[ShapeInferenceError] Inferred shape and existing shape differ in dimension 0 #1467
Conversation
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #1467 +/- ##
==========================================
+ Coverage 76.84% 76.88% +0.04%
==========================================
Files 209 209
Lines 22461 22461
Branches 3811 3811
==========================================
+ Hits 17260 17270 +10
+ Misses 4482 4473 -9
+ Partials 719 718 -1 ☔ View full report in Codecov by Sentry. |
|
Right. So far, I did not fix onnx-script. I just created the unit test which fails. |
|
We always have torch so I think we are ok |
| import torch | ||
| import torch.nn as nn | ||
| import torch.onnx | ||
| from onnx.inliner import inline_local_functions |
There was a problem hiding this comment.
| from onnx.inliner import inline_local_functions | |
| import onnx.inliner |
| def inline_function(*args, **kwargs): # pylint: disable=unused-argument | ||
| first_model_proto = args[0] | ||
|
|
||
| next_model = inline_local_functions(first_model_proto) | ||
|
|
||
| del first_model_proto.graph.node[:] | ||
| del first_model_proto.functions[:] | ||
| del first_model_proto.graph.initializer[:] | ||
| del first_model_proto.opset_import[:] | ||
| first_model_proto.graph.node.extend(next_model.graph.node) | ||
| first_model_proto.functions.extend(next_model.functions) | ||
| first_model_proto.graph.initializer.extend(next_model.graph.initializer) | ||
| first_model_proto.opset_import.extend(next_model.opset_import) | ||
|
|
||
| return first_model_proto |
There was a problem hiding this comment.
Avoid nested declarations of functions if it doesn't close over local variables?
| # torch.library.impl_abstract was renamed to torch.library.register_fake. | ||
| warnings.simplefilter("ignore", (UserWarning, DeprecationWarning)) | ||
|
|
||
| def assertONNX( |
There was a problem hiding this comment.
I would consider a more descriptive name. Does it assert on that the output of the onnx model will match the torch output, for example? Also snake cases for all function names
| def make_aot_ort( | ||
| dynamic: bool = False, | ||
| ort_optimization_level: Optional[str] = None, | ||
| ) -> tuple: |
There was a problem hiding this comment.
| ) -> tuple: | |
| ) -> tuple: |
Is the return type accurate?
|
|
| ) | ||
|
|
||
| baseline_result = model(*args) | ||
| result = compiled_model(*args) |
There was a problem hiding this comment.
@xadupre I tried to put everything on CUDA and run the code, but even though compiled_model is on CUDA, this result will not be on CUDA. Do you know why?
Creates a unit test running DORT and showing the following error:
Can be used to replicate issue #1443