TDA2: Output dimension of argmax layer is same as input layer

Navin Prashath

Part Number: TDA2

Hi,

I am using LeNet model for classification. The final layer is argmax with input of 1x1x1x40. When I import the model, the output of argmax is also 1x1x1x40. I expected it to be 1x1x1x1. And all the values of the argmax layer are zero.

Regards,

Navinprashath.R.R

over 7 years ago

0 Praveen Eppa1 over 7 years ago

TI__Genius 17580 points

Hi Navin,

Please check the limitations of ArgMax layer in the section 3.9 (TIDL Limitation) of the user guide.

Thanks,
Praveen

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

Intellectual 760 points

Hi Praveen,

Thanks. Looks like argmax is supported only across the channels.

Regards,
Navinprashath.R.R

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

Intellectual 760 points

Also for the same network, if I change the numParamBits from 8 to 12 import fails. Any idea why this is happening?

Regards,
Navinprashath.R.R

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

TI__Genius 17580 points

Hi Navin,

I am not sure, ideally it should work. What kind of failure are you seeing?

Thanks,
Praveen

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

TI__Genius 17580 points

Also, can you try with numParamBits = 10 ?

Thanks,
Praveen

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

Intellectual 760 points

With numParamBits= 10 also, the import module throws error. I get a windows pop up telling eve_test_dl_algo.out.exe has stopped working.
The default CIFAR10 example works fine with both 8 and 12 bits.

Regards,
Navinprashath.R.R

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

Intellectual 760 points

Hi Praveen,

Compared to the usual LeNet network, we have two Eltwise Sum layers.

Regards,
Navinprashath.R.R

0 kumar.desappan over 7 years ago in reply to Navin Prashath

TI__Mastermind 22145 points

The issue your observing may be because of the limitation mentioned in the below thread. Update your network to handle this

e2e.ti.com/.../694032

Processors

Processors forum

TDA2: Output dimension of argmax layer is same as input layer