SK-AM62A-LP: TIDL quantized model - image classification

Cristian Patru

Part Number: SK-AM62A-LP

Tool/software:

I am currently working with the AM62A processor and aim to utilize its deep learning capabilities for image classification tasks. Specifically, I need to run "tidlrt_classification" with a quantized model on this processor.

I have the following questions:

Running "tidlrt_classification" on AM62A: - Is it possible to run "tidlrt_classification" on the AM62A deep learning accelerator with a quantized model?

- If so, could you provide detailed instructions or documentation on how to set up and execute this?

Precompiled Models for Classification: - Are there any precompiled int8 models available for classification use cases that are compatible with the AM62A?

- Specifically, are there precompiled versions of popular models like MobileNet v1 or v2 that I can use?

20 hours ago

0 Reese Grimsley 17 hours ago

TI__Expert 7606 points

Hello,

Yes, this processor supports image classification. I would recommend against the TIDL_RT interface, and instead opt for APIs through Tensorflow-Lite or ONNX Runtime. These are better supported and have more useful features like offloading unsupported layer types to the Arm CPU cores.

Please find our Edge AI documentation on Deep Learning models. Our model zoo has several pretrained and quantized classification models to choose from. These do include mobilenetv1 and mobilenetv2

- https://software-dl.ti.com/processor-sdk-linux/esd/AM62AX/09_02_00/exports/edgeai-docs/common/inference_models.html

You can run these models with sample applications the target SOC

- https://software-dl.ti.com/processor-sdk-linux/esd/AM62AX/09_02_00/exports/edgeai-docs/common/sample_apps.html

For more information about compiling your own models, I would recommend looking at the edgeai-tidl-tools repo:

- https://github.com/TexasInstruments/edgeai-tidl-tools

BR,
Reese

Processors

Processors forum

SK-AM62A-LP: TIDL quantized model - image classification