• State Resolved
  • Locked Locked
  • Replies 4 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 403 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.

[TDA4M] Scale layer automatically replaced by batchnorm.

Sungan Gweon
Prodigy 65 points


Hello,

When we call scale layer without batchnorm layer in caffe prototxt (e.g, to multiply 3 to output of fully connected layer), it is automatically replaced by batchnorm layer on the device.

How we can use standalone scale layer on the device? I couldn't find any option in the configuration file regarding this problem.

Thank you.

  • 0
    TI__Guru 55685 points

    Hi Sungan,

    TIDL doesn't support scale as a standalone layer.

    Hence at the time of import it is being replaced by batchnorm and there is no option to enable scale as standalone.

    You can refer to tidl_j7_01_00_00_00/ti_dl/docs/user_guide_html/md_tidl_layers_info.html for more information.

    Regards,

    Rishabh

  • 0 in reply to Rishabh Garg
    TI__Guru 55685 points

    Hi Sungan,

    Are you facing any issue when scale is replaced by batchnorm?

    Regards,

    Rishabh

  • 0 in reply to Rishabh Garg
    Prodigy 65 points

    Hi Rishabh,

    Well, I'm not sure whether it solved my problem. My intention was to multiply feature map by a number, for example, MobileNet_conv3 * 5. However, I found it out that it there is no scale layer, but batchnorm layer when I analyzed the visualized graph. I have to check the value first to see whether it's right or not. But I'm worrying that there is performance issue since batchnorm uses redundant operation than scale.

    Thank you.

     

  • 0 in reply to Sungan Gweon
    TI__Guru 55685 points

    Hi Sungan,

    I would suggest you to first verify the functionality.

    I agree that batchnorm operation can take more time compared to standalone scale but the difference will not be too much, you can consider having custom layer implementation for scale layer while looking at performance optimization.

    Regards,

    Rishabh