TDA4VM: GlobalAveragePooling not working as expected

Sharon Ngu

Part Number: TDA4VM

Hi,

I am trying to run inference on target HW using TI's TFLite delegate and GlobalAveragePooling2D() does not behave as expected. With input of 1x4x4x3, I got the expected output size of 3 but only the first output is correct and the other two outputs seem to be junk. Are there any limitations on GlobalAveragePooling2D()?

Thanks.

Kind regards,

Sharon

over 3 years ago

0 Anand Pathak over 3 years ago

TI__Genius 9065 points

Hi Sharon,

This is supposed to work. Is the output correct in PC emulation mode and is the issue specific to hardware?

Regards,

Anand

0 Sharon Ngu over 3 years ago in reply to Anand Pathak

Prodigy 30 points

Yes, the output is correct in PC emulation mode but not on hardware.

0 Anand Pathak over 3 years ago in reply to Sharon Ngu

TI__Genius 9065 points

Ok Sharon, thanks. Can you please share the above representative network, we need to investigate this further.

Regards,

Anand

0 Sharon Ngu over 3 years ago in reply to Anand Pathak

Prodigy 30 points

Would the representative network affect the output in this case since it is a mean computation? I used a random 4x4 RGB image for inData.

0 Anand Pathak over 3 years ago in reply to Sharon Ngu

TI__Genius 9065 points

Hi Sharon,

I created a network with global average pooling layer and tried to run it on my end and I am not able to reproduce this issue. So I am requesting for your network just to make sure the base network we are trying to debug is same as your network where issue is seen. I don't think this will have anything to do with the inData though.

Regards,

Anand