Shape Inference code for the DepthToSpace operator #903
Conversation
Signed-off-by: Ettore Tiotto <[email protected]>
Signed-off-by: Ettore Tiotto <[email protected]>
There was a problem hiding this comment.
@etiotto my understanding is that both depthofspace and spaceofdepth requires the input to be of rank 4. That is a perfect example for introducing a verify method.
One can be added by adding the op in utiles/gen-onnx-mlir in the python array OpsWithVerifier.
You can look at other verifiers in OnnxOps.cpp, should be a trivial change.
You have to run the onnx script that rebuilds the table gens, see https://github.com/onnx/onnx-mlir/blob/master/docs/ImportONNXDefs.md
|
@AlexandreEichenberger I see you have approved this PR. Can I open a separate PR to add the custom |
|
@etiotto sure, a different PR is ok too. |
|
Jenkins Linux s390x Build #2172 [push] Shape Inference code for... started at 10:14 |
|
Jenkins Linux amd64 Build #2173 [push] Shape Inference code for... started at 09:14 |
|
Jenkins Linux ppc64le Build #1282 [push] Shape Inference code for... started at 10:15 |
|
Jenkins Linux s390x Build #2172 [push] Shape Inference code for... passed after 50 min |
|
Jenkins Linux amd64 Build #2173 [push] Shape Inference code for... passed after 54 min |
|
Jenkins Linux ppc64le Build #1282 [push] Shape Inference code for... passed after 1 hr 2 min |
The DepthToSpace operator accepts an input tensor and has a blocksize attribute. The operator reshapes the input tensor according to the blocksize attribute. Specifically, given an input tensor of dimension [N,C,H,W] (where N is the batch axis, C is the channel or depth, H is the height and W is the width) and a blocksize, the output tensor will have shape [N, C / (blocksize * blocksize), H * blocksize, W * blocksize].
This PR implements the shape inference code for the operator.
Signed-off-by: Ettore Tiotto [email protected]