• State
  • Locked Locked
  • Replies 5 replies
  • Subscribers 98 subscribers
  • Views 299 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: TDA4VM

TDA4VM: The TDA4's addition layer outputs feature maps with mosaic artifacts.

tianlong Tan
Prodigy 10 points
Part Number: TDA4VM


When we export the model to ONNX and deploy it on the TDA4 board (0804 version), we noticed inconsistencies between the feature maps generated on the x86 side and those on the board side. Upon comparing each layer of the network, we identified the issue occurring during the element-wise addition on the board side. After the addition operation, the feature maps exhibited a significant amount of mosaic artifacts.

  • 0
    TI__Mastermind 24041 points

    Hi,

    There could be possibility of issue that you have described above.

    Could you able to throw some light on how are you comparing the results ?

    x86 OSRT float vs TIDL fixed point ? 

    There could be possibility of issue as its older sdk, is it possible to try out the same experiment with latest tool chain, we have added significant amount of feature addition and fixed on latest 9.2 tool chain.

    Look forward to your observations on 9.2 sdk 

  • 0 in reply to Pratik Kedar
    Prodigy 10 points

    Sorry, I might not have made it clear.

    The x86 marked in the image refers to the int8 feature map generated by the TIDL importer. The TDA4 in the image refers to the feature map saved by the TIDL node running on the board end. We compare the results of these two, and find a significant difference after the add layer.

    Due to company plans, we can only proceed with this on 0804.

  • 0 in reply to tianlong Tan
    TI__Mastermind 24041 points

    Thanks for making the test procedure clear.

    Here is my understanding, the results are being compared int8 host emulation vs target for above experiment, for eltwize add operator mentioned above you are observing nuances, correct me incase.

    tianlong Tan said:
    Due to company plans, we can only proceed with this on 0804.

    I understand that, however my suggestion was try out (and not migrate) the same model inference on 9.2 sdk tools and not to migrate as you have the setup ready at your end.

    There could be one of the two things that could possibly come as conclusion.

    1. This issue could have resolved in latest 9.2 sdk release (we can have next level of discussion post this)

    2. Issue still visible on latest 9.2 sdk (i can investigate for RCA and can be taken further)

  • 0 in reply to Pratik Kedar
    Prodigy 10 points

    We don't want to switch from TIDL SDK 0804 to the latest version 9.2. I might not have been clear enough, which may have confused you. I don't think the differences I mentioned are caused by TIDL SDK. What we want to know is why the results on the end are different from the TIDL results. In end inference, there is a mosaic problem in the feature map, as seen in the top left corner of the image. We hope you can help us locate the issue in 0804 or provide some debugging ideas.

  • 0 in reply to tianlong Tan
    TI__Mastermind 24041 points

    Hi,

    tianlong Tan said:
    In end inference, there is a mosaic problem in the feature map, as seen in the top left corner of the image. We hope you can help us locate the issue in 0804 or provide some debugging ideas.

    As mentioned earlier there could be possibility of functional miss match for eltwise add layer in 8.4 sdk version.

    It would be easier for us if could consider sharing observations on,

    Pratik Kedar said:

    There could be one of the two things that could possibly come as conclusion.

    1. This issue could have resolved in latest 9.2 sdk release (we can have next level of discussion post this)

    2. Issue still visible on latest 9.2 sdk (i can investigate for RCA and can be taken further)

    Thank you and look forward to your reply on above experiments.