Conversion from tf.Tensor to numpy is slow in Tensorflow2.2.0 + nv20.8

Autonomous Machines Jetson & Embedded Systems Jetson TX2
,
1

Conversion from tf.Tensor to numpy is slow in Tensorflow2.2.0 + nv20.8
After inference, numpy conversion takes a long time (about 0.5sec)
Please tell me if the consistency is not good in the following environment
・ Jetpack 4.4
・ Tensorflow 2.2.0 + nv20.8

Jetpack 4.3 works without delay, but please tell me what’s the difference

3

Hi,

Could you monitor the device with tegrastats to see if any swap memory is used first?

$ sudo tegrastats

Thanks.

4

I was able to confirm the consumption of swap memory.
What are the main causes of such delays?

5

Hi,

Swap memory is a virtual memory with storage rather than real physical memory.
So the bandwidth is expected to be slower.

In general, we found that TensorFlow 2.x consumes more memory than v1.x.
Do you use the same script or TensorFlow version for JetPack4.3 and JetPack4.4?

Thanks.