• State
  • Locked Locked
  • Replies 1 reply
  • Subscribers 98 subscribers
  • Views 215 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: Output of Pointpillars without TIDLExecutionProvider in ONNXRUNTIME

David Pujol
Prodigy 30 points
Part Number: TDA4VM

Hi all!

I am working on the PointPillars model that TI provides to be compatible with the TDA4 and I have some doubts regarding the output with 1 class.

The penultimate layer, before the concat, has 3 outputs which are the following:

cls_preds = shape=(1, 248, 216, 2 * num_classes * num_classes)
box_preds = shape=(1, 248, 216, 14 * num_classes)
dir_cls_preds = shape=(1, 248, 216, 4 * num_classes)

But my question is regarding the reshape layer. What is the final output? I think that the total number of preds are 107316 
but the vector of 10 values I don't know it

Moreover, if I run the example notebook in the edgeai cloud with the TIDLExecutionProvider with onnxruntime,
the output is totally different and it's based on the prototxt with a final vector of (50,9) where 50 is the number of 
predictions and 9 is the expected bbox3d with the class, confidence and yaw.

Thank you in advance,
David
  • 0
    TI__Expert 4725 points

    Hi David,

    pointPillar model is not supported with onnxruntime for now.

    The output of concat layer is raw output which goes to detection layer. Detection layer of point pillar model was getting exported into complicated graph hence we avoided keeping it in onnx model. we can try now if the graph of detection layer is simplified in mmdetection3D repo or not. 

    For TIDLExecutionProvider to understand this graph we inserted "Reshape" layer in place of "DetectionLayer". This is our internal hard coding. Shape of output of "Reshape" Layer is forced to match the shape of TI Detection Layer. And it is 1 + 9 actual output, where 1 is indices of the detection.

    If it is really blocking you then you can try exporting complete onnx model with detection layer from below repository. 

    https://github.com/TexasInstruments/edgeai-mmdetection3d

    Please let me know your findings. will be happy to work with you to correct our internal hard coding.

    regards

    Deepak Poddar