TDA2: Any limit for eltwise in tidl?

jerrytte

Prodigy 220 points

Part Number: TDA2

Hi,

I use the the following block in cifar-10 network, but it always go wrong.

And I try to change it to the following block, and it works well, but accuracy has a large drop

There is something wrong with the first block or something limits for eltwise block?

Thanks

over 7 years ago

0 kumar.desappan over 7 years ago

TI__Mastermind 22145 points

Both the cases are suported by TIDL. are your using caffe or tensorflow trained model?
For the first case are you oberserving import issue or wrong inference output?

Can you share the model, we will have look.

Rgerda,
Kumar.D

0 jerrytte over 7 years ago in reply to kumar.desappan

Prodigy 220 points

Hi,

Please help to check the model, thanks

eltwise_without_relu_previous.tar.gz eltwise_with_relu_previous.tar.gz

0 jerrytte over 7 years ago in reply to kumar.desappan

Prodigy 220 points

Hi ,

Do you have any updates about this issue? Thanks.

Jerry

0 Praveen Eppa1 over 7 years ago in reply to jerrytte

TI__Genius 17580 points

Hi Jerry,

I am working on it, I will post my response here as soon as I have some update.

Thanks for your patience.

Regards,
Praveen

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

TI__Genius 17580 points

Hi Jerry,

As indicated in the TIDL user guide section 3.9 TIDL Limitation, Eltwise layer only supported on inputs of same types,
but in the first case (eltwise_without_relu_previous), one input "layer_64_1_bn3" is signed and other input "conv1_pool" unsigned, so it is not supported and hence not working.
Please make sure both the input should be of same type (either both are unsigned or signed inputs).

Thanks,
Praveen

0 jerrytte over 7 years ago in reply to Praveen Eppa1

Prodigy 220 points

Hi Praveen,

Got it.
I'll modify the net structure to satisfy this limitation, and make a validation later.
thanks for your help.

Thanks,
Jerry

Processors

Processors forum

TDA2: Any limit for eltwise in tidl?