Part Number:TDA2EXEVM
Tool/software: TI-RTOS
Hi TIDL experts,
I am using TIDL OD usecase, and I have SSD 512x512 Jacinto-based model.
I modified it from the original TI provided model, and my model now contains more layers.
And also it detect more classes, original model is 4 but now I modified it to 5 to detect one more class.
In my experiment on TDA2x, original 512x512 model can run at 13 FPS without sparsity, and can run at close to 24 FPS with sparsity at 70%.
It is almost two times faster.
But new model, can run at 11 FPS without sparsity, but can only run at 14 FPS with sparsity at 70%
It only increase a little, I expected it to around 20.
I conclude the information below:
Original TI provided model 512x512(4 classes, 6 heads):
(1) MAC (from import tool): 2.6 giga MAC
(2) FPS without sparse : 13FPS
(3) FPS with sparse : 24 FPS
New model (5 classes 7 heads):
(1) MAC (from import tool): 3.4 giga MAC
(2) FPS without sparse: 11FPS
(3) FPS with sparse : 14FPS
EVE frequency 635 MHz, DSP 750 MHz
I do some experiments on "conv2dKernelType" parameters in config file.
I follow the rule that only the feature maps bigger than 32x32 use sparse convolution, and change it to see the difference on FPS.
In new model, it does not influence a lot by this parameter, eve though the sparsity of specific layer is pretty high.(80 %)
Any possible reason that the increase of FPS only a little?
Is that the limitation of EVM?
Please give me some suggestion.
Thank you so much.
Best Regards,
Eric Lai