Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Use higher opset_version when exporting ONNX models #3964

Closed
xiaowu0162 opened this issue Jul 20, 2021 · 0 comments · Fixed by #3968
Closed

Use higher opset_version when exporting ONNX models #3964

xiaowu0162 opened this issue Jul 20, 2021 · 0 comments · Fixed by #3968
Assignees

Comments

@xiaowu0162
Copy link
Contributor

xiaowu0162 commented Jul 20, 2021

What would you like to be added:

  1. A dummy_input argument for the export_model in Compressor.
  2. Allow using opset_version=10 or higher when calling torch.onnx.export inside the export_model in Compressor.

Why is this needed:

  1. Some models need a tuple instead of a tensor as the model input. For example, transformer models require both input and input mask as the model input. In these cases, it seems that the original input_shape argument is not enough to specify the dummy input.
  2. Some models that use dynamic indexing (e.g., transformers) does not work with the default opset_version (9) of torch.onnx.export. The reported error looks like RuntimeError: Unsupported: ONNX export of Slice with dynamic inputs. DynamicSlice is a deprecated experimental op. The error disappears when opset_version=10 is used.

Without this feature, how does current nni work

It does not work for exporting transformers to ONNX.

Components that may involve changes:

  1. export_model function in Compressor
  2. Model Speedup may be affected

Brief description of your proposal if any:

Allow a dummy_input argument and possibly an opset_version argument.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging a pull request may close this issue.

1 participant