Tool/software:
Tool/software:
The sdk version I use now is ti-processor-sdk-rtos-j784s4-evm-09_02_00_05
There are differences in the quantization coefficients between SDK 9.2 and SDK 8.6, as shown in the figure below. The scale values of the same model after quantization are different, which are quite significant. Is this a normal phenomenon?
We have provided the data of the issue, and we hope it can be resolved as soon as possible.
Hi Guang,
Based on the data you provided and my beyond-compare review, the output looks OK. Version to version will have some differences in scale from layer to layer, probably even from run to run, as there is a certain entropy factor with ML. Is the model not performing adequately? If so, may I look at our model or the layer-by-layer outputs (float/non-float) with a debug set to 4?
I have attached my BC output and other than two extra dimensions added from 8.6 to 9.2; nothing looks like it is off in the weeds.
Again, I will need to see the model or the layer outputs. Unless the model is not performing correctly, this looks fine. Something to point out is that the model inference time may be a few percent slower in 9.2 from 8.6. There are patches to address the inference time, but the accuracy should be approximately the same.
Regards,
Chris
