TDA2EXEVM: TIDL OD usecase failed with new model

jerrytte

Part Number: TDA2EXEVM

Hi,

I have a SSD based model with 157 layers network(1.38G MACs), and the same problem happens like this issue https://e2e.ti.com/support/arm/automotive_processors/f/1021/p/705745/2604130?tisearch=e2e-sitesearch&keymatch=prototxt#2604130

There is no output, and the monitor is black.

And I see the root cause is that enabling dense convolution for some specific convolution layers causing this crash.

So could you please provide more inflammations about the usage limitation of dense convolution and sparse convolution, maybe it helps for my problem.

By the way, all my convolution settings obey the following guide

– We have tested the kernel size up to 7x7 (Shall work for higher values also, but not validated)
– Dilation is tested with 1,2,4.
– We support only stride 1 and 2. Any value higher than 2 is not supported.
– Dense convolution flow is supported for only 1x1 and 3x3 kernels with stride = 1 and dilation =1
– Maximum number of input and output channel supported in 1024

-Use dense convolution for Convolution layers with width x height < 64x64, as dense convolution is optimal for small resolutions and this value is ignored for non-conv layers.

Please give me a help

Thanks

over 7 years ago

0 Praveen Eppa1 over 7 years ago

TI__Genius 17580 points

Hi,

We found that enabling dense convolution flow for convolution layers with small number of channels (<32) is creating this issue. So I would suggest to enable dense conv flow for layers with more number of channels (>=32).

I can help you in checking this setting of dense or sparse flow in your import config for your network if share your prototxt and model along with the modified import config file.

Thanks,
Praveen

0 jerrytte over 7 years ago in reply to Praveen Eppa1

Prodigy 220 points

Hi Praveen,

I have tried to modify the import file, but it seems no work.

So please help me to have a check, thank you very much.

Best regards,

Jerry

ssd576x288_files.tar.gz

0 Praveen Eppa1 over 7 years ago in reply to jerrytte

TI__Genius 17580 points

Hi Jerry,

It looks like there is some issue with your model, it is not a problem of enabling dense flow.

To confirm that, I disabled dense flow in the import config and enabled sparse flow for all the layers by setting all the values to zero. But still I did not see any detection's in the output.
So, can you double check your model and prototxt ?

Thanks,
Praveen

0 jerrytte over 7 years ago in reply to Praveen Eppa1

Prodigy 220 points

Hi Praveen,

Thanks for your checking.

But I test the model with pycaffe in pycharm, it works well.

And I also check the trace dump generated by tidl_model_import.out.exe, it seems OK for convolution output.

Do you have any advice for checking my issue? Thanks

Best regards,

Jerry

ssd576x288_issue.tar.gz

0 Praveen Eppa1 over 7 years ago in reply to jerrytte

TI__Genius 17580 points

Hi Jerry,

Please note that TIDL supports only networks trained with caffe-jacinto frame work.

And for debugging, you can refer to section 3.8 (Matching TIDL inference result) in the user guide.

Thanks,
Praveen

0 jerrytte over 7 years ago in reply to Praveen Eppa1

Prodigy 220 points

Hi Praveen,

I train the network with caffe-jacinto, and the model file works well in pycharm.

I check my network layer with user guide, and find nothing that violate the user guide.

I think there is something mismatched between my network with TIDL, but this mismatch is unknown that is not presented in the user guide

0 Praveen Eppa1 over 7 years ago in reply to jerrytte

TI__Genius 17580 points

Hi,

We are not sure which layer is causing the mismatch. So you need to match layer by layer outputs to find the layer causing the issue.

Thanks,
Praveen

Processors

Processors forum

TDA2EXEVM: TIDL OD usecase failed with new model