SK-AM62A-LP: Arc Fault Model

Part Number: SK-AM62A-LP
Other Parts Discussed in Thread: TIDA-010955

Tool/software:

Hello TI Dev,

EdgeAI model composer provides a training platform for arc fault where there are several model selection for arc fault as shown in figure below. I am wondering whether those numbers (200,300,700,1400) are referred to. Do they represent time step?

Sincerely yours,

Key

  • Hello,

    I have brought this thread to the attention of the Edge AI Studio experts. Please note that due to the local holiday today, responses will likely be delayed.

    Thanks

    ki

  • Hello Key,

    The numbers in the model name represent the number of parameters, memory footprint and computational requirements.   See below

    • ArcFault_model_200: smallest memory footprint and computation requirement.
    • ArcFault_model_300: more model parameters for improved performance.
    • ...
    • ArcFault_model_1400: most model parameters for best performance.

    This allows you to trade off between the memory, computational and performance requirements of  your application. 

    More details are available in the TIDA-010955 - Machine Learning based Arc Fault Detection Reference Design  at 

    https://dev.ti.com/tirex/explore/node?node=A__AJmsZHbxSEUwh2-Fo6fUqA__digital_power_c2000ware_sdk_software_package__5C6SEVO__LATEST

    Best regards,

    Martin

  • Hello Martin,

    Does it relate to data preprocessing on input features? If my understanding is correct, it is similar to applying time step on on input features. Short time step will give less computation and parameter, and vice versa. 

    Regards,

    Key

  • Just a quick update that the Edge AI team will respond shortly.

  • Hello Sopheap,

    The numbers (200,300,700,1400) refer to the approximate number of parameters in the model.

    The smallest model (ArcFault_model_200_t) has the smallest number of parameters, which implies they will take the least amount of time to execute (fastest inference time), but the downside is also that since there are only 200 learnable parameters available, the model has limited learning capability. This means that this is best suited for simpler datasets 

    On the other hand, the model_1400_t has about 1400 parameters. It is the slowest to run out of the 4 models, but has the best learning capability. This would best be suited for complex environments. 

    To summarize:

    In terms of model performance (inference time): 200_t(fastest)-->300_t-->700_t-->1400_t(slowest)

    In terms of accuracy (on a very complex dataset): 1400_t(best)>700_t>300_t>200_t(worst)

    On a simpler dataset, you may get a perfect (or nearly perfect) accuracy even on a smaller model (such as 200_t/300_t)

  • Hello Adithya,

    I am aware of this information. What I am curious and not sure is the detail of input tensor shape for the training. Could you please enlighten me on this matter? After training on model composer for all model (200,300,700,1400), a compiled artifact and dataset used for evaluation are provided. After I carefully checked the dataset for evaluation from model composer on EdgeAI studio, I found that the input tensor for all 4 model is having the same dimension (1,1,256,1). Does it mean that model_1400_t consists of 1400 total training parameters? 

    Regards,
    Sopheap

  • Hi Sopheap,

    The model parameters is not dependent on the input tensor size. Any of the models (200/300/700/1400) can accept any input tensor size.

    The input size is determined by the feature extraction parameters chosen on the GUI. (for example, the frame size, binning, feature extraction method etc)

    The model parameters on the other hand is constant.