• State Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 1589 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.

Be applied the regression model on TIDL Library.

Ahan Tseng
Prodigy 190 points

Hi,

I have some questions about applied TIDL library.

I applied a regression model on TIDL.

The output are quantized with TIDL library. 

However, I need to know the relationship between the TIDL results (quantized to 8-bits) and the caffe jacinto results (float opration) to predict.

Can you give me some advice to realize the relationship?

Ahan

 

  • 0
    TI__Mastermind 22145 points

    Ahan,

      The output is 8 bit fixed point representation with the scaling factor. The 8 bit value can be divided by scaling factor to get the floating point value. The scaling factor is available in dataQ of (sTIDL_DataParams_t).

    Thanks and Regards,

    Kumar.D

  • 0 in reply to kumar.desappan
    Prodigy 190 points

    Hi, Kumar

    Now I used TIDL sample model (jsegnet) to validate the scaleing factor.

     Layer    1 : Max PASS :    19211 :    15301 Out Q :      254 ,    43861, TIDL_BatchNormLayer, PASSED  #MMACs =     1.57,     0.00,     1.57, Sparsity :   0.00, 100.00
     Layer    2 : Max PASS :   106321 :   108432 Out Q :     7934 ,   108857, TIDL_ConvolutionLayer, PASSED  #MMACs =   314.57,   261.23,   278.92, Sparsity :  11.33,  16.96
     Layer    3 : Max PASS :    19248 :    22842 Out Q :     8500 ,    22932, TIDL_ConvolutionLayer, PASSED  #MMACs =   301.99,   257.43,   267.39, Sparsity :  11.46,  14.76

    1.  Can I think the "scaling factor" is the Out Q?

    2. If yes, for the jsegnet, the first layer is bias layer. The output were -128 to 127 (quantization to 8 bits), then, they divided by the scaling factor will be -0.5 to 0.5.

    But, the output on caffe jacinto was -128 to 127. How can I explain the result ?

    3. If not, can you tell me how to get the scaling factor ?

    Ahan

  • 0 in reply to Ahan Tseng
    TI__Mastermind 22145 points

    Hi Ahan,

     The Out Q (scale factor) is in Q8 fixed point format. This needs to be divided by 256 to get the scale factor in floating point.

    Thanks and Regrads,

    Kumar.D