TDA2: Running caffe-jacinto object detection using Vision SDK 3.03 TIDL_OD usecase

Hi Eric,

Thanks for the inputs. After setting "keep_top_k" to 20 and "confidence threshold" to 0.15, I am able to run both the initial and sparse model at 3fps.

When I tried to change the FPS of sparse model to 6, again there is no stream(doesn't work for 5fps also).
In the caffe jacinto github, it is mentioned that sparse model achieves 2.5x speedup, so I thought sparse model should run at least 2x as initial model. Kindly let me know your inputs on this.

Regards,
Navinprashath.R.R

Hi Eric,

Thanks for the link. I saw this thread earlier. When running in FILE IO Mode, either EVE or DSP is selected (Also in TIDL OD usecase the whole model is configured to run only on 4 EVE cores).
So how these changes on setting layersGroupId for extra layers will take effect?
Do these layers run on DSP, when layersGroupid configured as 2 irrespective of EVE or DSP being selected in usecase?

Regards,
Navinprashath.R.R

Hi Navin,

I am not sure I am correct.
I read some of the TIDL source.
Most of the layer implemented by TI can run both on the EVE and DSP.
But the performance may be very different.
For example, the floating point operation should run on DSP for better performance.
So yes for OD usecase , if you set the layergroup to 2,it will run on the DSP.
(i think the layer group parameter may stored in the NET.BIN
and in the usecase it will connected to the TIDL library and run on specific core)
If there is any mistake , please correct me.

Best Regards,
Eric Lai

Hi Eric,

Thanks for the inputs. For which usecase did you get 10 fps,Object detection using FILEIO usecase or TIDL OD usecase?

Regards,
Navinprashath.R.R

Hi Navin,

I run the OD usecase with the caffe-jacinto 512x512 model.

Best Regards,
Eric Lai

Hi Eric,

I replaced layersGroupId and conv2Dkerneltype with the one shared on the thread you have shared. But I'm getting error when importing. The import tool crashes.

Here is my import file:


# Default - 0
randParams = 0

# 0: Caffe, 1: TensorFlow, Default - 0
modelType = 0

# 0: Fixed quantization By tarininng Framework, 1: Dyanamic quantization by TIDL, Default - 1
quantizationStyle = 1

# quantRoundAdd/100 will be added while rounding to integer, Default - 50
quantRoundAdd = 25

numParamBits = 8
# 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType = 0

inputNetFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\object_detection\voc0712\JDetNet\sparse\deploy.prototxt
inputParamsFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\object_detection\voc0712\JDetNet\sparse\voc0712_ssd.caffemodel"
outputNetFile = "..\..\test\testvecs\config\tidl_models\sparse\NET_OD.BIN"
outputParamsFile = "..\..\test\testvecs\config\tidl_models\sparse\PRM_OD.BIN"

rawSampleInData = 1
preProcType = 4
sampleInData = "..\..\test\testvecs\input\tsr_512_512.y"
tidlStatsTool = "..\quantStatsTool\eve_test_dl_algo.out.exe"
layersGroupId = 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 2 2 2 0
conv2dKernelType = 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

Regards,
Navinprashath.R.R

Hi Navin,

How it crash?
I think you should ask TI expert about this part.
Maybe the number of class is too many?
For me, my model is to detect 4 classes.
and the import is okay.
How about just replace the layersGroupId and conv2dKernelType with your own model?

Best Regards,
Eric Lai

Hi Eric Lai,

Thanks for your help here..

Regards,
Praveen