TIDL 9.1 and 9.2 support transformer models and DeiT and DETR is provided as examples. Are there any support for ViT-S/B and what latencies should be expected for those models ?
This thread has been locked.
If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.
TIDL 9.1 and 9.2 support transformer models and DeiT and DETR is provided as examples. Are there any support for ViT-S/B and what latencies should be expected for those models ?
Hi,
Due to a holiday, half of our team is currently out of the office. Please expect a 1~2 delay in responses.
Apologies for the delay, and thank you for your patience.
-Fabiana
Hi,
We have only validated DeiT and DETR models end to end using timm based lib
And we support different transformer modules like attention, patch enabling etc and Window shifting and patch merging specific to SWIN.
I recommend to do diligence of TIs transformer support here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_06_00/docs/tidl_fsg_vtfr.md
Thank You
Hi Roman,
Is there any intention for ViT ?
You can find the roadmap of our transformer model support here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/09_02_06_00/docs/tidl_fsg_vtfr.md
And what are the latencies for DETR?
Do you mean benchmark ?
Do you mean benchmark ?
Yes, I would like to know which fps I should expect in a real-time system (typical time of inference) Also, accuracy degradation due to quantization would be great.
Hi,
The transformer model benchmarks are not available yet in public domain.
Per my internal alignment team is working on it, it will shared in public soon.
Thanks