CCS: TDA4x: Tensorflow based model Signed layer output optimization issue on board

Hi Vyom,

Are you refring to below

or 

Below issue

Dear Sir,

The issue described in the second thread i.e., https://e2e.ti.com/support/processors/f/791/p/855997/3167091#3167091

Thanks and Regards,

Vyom Mishra

Hi Vyom,

If yes,

Can you provide mode details of layer 78. Things like the resolution, kernel dimension other properties of the convolution.

I could not find theses details on the other thread.

Dear Sir,

Please find the below thread for the complete progress on 78th layer information and further discussion:

In the last response from TI side

Anshu replied us 

" The issue is coming because the output of this layer is coming as signed output, which is currently not handled in our optimized implementation. To move forward can you have a Relu layer after this convolution? In the mean time I will figure out if we can support this or not ( I think there should not be an issue in supporting this but I want to check with our team here)."

My concern is regarding the above statement, Kindly confirm if signed output coming from the layer is handled in your optimization implementation of latest release?

Thanks and Regards,

Vyom Mishra

Hi Vyom,

   What is the version of MMALIB you are using? This issue is fixed in MMALIB 1.1. Can you confirm?

Regards,

Anshu

Hi Vyom,
   Can we close this thread?

Regards,

Anshu

Dear Sir,

We tried to run the model on target(StandAlone-CCS) on below versions:

Processing config file #0 : testvecs/config/infer/tidl_infer_DASNet.txt

Syncd

MMALIB_CNN_convolve_col_smallNo_ixX_ixX_oxX_init 1 :FAILED

Algorithm Init failed with error number: -1

Error at line: 496 : in file src/tidl_tb.c, of function : tidlMultiInstanceTest

Invalid Error Type!

Hi Vyom,

   Is there a small dummy model which you can share for us to reproduce this issue?

Regards,

ANshu

Hi Vyom,

   I tried this model on the release mentioned here and I am not seeing any such error message. Can you specifically tell which layer is this? Btw the model which was shared was not importing as it is, i have to skip one error message in concat layer ( this shouldn't change any behavior).

Regards,

ANshu

Dear Sir,

Thanks for trying out on your side.

Yeah, we need to skip the error message and it will not affect the behaviour.

As of now, we are failing with this same dummy model on Target side.

I request if you can mail us the imported model files with inference and import config files too.

We can also check on our side.

Thanks and Regards,

Vyom Mishra

Vyom,

     The failure you are observing is because we don't support 2x2 depthwise separable convolution. Is it possible to change this?

regards,

Anshu

Dear Sir,

It is a little confusing for us.

If 2x2 depthwise Conv. is not supported then how you ran the dummy model on the target.

Till then i ll check with the team for this change.

Kindly do the needful.

Thanks and Regards,

Vyom Mishra

Vyom,
  I am not able to run on target because I get the following error :

Hi Anshu,

Can you please clarify if the model ran on the Target or not. If it did then why are you claiming the error is due to and unsupported layer which ran on the target.

Best Regards,

Sankalp

Sankalp,

  This model will not run on target because of the error mentioned above.

Regards,

ANshu

The model worked till PC emulation. Why is it that after PC emulation the model is failing on target? if the support is not present on target then why has it been allowed on PC emulation. Do you mean that the operations supported in PC emulation and not on Target?

Sankalp,

   PC emulation mode runs the reference code for each layer, here when I say reference code it means that there is no DMA, MMA involved, its a C/C++ implementation of each layer. Now reference code can be available for various combinations of a particular layer but  all combinations may not be having optimal implementation available. You can refer the following section for supported layers :

http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_01_00_10/ti_dl/docs/user_guide_html/md_tidl_layers_info.html

  We have added more warnings for such cases in import tool and in our upcoming release you should see warnings for this case during import itself.

Regards,

Anshu 

Dear Anshu,

Thank you for adding warnings. 

The error was coming due to signed values earlier as per you and not due to the depthwise separable conv layer not being supported.

We had waited for you to resolve this at your end. In case it wasn't the signed output issue we could have moved ahead faster on this model.

Regards,

Sankalp

Sankalp,
  The last time similar issue was reported that time the layer had 7x7 depthwise separable convolution but that time the issue was because of signed values. That issue is now fixed. But the model which is shared this time has a 2x2 depthwise separable convolution and not 7x7 depthwise separable convolution, which is not what supported so it has nothing to do with the previous reported issue. Hope this clarifies.

Regards,

Anshu