• State TI Thinks Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 479 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.

TDA2PXEVM: Questions regarding image processing tools and optimization of semseg on EVE's

Hafiz Muhammad Owais
Expert 1065 points
Part Number: TDA2PXEVM
Other Parts Discussed in Thread: TDA2

Hi Dear TI,

I have some thought provoking questions regarding tda2 and tidl. I have successfully implemented semseg with 2EVEs on tda2px. Now I have these following questions for which I need detailed answers.

1. As far as I understood, I cannot use opencv due to performance issue for image processing. Therefore, the only option I am left with is either vlib or eve_sw. I wanna ask, among them, which one is more suitable and closer to opencv. This arises another question

2. If I use eve_sw, my computational requirement is already occupied by the semseg, so I believe that I can only use vlib as it is for dsp. Please put some light on it.

3. Now, the third question is: How can I optimize the semseg so that I can use only one EVE to make it run on more than 10fps. What are the possible way which can contribute to fulfill this purpose.

4. Finally, I wanna ask, how can I check the memory consumed by tda2 while running the network or any usecase. I can calculate the theoretical value by converting it from the PC memory usage while inference and converting it for the tda2. However, I actually wanna see the practical value printed by the tda2 on the fly. 

Thank you once again, will be waiting for your detailed answers.

With best regards,

H. M. Owais

  • 0
    TI__Genius 17580 points

    Hi Owais,

    1 & 2. You can use vlib.

    3. There is no further optimization possible in the Semseg use case on one EVE, you need to more EVE's in parallel to improve fps.

    4,  You can refer to "tidl_tb.c" for all the allocated buffers.

    Thanks,

    Praveen  

  • 0 in reply to Praveen Eppa1
    Expert 1065 points

    Hi Praveen,

    Thank you for reply. Regarding your answer 3, as per my understanding the network structure contributes towards the performance on tda2. So I want to rephrase my question:

    The question is whether the network structure can enhance the performance (i.e. the fps), using the same usecase? If so, would you like to give any suggestions on it?

    If, we build another usecase, what are the things need to considered to increase the performance? Once again, my goal is achieve the highest fps on minimum number of EVEs. What are the possible choices do I have on your current platform.

    With best regards,

    H.M. Owais

  • 0 in reply to Hafiz Muhammad Owais
    TI__Genius 17580 points

    Hi Owais,

    Yes, the network structure can enhance the performance. For that you can use example scripts for training sparse models from Caffe-jacinto github link (https://github.com/tidsp/caffe-jacinto-models) and try couple of thimgs suggested below,

    1. You can try to increase sparsity in the model

    2. You can try to decrease the number of layers and use more convolution layers

    3. You try to reduce the resolution of feature sixe if possible use sizes multiple of 32 

    Thanks,

    Praveen