Efficientdet models do not work on hexagon delegate about tensorflow HOT 4 OPEN

Hi @Hozzu ,

I don't have a device with hexagon delegate so i tried out your model on the neuron delegate by mediatek and it failed there also wit the error:
java.lang.IllegalArgumentException: Cannot copy from a TesorflowLite tensor (statefulPartitonedCall_2:1) with 400 bytes to a Java Buffer with 100 bytes.

Can you please provide your crash logs from the hexagon delegate?

from tensorflow.

Hi @Hozzu ,

I don't have a device with hexagon delegate so i tried out your model on the neuron delegate by mediatek and it failed there also wit the error: java.lang.IllegalArgumentException: Cannot copy from a TesorflowLite tensor (statefulPartitonedCall_2:1) with 400 bytes to a Java Buffer with 100 bytes.

Can you please provide your crash logs from the hexagon delegate?

Hello.
It does not crash.
It runs but the output value is invalid.

When I use cpu or gpu delegate, the output values are valid.
But when I use hexagon delegate, the output value is different from the cpu and gpu ones and invalid.

The other models like yolov8 and mobilenet v2 work well on hexagon delegate.
However, efficientdet which made from google does not work on the hexagon delegate.

from tensorflow.

Hi @pkgoogle ,

Can you please take a look ?

from tensorflow.

I also do not have a device that can properly reproduce this, @arfaian can you please take a look? Thanks.

from tensorflow.