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.

TDA4VH-Q1: "Unknown input dimension, not supported by TIDL" even though the dimension is static in ONNX

Part Number: TDA4VH-Q1


Tool/software:

Hi!

I'm trying to compile a model with a Softmax layer that I want to offload to the DSP. However, during compilation, TIDL reports the error "Layer 510 - op type Softmax, Unknown input dimension, not supported by TIDL". To me, this sounds like TIDL refuses a node with dynamic shape, but I have run the model through ONNX's shape inference, so the shape should be static. This is also what I see when I look at the computational graph in Netron. What could be some possible causes for this error?

I have tried softmaxing over different dimensions and reshaping the input tensor to have fewer dimensions, but without success.

Best regards

Olof

  • Hi;

    Could you provide some details about your model with a softmax layer? Also, could you share what you saw in the Netron, when you upload your model?

    Thanks and regards

    Wen Li

  • Hello!

    I don't seem to be able to upload the ONNX file directly. Instead, I have cut out a subdivision of the model and attached a screenshot of it and the details regarding the softmax node. This subdivision still reproduces the error. I hope that this is workable.

    I found out that eliminating the transpose node and changing the input accordingly gets rid of the error on this subdivision. However, this would probably be very time-consuming to achieve on the full model.

    Have I done something wrong with the transpose node? I don't see any conflicts with the requirements stated in your documentation (link).

    Best regards

    Olof

  • Hi again!

    I noticed that by putting another reshape after the transpose, the error vanishes, and the softmax node can be run on the DSP. This is probably a workaround that would suffice for me, but I guess that this behavior could warrant some bug hunting in TIDL unless there is an error that I made.

    Best regards

    Olof

  • In the full model there were some concat nodes preceded by transpose nodes that resulted in the same error. They were also mitigated by adding a dummy reshape between the transposes and the concats.

    Olof