• State TI Thinks Resolved
  • Locked Locked
  • Replies 1 reply
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 102 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:

PROCESSOR-SDK-AM62A: Can I stream 3280x2464 resolution video through apps python

PROCESSOR-SDK-AM62A: Model training above the resolution 224x224

Sajan Saji
Intellectual 510 points
Part Number: PROCESSOR-SDK-AM62A


Tool/software:

Hello Experts,

I need to train the model above 224 resolution. Does that improve the accuracy of output.

Below information is from conversation with Qutaiba in e2e.

For models which are not supported by ModelMaker, you can use any other non-TI tool such as pytorch to conduct the training. When done, you can use the edgeai-tidl tool to port/compile the trained model to work on AM62A.

Can you please give me the necessary step I should follow.

Warm Regards,
Sajan

  • 0
    TI__Genius 15686 points

    Hi Sajan,

    This is not a TI-specific topic, and you can find many online resources for training models and learning about the impact of resolution on accuracy. I will provide a cursory response.

    Sajan Saji said:
    train the model above 224 resolution. Does that improve the accuracy of output.

    Generally speaking, yes, at the cost of increased computation requirements. There will be diminshing return's w.r.t the tradeoff of compute requirements vs. accuracy.

    Sajan Saji said:
    Can you please give me the necessary step I should follow.

    This is outside TI's support provisions for such open-source tools. Please see PyTorch documentation and examples for this. You will need to export into ONNX when training is complete.

    I will note that there is a TIDL wrapper [0] for pytorch that applies optimizations to replace unsupported layers (or ones that are generally difficult for 8-bit quantization) with supported layers. The workflow here is to train the baseline, source model until accuracy converges, apply the wrapper to replace layers, and then continue training (reduced number of epochs / until convergence).

    [0] https://github.com/TexasInstruments/edgeai-tensorlab/blob/main/edgeai-modeloptimization/torchmodelopt/edgeai_torchmodelopt/xmodelopt/surgery/README.md