• State TI Thinks Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 1 answer
  • Subscribers 99 subscribers
  • Views 473 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.

Original question:

SK-AM62A-LP: edge-tidl-tools TIDLRT_invoke Stuck When the model has multiple outputs

SK-TDA4VM: TIDLRT multiple outputs

Neal Yan
Prodigy 160 points
Part Number: SK-TDA4VM
Other Parts Discussed in Thread: TDA2

Hi,

I'm trying to deploy a model with multiple output tensors, I noticed this similar thread, my questions are:

1. I'm not confident about how to access multiple output tensors, based on this thread, the key is changing of "j" right?

2. I remember a long time ago, when deploying models, I need to deal with the paddings (both inputs and outputs) but I didn't find anything in this tidlrt code.

3. I also remember I need to deal with de-quantization scalar, how to get this number for each output tensor?

My model is in caffe format and I had a hard time converting it to onnx, so have to stick with tidlrt.

I guess a standard 3 output tensor OD example would cover all these questions, can someone help?

Thanks,

Neal

  • 0
    TI__Genius 17026 points

    Hi Neal,

    1. You will need to allocate an output tensor and then assign a pointer that TIDL-RT can use for copying memory out from the accelerator.

    2. For paddings, could you give an example to clarify? The input tensor shape needs to match the model - this can be accomplished several ways, including padding, resizing/interpolation. For the outputs, do you mean paddings from the end of one output tensor to the next one if everything is in the same area of contiguous memory? If you're assigning output memory locations like above, then padding shouldn't be an issue

    3. I don't believe you need to apply a de-quantization function on the output; that should already be handled on the output of TIDL. The outputs are most likely in fp32

    Link to TIDL-RT documentation (this is sometimes difficult to find or search shows old versions):

    software-dl.ti.com/.../index.html

    Best
    -Reese

  • 0 in reply to Reese Grimsley
    Prodigy 160 points

    Thank you for the detailed information, I guess point 2 and 3 are tda2 era deployment routine. I don't have questions for now, I will try to follow your instructions first.

  • 0 in reply to Neal Yan
    TI__Mastermind 24041 points

    Sure Neal.

    Closing this thread.

    Regards,

    Pratik