Linux/TDA2PXEVM: Regarding TIDL training

Heechang Kim

Expert 2700 points

Part Number: TDA2PXEVM

Tool/software: Linux

Dear,

I used usecase of the "Object Detection" by TIDL.

But, it cannot detect person.

So, I'd like to improve detection.

How can I do to improve detection for TIDL?

Is there any tools for training the TIDL?

If there is, please let me know the process of the training.

Best regards,

Heechang

over 6 years ago

0 Heechang Kim over 6 years ago

Expert 2700 points

In SDK, there are some file like regarding TIDL, vision_sdk/apps/tools/TIDL_files.
But, I cannot find any document.

0 Anuj Gupta over 6 years ago in reply to Heechang Kim

TI__Genius 15985 points

Hi,

Please go through the below userguide.

ti_components\algorithms\REL.TIDL.01.01.03.00\modules\ti_dl\docs\TIDeepLearningLibrary_UserGuide.pdf

Regards,
Anuj

0 Heechang Kim over 6 years ago in reply to Anuj Gupta

Expert 2700 points

Hi,

I'm checking the document in "ti_components\algorithms\REL.TIDL.01.01.03.00\modules\ti_dl\docs\TIDeepLearningLibrary_UserGuide.pdf"

In 3.6.4, I downloaded "caffe-jacinto-models".

I run the "tidl_model_import.out.exe tidl_import_j11_v2.txt".

But, I cannot open the parameter of input file, which is "imagenet_jacintonet11v2_iter_160000.caffemodel".

It exists correct path.

Best regards,

Heechang

0 kumar.desappan over 6 years ago in reply to Heechang Kim

TI__Mastermind 22145 points

Please update the path of "imagenet_jacintonet11v2_iter_160000.caffemodel" in " tidl_import_j11_v2.txt". You can provide absolute path if you are finding issues with relative path.

0 Heechang Kim over 6 years ago in reply to kumar.desappan

Expert 2700 points

Hi,

Thanks.

I have a question.
I'm using your library but it cannot detect pedestrians well.
So, I think that "tidl_algo.lib" should be updated, right?

How can I create the "tidl_algo.lib"?
If I improve the detecting the pedestrians, what input images use?
Should I use the CCS?

I don't know well cause I'm not deep-learning engineer.

BR,
Heechang

0 kumar.desappan over 6 years ago in reply to Heechang Kim

TI__Mastermind 22145 points

TIDL library is not specific to any algorithm (Like detection or segmentation). it is generic CNN network acceleration software.
Please improve your accuracy of CNN network by training with more data to improve accuracy of any algorithm.
You can use caffe-jacinto (github.com/.../caffe-jacinto) to train your models. Refer the documentations in the same for more information on training.

0 Heechang Kim over 6 years ago in reply to kumar.desappan

Expert 2700 points

Hi,

In section 3.6 by TIDL Userguide, if I use caffe model, then the output of the "*.bin" are 2 files(net and parameter file).
After I trained and update the “{TIDL_install_path}\test\testvecs\config config_list.txt”, if I build the vision_sdk for app.out for object detection usecase then the app.out file is updated which is trained?

Best regards,
Heechang

0 kumar.desappan over 6 years ago in reply to Heechang Kim

TI__Mastermind 22145 points

Please go throught the user guide completely.
1. Update import config
2. Run the import tool with the updated config file to generate TIDL models.
3. Update vision_sdk config filw with Generated TIDl model file to use the newly trained model in the app

0 Heechang Kim over 6 years ago in reply to kumar.desappan

Expert 2700 points

Hi,

1. Update import config
-> tidl_import_j11_v2.txt - input net and param files are caffe-jacinto-models.

2. Run the import tool with the updated config file to generate TIDL models.
-> tidl_model_import.out.exe tidl_import_j11_v2.txt - output files are *.bin

3. Update vision_sdk config filw with Generated TIDl model file to use the newly trained model in the app
-> “{TIDL_install_path}\test\testvecs\config config_list.txt”

Is this right?
Then, is it OK to build vision_sdk for app.out of the newly trained model?

Best regards,
Heechang

0 kumar.desappan over 6 years ago in reply to Heechang Kim

TI__Mastermind 22145 points

First two steps are right. you 3rd step is for standalone Test application.
For vision_sdk app, the 3rd step shall update use case config file.
Anuj, can you point the right use case config file

0 Heechang Kim over 6 years ago in reply to kumar.desappan

Expert 2700 points

Hi Aunj,
Please let me know about the 3rd step.
BR,
Heechang

0 Anuj Gupta over 6 years ago in reply to Heechang Kim

TI__Genius 15985 points

Hi,

You just need to place the new updated model files in sd card.

Please go through the chapter 5 of below doc.

vision_sdk\docs\FeatureSpecificUserGuides\VisionSDK_UserGuide_TI_DeepLearning.pdf

Regards,

Anuj

Processors

Processors forum

Linux/TDA2PXEVM: Regarding TIDL training