model Optimizer certain parameters such as scale_values and mean_values must be set to maintain the accuracy of a model. 

Where/how would I go about finding these values to make my IR as accurate as possible?  Are there generic values for ssd_mobilenet that will make an IR more accurate, or do I need to look at my Graph in Tensorboard where I did my own training?

[Mark_L_Intel1]

Hi Serdjan,

I didn't follow the path you did, but in theory, you should find those values in the output of the training process.

Mark

[Serdjan_R_Intel]

Hi Mark,

I forgot to post that my problem was solved.  There was a big in the MO for ssd_mobilenet object detection models.  I was sent the R3 version of OpenVINO and now my model runs fine.

[Mark_L_Intel1]

Good to know and you have fast paced work!

Mark

[Monique_J_Intel]

Hi Serdjan,

Just a reminder to please use the OpenVINO internal forum for all your questions.

Kind Regards,

Monique Jones