TDA2EXEVM: How to run a TIDL OD model in a different resolution?

Hi Gary,

>> But the there were nothing detected. The result is as follow

Can you attach this again as I am unable to see the image?

>> I also tested on host emulation, the result displayed correctly.

Can you attach this again as I am unable to see the image?

Thanks,

Praveen

Hi Praveen,

I re-uploaded the missing pictures.

Thank you for your help,

Gary

Hi Gary,

The modified files for 512x512 size looks fine. Did you get proper output from import tool with 512x512 size?

Thanks,

Praveen 

Hi Praveen,

I just tried, the output from import tool was correct.

Will assign layersGroupId in import setting file relate to this issue?

I checked import file for 768x320, it set some layers to DSP.

But I'm not sure which layer should be set to DSP.

Are there any other way to test?

Thank you,

Gary

Hi Gary,

Run only last layer (Detection output layer) in DSP and rest all the layers on EVE.

Thanks,

Praveen

Hi Praveen,

>> Run only last layer (Detection output layer) in DSP and rest all the layers on EVE.

This is the setting I used for all above results.

My layersGroupId setting:

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    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1    1   2    0

Thank you,

Gary

Hi Gary,

Can you try with this below setting for layersGroupId ?

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    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   2    0

Thanks,

Praveen

Hi Praveen,

After changing the layersGroupId , the screen becomes black.

Nothing shows up.

Thank you,

Gary

Hi Gary,

Is this with 521x512 size?

Does the default size 768x320 worked for you on TDA2x?

Thanks,

Praveen

Hi Praveen,

>> Is this with 521x512 size?

Yes.

>>Does the default size 768x320 worked for you on TDA2x?

Yes.

Thank you,

Gary

What modifications you did in the use case for 512x512 size?

Thanks,

Praveen

Hi Praveen,

All modified files are copy from raw TI caffe-jacinto model ( ../ssd512x512_ds_PSP_dsFac_32_fc_0_hdDS8_1_kerMbox_3_1stHdSameOpCh_1/sparse/ )

I list modified parts in below.

deploy.proto

1. Changing  keep_top_k and confidence_threshold to 20 and 0.15 in DetectionOutput layer, respectively.

caffemodel

Nothing changes here, I used raw TI sparse 512x512 model located in above path.

TIDL_SSD_CFG.txt (in SD card)

1. Changing input height/width from 768x320 to 512x512

2. Changing net/param file names to fit the bin file names generated from import tool.

3. Setting fps and threshold  to [10, 0.15], respectively.

chains_tidlOD.c

1. Changing GRAPHICS_HEIGHT  value from (1080-64) to (1080 - 10).

2.  Changing pUcObj->DisplayPrm.rtParams.posY value from 64 to 10.

Note: The recompiled MLO and AppImage work fine with resolution 768x320.

import setting 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       = "D:\Download\19Oct14-18\import_tool_test_clean\deploy_512x512_sparse.prototxt
inputParamsFile    = "D:\Download\19Oct14-18\import_tool_test_clean\voc0712_ssdJacintoNetV2_iter_104000_sparse_512x512.caffemodel
outputNetFile      = "D:\Download\19Oct14-18\import_tool_test_clean\tidl_net_ssdJacintoNetV2_sparse_512x512.bin"
outputParamsFile   = "D:\Download\19Oct14-18\import_tool_test_clean\tidl_param_ssdJacintoNetV2_sparse_512x512.bin"

rawSampleInData = 1
preProcType   = 4
sampleInData = "D:\Download\19Oct14-18\import_tool_test_clean\512x512.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    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   2    0
conv2dKernelType = 0	0	0	0	0	0	0	0	0	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. Changing input & output path.

Here I attach related files I used.

import_tool_test_512x512.rar

Thank you

Gary

Hi Gary,

Thanks for sharing the files, I will get it checked by VSDK experts and come back.

Please expect delays in the response as many people are on leave this week because of local holidays here.

 Regards,

Praveen

Hi Gary,

Can you press p after running the usecase for 1 min.

Please share entire log of this experiment.

Regards,

Anuj