Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6392 Discussions

Unable to run converted ssd V2 model ERROR-"unsupported output"

Petrica
Beginner
840 Views

error_output.png

I have successfully converted my model, but I cannot run it because of the error "unsupported model output" 

 

I will attach the saved model and the parameters that I have used for conversion.

 

This is how I tested the model (using the sample file from the toolkit):

python3 object_detection_demo.py -m ./TFOD_40/transform/saved_model.xml -at ssd -i ./mitoi.mp4 -d MYRIAD

 

And this is how i converted my model:

mo_tf.py --saved_model_dir ./TFOD_40/saved_model --tensorflow_object_detection_api_pipeline_config ./TFOD_40/pipeline.config --output_dir ./TFOD_40/transform 


 

 

0 Kudos
4 Replies
Peh_Intel
Moderator
763 Views

Hi Petrica,


Thanks for reaching out to us.


The object_detection_demo is expecting detection_out with shape [1,1,N,7] and each detection is expected in the following format: [image_id, label, conf, x_min, y_min, x_max, y_max]. You can refer to SSD with MobileNet V2, one of the validated model for the demo.


Besides, I also notice that your IR files are empty which need you back to the conversion process again. By the way, I also fail to convert your model. In the previous discussion, you mentioned that you was able to convert the model with the same pipeline config file in the past. I would like to know what have been changed which led to conversion failure.



Regards,

Peh



0 Kudos
Petrica
Beginner
755 Views

In the past, like 2 months ago I successfully converted  with a similar pipeline config file, just replaced this training_config argument " data_augmentation_options {
random_horizontal_flip {
}
}" 

with  the following parameters:

"data_augmentation_options {
random_adjust_contrast {
}
}
data_augmentation_options {
random_adjust_saturation {
}
}
data_augmentation_options {
random_adjust_brightness {
}
}"

And since then, I am not able to convert the old model with the  transformations config (.json) file. Probably I used another parameter when running the conversion script, but I am not able to remember what I have used, exactly.

From your answer I understand that my model cannot be converted, or is it a different way for me to use the Intel Neural Compute Stick with my model.

 

0 Kudos
Peh_Intel
Moderator
697 Views

Hi Petrica,

 

You can refer to the parameter in the config files here.

 

Based on your current model (saved model format), the node names of your model does not match with the node names in the JSON file. These node names are stated under the "start_points" and "end_points" of the "id: ObjectDetectionAPIProposalReplacement" in the JSON file.

 

Hence, if you proceed to convert your model without the transformation config file (JSON file) will result in empty IR.

 

wrong_model.JPG

 

Besides, I also notice that the input of your model does not attach to the graph. It looks weird for me. Please verify on your side that your model is trained completely.

 

input_not_attach.JPG

 

You can use Model Optimizer to visualize the input graph of the model in TensorBoard.

1.      Dump the input graph of the model.

mo_tf.py --saved_model_dir=<saved_model> --tensorboard_logdir=<TENSORBOARD_LOGDIR>

2.      Visualize the input graph of the model in TensorBoard.

tensorboard --logdir=<TENSORBOARD_LOGDIR>

3.      Copy and paste the output URL in the browser.

 

 

Regards,

Peh

 

0 Kudos
Peh_Intel
Moderator
652 Views

Hi Petrica,


Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored. 



Regards,

Peh


0 Kudos
Reply