onnx
Schedule for moving onnx optimizers to this new repo
@daquexian @fumihwh : Could you add details regarding the planned schedule for moving optimizers to this repo along with ETAs. Thanks!
@askhade Thanks for reminding me!
As discussed in infra SIG meeting (https://github.com/onnx/sigs/blob/master/infra/meetings/002-20200806.md), optimizer code in onnx/onnx repo will be moved to this repo and maintained actively by the community. After discussing with @fumihwh, here is the planned timeline:
| Date | Content |
|---|---|
| 08.31 | Copy optimizer code from onnx/onnx repo to onnx/optimizer repo, ensuring the same functionality. Build an "onnx-optimizer" (or another better name) python package. Add warning message in onnx/onnx repo to advise users to use the new package. |
| 10.10 | Fix critical bugs: errors when IR version >= 4 fuse_bn_into_conv breaks the graph or introduces wrong parameters |
| 1.8 Release | |
| After 1.8 Release | New feature development, including (still need some discussion): New user-friendly API New IR for optimizers. Replace tensor.h with a mature tensor library like xtensor Integrate with ONNX Function operators (e.g., unfold function ops and do more optimization) Pattern matching like tensorflow-onnx |
| Before 1.9 Release | remove optimizer codes in ONNX repo |
Update: About tensor.h, I have a new idea of separating graph writing and constant folding ( https://github.com/onnx/optimizer/issues/9). We will not need tensor.h or a tensor library if we choose this way.
@daquexian : Add warning message in onnx/onnx repo to advise users to use the new package. When will this be done?
| 08.31 | Copy optimizer code from onnx/onnx repo to onnx/optimizer repo, ensuring the same functionality.Build an "onnx-optimizer" (or another better name) python package. Add warning message in onnx/onnx repo to advise users to use the new package. |
|---|
I think the optimizer package creation should be done only after fixing all the pending issues specially the IR gap. I do not see a point in publishing a package when the underlying code is not really useful because of bugs and incomplete support.
I think the optimizer package creation should be done only after fixing all the pending issues specially the IR gap. I do not see a point in publishing a package when the underlying code is not really useful because of bugs and incomplete support.
@askhade I don't think so. This repo has at least the same functionality with the optimizer in onnx/onnx repo. We should encourage users to use this repo so that they can benefit from the coming bugfixes (like #4).
I have published a package named onnxoptimizer on pypi yesterday (I had to wait for the merging of #2 so delayed a bit), and will submit a PR to add warning message in onnx/onnx repo.