SSD based TIDL OD use case output for 21 classes

Gouri Patil

Hello,

I have validated the SSD based ti_dl Object Detection output with pre-trained model for 21 classes.

And I observed that, the bounding box/rectangle of object detection is not consistent.(It is not continuously detecting the object even on the same input image).

Please find attached video file for your reference.

Please guide to find out the reason.

tidl_OD_Output21Class.zip

Thanks & Regards,

Gouri

over 7 years ago

0 Praveen Eppa1 over 7 years ago

TI__Genius 17580 points

Hi Gouri,

This could be an issue with quantization parameters setting.

Please refer to section 3.8 Matching TIDL inference result in the user guide and try different parameters that controls quantization and check.

Thanks,
Praveen

0 Gouri Patil over 7 years ago in reply to Praveen Eppa1

Intellectual 505 points

Hello Praveen,

I have gone through the section 3.8- Matching TIDL inference result in the user guide.
So, I want to know that is there any script for generating layer by layer Caffe inference?

Regards,
Gouri

0 Praveen Eppa1 over 7 years ago in reply to Gouri Patil

TI__Genius 17580 points

Hi Gouri,

I don't have any script to generating layer by layer caffe inference.

Thanks,
Praveen

0 Gouri Patil over 7 years ago in reply to Praveen Eppa1

Intellectual 505 points

Hello Praveen,

Then how TI itself validate this result? And if I want to validate then How should I Validate?

Regards,
Gouri

0 Praveen Eppa1 over 7 years ago in reply to Gouri Patil

TI__Genius 17580 points

Hi Gouri,

I will check with team and come back.

Thanks,
Praveen

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

TI__Genius 17580 points

Hi Gouri,

Please find the script in the below link, but it is not up to date so use it is a reference.

github.com/.../infer_blobs.py

Thanks,
Praveen

0 qiuhui jiang over 7 years ago in reply to Praveen Eppa1

Intellectual 260 points

hi,

I want to ask which guide you refer , when you run a network model，I not find this doc in vsdk.

and I have downloaded the PROCESSOR_SDK_VISION_03_04_00_00 for TDA2XX,

could you help me, thanks.

0 Praveen Eppa1 over 7 years ago in reply to qiuhui jiang

TI__Genius 17580 points

Hi,

The mentioned user guide (TIDeepLearningLibrary_UserGuide.pdf) is in TIDL release under below folder in VSDK,
"ti_components\algorithms\REL.TIDL.01.01.01.00\modules\ti_dl\docs"

Thanks,
Praveen

0 qiuhui jiang over 7 years ago in reply to Praveen Eppa1

Intellectual 260 points

thanks very much~
Do you use the CCS?

0 Praveen Eppa1 over 7 years ago in reply to qiuhui jiang

TI__Genius 17580 points

Hi,

This question is answered in the below thread, so I am closing this thread. Please post your followup questions in the below thread.
e2e.ti.com/.../712872

Thanks,
Praveen

0 Gouri Patil over 7 years ago in reply to Praveen Eppa1

Intellectual 505 points

Hello Praveen,

I have gone through that script. By doing some modifications I am able to generate some layers .bin files.

But I am bit confused in following points:

1. While importing the attached .prototxt and .caffemodel on board using the visionSDK import tool, it generates layer by layers 44 trace_dump.y files.

And infer_blob.py script generates some layers . bin files. For matching tidl inference with the Caffe inference result, I need same 44 .y files of Caffe inference.

So, for that which layers names should I specify from deploy.prototxt to get the same results? And how should I convert Caffe inference .bin files to .y files.

Please find attached .prototxt, .caffemodel, updated infer_blobs.py and input image.

Please guide.

Matching_inference_result.zip

Regards,

Gouri

0 Praveen Eppa1 over 7 years ago in reply to Gouri Patil

TI__Genius 17580 points

Hi Gouri,

As TIDL import tool merges some layers processing to speed up the execution. So, each layer in caffe output does not have one to one mapping with the TIDL imported layers. Therefore refer to import tool log to get the layer names to map to caffe layers.
Also, refer to FAQ 9 in the user guide for more details on this layers merging.

TIDL can also dump layer by layer outputs in floating point also to easily compare with caffe outputs, so for that define "ENABLE_FLOAT_TRACE" to 1 in the test bench (tidl_tb.c) file.

Thanks,
Praveen

0 Gouri Patil over 7 years ago in reply to Praveen Eppa1

Intellectual 505 points

Hello Praveen,

From import tool log I am getting layer names to map to Caffe layers. I have added all the respective layers in the script and trying to generate .bin file but it throws an error
" IOError: [Errno 2] No such file or directory: 'ctx_output5/relu_mbox_loc_flat_uint8.bin"(because of '/' in between layer name).
so, what should I do to overcome this issue?

To dump layer by layer outputs in floating point if I define "ENABLE_FLOAT_TRACE" to 1 then where I use this macro specifically to get output in floating point.

Please find previously attached deploy.prototxt and infer_blob.py.

Regards,
Gouri

0 Gouri Patil over 7 years ago in reply to Gouri Patil

Intellectual 505 points

Hello Praveen,

Any update on this?

0 Praveen Eppa1 over 7 years ago in reply to Gouri Patil

TI__Genius 17580 points

Hi,

>> " IOError: [Errno 2] No such file or directory: 'ctx_output5/relu_mbox_loc_flat_uint8.bin"(because of '/' in between layer name).
so, what should I do to overcome this issue?

I am not sure about this error, I am checking with experts on this .

>> To dump layer by layer outputs in floating point if I define "ENABLE_FLOAT_TRACE" to 1 then where I use this macro specifically to get output in floating point.

Just #define ENABLE_FLOAT_TRACE (1) in tidl_tb.c file, will dump the TIDL outputs in floating point instead of fixed point, which will be easy to match with caffe reference floating point outputs.

Thanks,
Praveen

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

TI__Genius 17580 points

Processors

Processors forum

SSD based TIDL OD use case output for 21 classes