AM62A Quantition

Hell Shuai Wang,

Thank you for your query. Could you also try this with 16-bit quantization and share the result of that? This is a good first step in diagnosing quantization-related accuracy.

Assuming accuracy is good with 16-bit quantization, we can use hybrid quantization so selectively use 16-bit at critical layers - often the first and/or last convolutional layers. We did a similar analysis when enabling yolov5: https://www.ti.com/video/6286792047001

Please also see the following document as well - debug_level 4 will give floating point and fixed point traces.

https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md#feature-map-comparison-with-reference

Best,
Reese