EDGE-AI-STUDIO: CUDA-compiled TIDL

Ahmed Mujtaba

Hi,

I am trying to convert the "onnx" model to generate artifacts for J721S2 processor using EdgeAI Benchmark. In the documentation of EdgeAi Benchmark, it is mentioned that if you have cuda compiled tidl on pc, then you can generate artifacts in parallel number of devices. Can you specify where I can get the CUDA-compiled artifacts?

Additionally, I cannot find any documentation or set a guideline related to the artifacts generation process. There is no concrete number of calibration images, iterations and bits for quantization. The overall artifacts generation process seems a little tedious and random. It would be very beneficial if I can find any documentation for the artifacts generation process.

Thanks,

Ahmed

over 2 years ago

0 Pratik Kedar over 2 years ago

TI__Mastermind 24041 points

Hi,

If you have pretrained model, you can use edgeai tidl tools for model compilation, please refer the extensive documentation here : https://github.com/TexasInstruments/edgeai-tidl-tools

Let me recheck the support on cuda compiled tools and get back to you.

0 Ahmed Mujtaba over 2 years ago in reply to Pratik Kedar

Prodigy 235 points

Thanks for the reply, Pratik.

Also, can you check about how to upload the annotated dataset upload in EdgeAI Studio? In video demos, I can see that the format should be compressed (.zip, tar.gz, etc.) but I don't know the internal directory structure. Therefore, wheneever, I try to upload the .zip file, it gives me error that annotated dataset is not found.

Thanks,

Ahmed