Do not see any inference time improvement on object detection model. about aws-neuron-sdk

[awsrjh]

Hello and thank you for providing feedback.

Object detection models are currently a limitation of the Neuron Compiler (please see https://github.com/aws/aws-neuron-sdk/blob/master/release-notes/neuron-cc.md#1046800 ) and thus the behavior of the Tensorflow-Neuron integration is to revert the whole graph to a CPU-only execution. This is why you see that "Number of operations placed on Neuron runtime: 0" as you pointed out.

We are working on this and expect to have support for the additional functions shortly (please see the roadmap here: https://github.com/aws/aws-neuron-sdk/projects/2 . The Issue #40 tracks this. Please feel free to to add additional information there.

[aws-taylor]

Hello whatdhack,

Is there any further assistance we can offer at this time?

Regards,
Taylor

[whatdhack]

Any timeline on when you might release object detection ( specifically Faster RCNN, Yolo V3 and SSD) support ?

[aws-taylor]

Hello whatdhack,

You can find our roadmap here. In particular, notice that Mask R-CNN is high on the priority list. I'd encourage you to to "watch" this issue for the most up-to-date information.

Is there any other help we can offer?

Regards,
Taylor

[micwade-aws]

Since we haven’t heard back in a while we are closing this issue, please reopen it if more support from us is needed.

[whatdhack]

Hi, any update on support for object detection models like maskrcnn ?

[AWSGH]

We just released a new Neuron version with support for SSD300, and plan to continue expanding our support in the future. Please have a look at: https://github.com/aws/aws-neuron-sdk/blob/master/src/examples/tensorflow/ssd300_demo/README.md