EDGE-AI-STUDIO: model format and how level of optimisation for the SoC?

GR,

For your first question, the models provided by Texas Instruments in Edge AI Studio are in ONNX format.  Downloading in other formats is not supported.

We will follow up on your other questions soon.

Thanks,

Martin

Hello GR,

Thank you for your queries. I'll respond to ones that Martin hasn't already addressed.

GR said:

Question2:

If we train and compile model in edge-AI-studio,

How level of optimised for C7xDSP?

If you can be more specific in what you mean by 'optimization level', that would be helpful.

Our TI Deep Learning software runs on this C7xMMA core (MMA = matrix multiply accelerator, and is where the 2 TOPS performance number comes from). Each layer we support for neural networks has been optimized to use the data movement mechanisms and MMA core -- most of these have been hand-optimized by our development team. Can you clarify what you mean by your question?

GR said:

Question3:

additional to question2

If there is anything to data looks like : 

  edge-AI-studio'model : including ~％ float, ~％ int

  model trained/compiled another OSS : including ~％ float, ~％ int

If I understand correctly, you want to see data about the accuracy of the floating point vs. fixed point version of our models, as well as the original OSS versions. The second point is relevant since we will make minor changes to model architectures to better utilize the accelerator.. These changes models should have a 'ti-lite' designator within the name. 

You can find information out the models we host in our model_zoo on that repository, but not all have fixed vs. floating point values. If it does not specify fixed or floating point, then it is safe to assume floating point. We aim for <2% accuracy loss for 8-bit fixed point, but most have smaller than 2% loss. 

https://github.com/TexasInstruments/edgeai-modelzoo/tree/main/models/vision/classification <-- specific page for classification models; there are other types in the parent directory.

https://github.com/TexasInstruments/edgeai-yolox/blob/main/README_2d_od.md <-- YOLOX nano information

https://github.com/TexasInstruments/edgeai-yolov5 <-- YOLOv5

I see that most of these are floating point, because they focus on the base model, rather than the implemntation on the SoC. You may find performance and accuracy values for models on a per-device basis at https://dev.ti.com/edgeaistudio/ -> Model Analyzer (login) -> Select AM62A -> Model Selection tool (left pane)

BR,
Reese