SK-AM62A-LP: Custom Model

Hello,

Yes, we have documentation and guidance for this topic. Many user guides and documentation pages for custom models is on the edgeai-tidl-tools repository on Github.

Please see here for custom models: https://github.com/TexasInstruments/edgeai-tidl-tools?tab=readme-ov-file#compile-and-benchmark-custom-model 

Let me know if you still have questions! For putting together and end-to-end imaging pipelines with a live video source (file, camera, internet stream), please see the Edge AI SDk documentation: https://software-dl.ti.com/processor-sdk-linux/esd/AM62AX/10_00_00/exports/edgeai-docs/common/edgeai_dataflows.html 

BR,
Reese

Hi Reese,

Thanks, These resources shows how to run a single model or few models in parallel. Do we have some resource in which explained how to use models one after the another?

Regards,

Marvin

Hi Marvn,

Yes, these resources show how to run a model, and they support running multiple models (for compilation or inference) in parallel. They do not show sequentially run models, largely because there is not anything special to account for when doing so.

For multiple models in sequence, the same APIs will be used. There is no restriction on the order in which models are initialized or called, so long as their weights fit into the device's RAM/LPDDR4.

To run multiple models one after another; 

• initialize each model using the preferred API, e.g. ONNX Runtime
• Within a loop
  • preprocess images as needed
  • Pass the image as input to the model with the DL runtime API, e.g. ONNX Runtime
  • If the output of one model feeds directly into the next, retrieve that output from the runtime, apply any further processing/transformation, and pass into the next model

The multiple sequential model use-case is not directly included in our examples. It is left to the developer to implement this. Our model zoo does not include models that require sequential operation in this way (e.g. Region proposal network --> classification network for 2-stage object detection). Instead, we support single-shot detection models. 

BR,
Reese