onnx-mlir
onnx-mlir copied to clipboard
onnx
Uplift BatchNormalization, CastLike, Bernoulli to opset 15
This PR helps move some ops to the latest opset. These are all ops whose latest redefinition are in opset 15, and haven't been updated or added in onnx-mlir
Ops introduced in Opset 15
- CastLike
- Bernoulli
Ops changes
- BatchNormalization
Note that BatchNormalization has a more relaxed definition in Opset15. It allows more flexibility in the types of its inputs.
Updated the branch, which will also automatically trigger the CIs since I did it. So you will have to do a pull, let me know if you prefer me not doing this and simply laughing the CIs for you.
I have not looked at the code in detail yet, our goal is to maintain backward compatibility with older models. So as you upgrade, make sure that the old version also works. Sometimes it is easy, as the old one works by default, sometime there is some minimal work (e.g. when an attribute becomes a input), and sometime we had to rename the old version of the op with an explicit version number (make this one up, ONNXSqueezeV13Op), migrate all our current code to that versioned op, and make the new op the default (without version number) with the new support it requires. I believe it is well documented in our docs folder.
@philass since you don't have rights to push down code (I think), let me know when you are ready to do so.
No change needed for lowering. The only change of BatchNormalization is that ReduceMean has to be float when the input is float16. That is used only in the train mode. We only lowered BathNormalization for inference mode.
@AlexandreEichenberger I'm ready to have this merged.
Thanks for the review. And thanks @chentong319 for the input
@chentong319 ready to approve? If yes, and @philass is happy with the state of the PR, I can merge it. Tx to all
@chentong319 are you ready to approve this PR, or suggest changes? Tx @philass there are some conflicts that need manual attention.