Community
cancel
Showing results for 
Search instead for 
Did you mean: 
116 Views

[NCS] YOLOv3 from Caffe to IR using extension layer

I want to create an intermediate representation from Caffe with regionyolo_extension, i want to create it by model optimizer.. my question is, how to do conversion from Caffe to IR and include regionyolo_extension? i have my own prototxt and caffemodel (darkNet to Caffe conversion), but i hope i can inference yolo layer in my NCS, not my PC..

say, if the output layer name in my prototxt is "layer83-yolo", "layer95-yolo"  and "layer107-yolo"

thank you

0 Kudos
4 Replies
David_C_Intel
Employee
116 Views

Hi Candra,

Thanks for reaching out. You can check this documentation on How to Convert YOLOv3 Model to IR. Try it and tell us if you manage to convert it, in case you encounter some issues, please send us your files for us to try it.

Regards,

David

116 Views

DavidC (Intel) wrote:

Hi Candra,

Thanks for reaching out. You can check this documentation on How to Convert YOLOv3 Model to IR. Try it and tell us if you manage to convert it, in case you encounter some issues, please send us your files for us to try it.

Regards,

David

 

Thank you David, but in that post it shows conversion model from Tensorflow.. what i want to do is converting from Caffe to YOLO with model optimizer

David_C_Intel
Employee
116 Views

Hi Candra,

Thank  you for your reply. Please take a look at this Converting a Caffe* Model documentation.

Let us know if you manage to get the IR files.

Best regards,

David  

 

Vladimir_Dudnik
Employee
116 Views

you may try command like this (it was tested to convert to IR yolo v3 model from caffe:

python mo.py --model_name yolo-v3 --output_dir <output dir> --framework=caffe --data_type=FP16 --reverse_input_channels '--input_shape=[1,3,416,416]' --input=data '--scale_values=data[255.0]' --output=layer83-yolo,layer95-yolo,layer107-yolo --input_model=<your_model.caffemodel> --input_proto=<your_model.prototxt> --caffe_parser_path=/opt/caffe/src/caffe/proto

Regards,
  Vladimir

Reply