about TIDL to process small ROI

Ahan Tseng

Hi,

I have some questions about TIDL to process small ROI.

On REL.TIDL.01.00.00.00 datasheet, I found that it supports small ROI processing by dense convolution.

When I build the tidl caffe model(net and param binary file) on REL.TIDL.01.00.00.00 library, I set the the "Conv2dKernelType" parameter to control the sparse (0) and dense convolution (1).

For the two settings about the parameter "Conv2dKernelType", I got different results on the same model and I will show the two results.

1. What is the "dense convolution" ?

2. If I want to process the small ROI on TIDL library, I only set the "Conv2dKernelType" parameter.

If yes, why I got different results with the parameter "Conv2dKernelType" to set 0 and 1?

If not, how should I set ?

Conv2dKernelType = 0, sparse convolution

Conv2dKernelType = 1, dense convolution

Thanks and Regards,

Ahan

over 8 years ago

0 Praveen Eppa over 8 years ago

Expert 1420 points

Hi Ahan,

1. Dense convolution means kernel coefficients with no Sparsity. Sparsity indicates the % of kernel coefficients having zero value, so in the dense convolution all the kernel coefficients are non-zero vlaues.

2. If you want to process the small ROI on TIDL library, please set the "Conv2dKernelType" parameter to "1".

Please let me know for any questions.

Thanks,
Praveen

0 Praveen Eppa over 8 years ago in reply to Praveen Eppa

Expert 1420 points

Hi Ahan,

The dense and sparse convolutions are functionally same.
Can you please share the configuration so that we can debug the issue.

Thanks,
Praveen

0 Ahan Tseng over 8 years ago in reply to Praveen Eppa

Prodigy 190 points

Hi Praveen,

About the test model, I tried to use the settings "Conv2dKernelType" to 1, dense convolution.

I got two different results on eve and dsp mode.

I also tried to use two settings of the "Conv2dKernelType" parameter with dense(1) and sparse(0) convolution, respectively, on eve mode.

It also got two different results.

I totally set the four different setttings with the "Conv2dKernelType" parameter and the different mode, dsp and eve.

Finally, I found I got two results on this settings.

(1)

Conv2dKernelType = 1, eve mode.

(2)

Conv2dKernelType = 1, dsp mode.

Conv2dKernelType = 0, eve mode.

Conv2dKernelType = 0, dsp mode.

Can you help me to resolve the question?

Thanks and Regards,

Ahan Tseng

The link is the caffe model and prototxt file for the test model.

https://drive.google.com/open?id=1PkICJr6F4c_ismkN2Mbuk_pFS37yKtTk

0 Praveen Eppa over 8 years ago in reply to Ahan Tseng

Expert 1420 points

Hi Ahan,

We could replicate the issue reported at our end and currently debugging this issue.
We will get back to you with fix soon.

Thanks,
Praveen

0 Praveen Eppa1 over 8 years ago in reply to Praveen Eppa

TI__Genius 17580 points

Hi Ahan,

We fixed this issue and will be sharing the fix in CDDS.

Thanks,
Praveen

0 Praveen Eppa1 over 8 years ago in reply to Praveen Eppa1

TI__Genius 17580 points

Hi,

Fix is available in the below CDDS link
cdds.ext.ti.com/.../emxNavigator.jsp

Please take fixes for bug3 from patch release folder.

Thanks,
Praveen

0 Charles Yeh over 8 years ago in reply to Praveen Eppa1

Intellectual 470 points

Hi,

The "TIDL/Patch_Release" folder is till empty.

Charles

0 Praveen Eppa1 over 8 years ago in reply to Charles Yeh

TI__Genius 17580 points

Hi Charles,

Can you try with this below link ?
cdds.ext.ti.com/.../emxTree.jsp

Thanks,
Praveen

0 Charles Yeh over 8 years ago in reply to Praveen Eppa1

Intellectual 470 points

Hi, Praveen,

We can not access the link.

Charles.

0 Praveen Eppa1 over 8 years ago in reply to Charles Yeh

TI__Genius 17580 points

Hi,

Please let us know the fix (provided in Patch release) resolved your issue or not?

Thanks,
Praveen

0 Ahan Tseng over 8 years ago in reply to Praveen Eppa1

Prodigy 190 points

Thanks!

The issue is already resolved.

0 Ahan Tseng over 8 years ago in reply to Praveen Eppa1

Prodigy 190 points

Hi,

I have a sparsity model and I tried to use the settings "Conv2dKernelType" to 1, dense convolution.

But, I got two different results on eve and dsp mode.

I also tried to use two settings of the "Conv2dKernelType" parameter with dense(1) and sparse(0) convolution, respectively, on eve mode.

It also got two different results.

I totally set the four different setttings with the "Conv2dKernelType" parameter and the different mode, dsp and eve.

Finally, I found I got two results on this settings.

(1)

Conv2dKernelType = 1, eve mode.

(2)

Conv2dKernelType = 1, dsp mode.

Conv2dKernelType = 0, eve mode.

Conv2dKernelType = 0, dsp mode.

Can you help me to resolve the question?

Thanks and Regards,

Ahan Tseng

The link is the caffe model and prototxt file for the test model.

drive.google.com/open

0 Praveen Eppa1 over 8 years ago in reply to Ahan Tseng

TI__Genius 17580 points

Hi Ahan,

We will replicate the issue reported at our end and will get back to you with an update soon.

Thanks,
Praveen

0 Praveen Eppa1 over 8 years ago in reply to Praveen Eppa1

TI__Genius 17580 points

Hi Ahan,

We could replicate the issue. The issue is because of dilation > 1 for dense flow. Currently we don't have support for dilation > 1 for dense flow. I suggest you to use Sparse flow when dilation >1.

Thanks,
Praveen

0 Ahan Tseng over 8 years ago in reply to Praveen Eppa1

Prodigy 190 points

Thanks!

Processors

Processors forum

about TIDL to process small ROI