Hi,
I'm trying to use different quantization bit when importing models.
And I found the document:
http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/06_01_01_12/exports/docs/tidl_j7_01_00_01_00/ti_dl/docs/user_guide_html/md_tidl_model_import.html
It indicates these two terms:
numParamBits:8bit,Bit depth for model parameters like Kernel, Bias etc. Max supported is 16. Any value > 8 is experimental in the current release.
numFeatureBits:8bit,Bit depth for Layer activation; Max supported is 16. Any value > 8 is experimental in the current release.
I just want to verify if numParamBits is for model parameter and numFeatureBits is for layer feature map value quantization.
Also, these two parameters are set in import.txt.
For example :
numParamBits = 16
numFeatureBits = 16
Other pages mention that "16-bit and 32-bit will come with significant performance penalty".
Shall we use 8-bit quantization rather than 16-bit in a usual case?
Thank you,
Kevin.
