• State TI Thinks Resolved
  • Locked Locked
  • Replies 1 reply
  • Answers 1 answer
  • Subscribers 99 subscribers
  • Views 214 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: Processors forum

Henry2333
Intellectual 490 points
Part Number: TDA4VM

Hi,friends.

I am now using the QAT model on the PC to get the intermediate results.

I found that the preprocessing result is wrong.

The settings on import.txt are as follows:

inDataNorm = 1
inMean = 0 0 0
inScale = 0.00392157 0.00392157 0.00392157

layer0 value is:

layer1 value is:

For example, the first value should be 175 * 0.00392157 = 0.68627475, right？

onnx model is 

tidl model is

Thank you

Henry

  • 0
    TI__Genius 11376 points

    Hi Henry, Sorry about the delay in replying.

    Inside TIDL, the mean subtraction and scaling is happening in fixed point. 

    inScale = 0.00392157 0.00392157 0.00392157

    will be quantized first to 8bits (if you are using 8bit mode in TIDL) and the operation will happen in fixed point.

    That is the reason why you are seeing this difference with the result of floating point operation.

    TIDL also has a floating point simulation mode in PC (won't work in EVM) and in that, this quantization is not there.