• State
  • Locked Locked
  • Replies 3 replies
  • Subscribers 97 subscribers
  • Views 256 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.

Original question:

TDA4VM: The trouble at "TVM/Neo-AI-DLR + TIDL Heterogeneous Execution"

TDA4VM: calibrate with a lot of image data

Tsurumoto
Genius 16975 points

Part Number: TDA4VM

Hi Zhao

I got an additional question related to 

https://e2e.ti.com/support/processors/f/791/t/979543

Customer agree that the reason of above E2E issue is lack of memory.

Question: What should they do to calibrate with a lot of image data?

 

The issue was happened when they try to do quantization  with 32GB Linux box.

Their custom model requires a large size of image data (480*640).

And also, the size of model itself is also big.

If they can do the calibration only with a few number of image data then it could affect accuracy.

Thanks and Best regards,

HaTa

  • 0
    TI__Expert 4105 points

    Hi HaTa,

        Do you have a setup on your end?  Can you try reproduce the problem?

        I tried the model you provided with 100 calibration images, it succeeded on my Linux box with 16GB memory.  Make sure the Linux box is quiet without other high-memory-consuming processes running in parallel.

        Oh, also increase the stack size on the Linux box, I am using 16MB now.  Please run this command: "ulimit -s 16384".

    -Yuan

  • 0 in reply to Yuan Zhao
    TI__Genius 16975 points

    Hi Yuan

    I'd like to know your current progress for this topic.

  • 0 in reply to Tsurumoto
    TI__Expert 4105 points

    Hi HaTa,

        Has the customer tried the recommended stack size increase?

        Can you clarify with a more concrete number for number of images being used for calibration?  I tried 100 images, the calibration succeeded.

        The number of calibration images does not really impact the memory usage during calibration process.  All calibration input data are written into file system, and are read into memory one image at a time during calibration process.

    -Yuan