Linux/TDA2EVM5777: Self converted JDetNet model doesn't show same FPS with pre-installed JDetNet model

Byeongheon Yoo

Expert 1240 points

Part Number: TDA2EVM5777

Tool/software: Linux

Hi,

I downloaded JDetNet model from github and converted that model.

In TIDL OD usecase, the installed JDetNet model that is contained in Vision SDK showed about 20 FPS.

However, the converted JDetNet model shows about 10 FPS and file size was different.

Moreover, the detected box of vehicle was pink color. The color of vehicle was green when is tested with original model.

I also changed objectDraw link as related post says.

The path of downloaded deploy.prototxt is 'ti_dl\test\testvecs\config\caffe-jacinto-models\trained\object_detection\voc0712\JDetNet\ssd768x320_ds_PSP_dsFac_32_hdDS8_0\test_quantize' and the name of pretrained model is voc0712_ssdJacintoNetV2_iter_120000.caffemodel.

Is it normal situation?

Regards,

Yoo

over 7 years ago

0 Praveen Eppa1 over 7 years ago

TI__Genius 17580 points

Hi Yoo,

Did you update the below parameters in the “deploy.prototxt” as indicated in the original thread

keep_top_k: 20

confidence_threshold: 0.15

Thanks,

Praveen

0 Byeongheon Yoo over 7 years ago in reply to Praveen Eppa1

Expert 1240 points

Hi Praveen,

Yes, i did it. I already did all things that were refered in original question. (https://e2e.ti.com/support/processors/f/791/t/689617)

Is the jetnet model that is provided when i install the vision sdk from the jdetnet model which is opened in caffe-jacinto git repo?

Regards,

Yoo

0 Praveen Eppa1 over 7 years ago in reply to Byeongheon Yoo

TI__Genius 17580 points

Hi Yoo,

The jdetnet model that is provided when install the vision sdk is not same as one in the caffe-jacinto git repo. The jdetnet comes with VSDK had only 4 classes and the one in caffe-jacinto had 21 classes.

Thanks,

Praveen

0 Byeongheon Yoo over 7 years ago in reply to Praveen Eppa1

Expert 1240 points

Hi Praveen,

Thanks for your useful information.

I got a one more question about whether you used any optimization method in training JDetNet model that is provided in vision sdk.

I used tidl_import_JDetNet.txt file that is provided in vision sdk.

However, as i say above post, the FPS was slower 2 times when i converted JDetNet caffe model from git repo (provided JDetNet 20 FPS, self converted JDetNet 10 FPS).

Is that caused only because of difference of number of classes?

Regards,

Yoo

0 Byeongheon Yoo over 7 years ago in reply to Praveen Eppa1

Expert 1240 points

Any updates here?

I also wonder whether i can get deploy.prototxt and recipe txt file of provided JDetNet.

Regards,

Yoo

0 Praveen Eppa1 over 7 years ago in reply to Byeongheon Yoo

TI__Genius 17580 points

Hi,

Yes, the degradation could be because of difference of number of classes.

I am not sure that I can share that publicly here, but only difference is if had four classes.

Thanks,

Praven

Processors

Processors forum

Linux/TDA2EVM5777: Self converted JDetNet model doesn't show same FPS with pre-installed JDetNet model