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

Model conversion

Hi,

I'm working on running blazeface and posenet models with inference engine. First, I used the following command to convert them:

- For Posenet:

python3 mo_tf.py --input_model posenet.pb --output_dir /users/thibault/Documents/posenet --data_type FP32 --disable_nhwc_to_nchw

- For Blazeface :

python3 mo_tf.py --input_model blazeface.pb --output_dir /users/thibault/Documents/posenet --data_type FP32

After running this model I get a bad display of landmarks for both two models. The code is good but I guess I missed something during tensorflowLite model conversion.

If someone has an idea how to convert correctly this two models.

Thanks

 

0 Kudos
4 Replies
Highlighted
Moderator
139 Views

Greetings,


If possible could you provide the link to the resources of blazenet and posenet so I could try to convert them myself?


Fyi, this is the official documentation on how to convert Tensorflow model:

https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Mode...


Sincerely,

Iffa


0 Kudos
Highlighted
Beginner
132 Views

Thanks for your reply,

 

I found the frozen models from here with script download:

Blazeface:

https://github.com/PINTO0309/PINTO_model_zoo/tree/master/30_BlazeFace/01_float32

Posenet:

https://github.com/PINTO0309/PINTO_model_zoo/tree/master/03_posenet/03_posenet/01_float32

I just noticed that the posenet model has been converted for openvino (FP32 and FP16) from this link 4 days ago. I'm going to try with this one too:

https://github.com/PINTO0309/PINTO_model_zoo/tree/master/03_posenet/03_posenet/05_openvino

Regards

0 Kudos
Highlighted
Moderator
112 Views

Greetings,


You can actually get the readily converted model in IR form by running the download script inside openvino folder (in Blazenet/Posenet folder).

Since this came with the Blazenet and Posenet model's package and clearly stated this is for openvino, this should already been correctly converted.



for Blazenet you will get 2 folders (face detection back and front)


Same thing applies for Posenet


Sincerely,

Iffa



0 Kudos
Highlighted
Moderator
100 Views

Greetings,


Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question.


Sincerely,

Iffa


0 Kudos