faster rcnn on cloud TPU

[ernstgoyer]

I'm trying to run faster rcnn resnet inception v2 on a cloud TPU instance, I know that there is no config in the object detection API that is compatible with TPUs, but I am getting always the following error:

Compilation failure: Detected unsupported operations when trying to compile graph _functionalize_body_20[] on XLA_TPU_JIT:Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, RandomShuffle, Where, Where, RandomShuffle, Where, RandomShuffle, Where, Where, RandomShuffle, Where, RandomShuffle, Where, Where, RandomShuffle, Where, RandomShuffle, Where, Where, CropAndResize, CropAndResizeGradImage, CropAndResizeGradBoxes

so I was wondering if there is a way to change these operations to be compatible with TPU

[aman2930]

Could you please share the portion of the code where these operators are being used? There must be corresponding TPU compatible operations, but I need to look into the code around these operations first.

[rishabhindoria]

@ernstgoyer https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tpu_compatibility.md

[dattatrayshinde]

@ernstgoyer Any help here with TPU compatible config for faster RCNN?