A Problem with Bilinear Resize op in Openvino R3

[Lee__Sangyun]

I'm working with Openvino R3 released recently.

It seems that there is a problem with Bilinear Resize op.

Bilinear resize output of openvino is differ from that of tensorflow.

The reason is because the op of the openvino is implemented as "align_corners=True", while default setting of TF is "align_corners=False"

I've confirmed that output of openvino is same with that of TF, if align_corners is set to True in TF.

Is there any options for align_corners=False in openvino??

or any plan to update for that?

Thanks, regards

[Severine_H_Intel]

Hi Sangyun, 

thank you for reporting this issue, we are aware of it and we will work to solve it. For now, only the case "align_corners=True" is possible.

Best, 

Severine

[Hendrich__Florian]

Hi,

i have another issue with the Resize Bilinear operation in the current R3 release. After successfully converting a trained TensorFlow model into the Intel IR format, the inference engine raises an "Following layers are not supported by the plugin for specified device CPU:" for the ResizeBilinear operation. I am trying to convert a image segmentation network which uses bilinear resize in tensorflow for usage on intel CPU. The test was done with the python API.

Any suggestions why this error is raised, is this operation not implemented yet for CPU?

Thanks!

[Lunscher__Nolan]

I have this same issue in the R4 release.

Hendrich, Florian wrote:

Hi,

i have another issue with the Resize Bilinear operation in the current R3 release. After successfully converting a trained TensorFlow model into the Intel IR format, the inference engine raises an "Following layers are not supported by the plugin for specified device CPU:" for the ResizeBilinear operation. I am trying to convert a image segmentation network which uses bilinear resize in tensorflow for usage on intel CPU. The test was done with the python API.

Any suggestions why this error is raised, is this operation not implemented yet for CPU?

Thanks!

[Rai__Avinash]

Hi,

I am facing the same issue with our Super Resolution model which uses ResizeBilinear Interp layer from custom kernels.

I tried in tensorflow with align_corners=true as well but no luck. The accuracy drop is high. Can anyone help me with this? I am using OpenVINO R3.1