-
Notifications
You must be signed in to change notification settings - Fork 446
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
Add mmdeploy export #2806
Add mmdeploy export #2806
Conversation
* added proper configs * minor * fix config * fix head, updated configs * fix pre-commit. Part 1 * fix pre-commit. Part 2 * fix pre-commit. Part3 * added default numbers for HW, mean and scale
…lkit#2797) * refactored export * change export to onnx * fix pre-commit. part 1
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It seems that there are many TODO or workaround in this PR. Could we make the separate backlog for those issues to avoid the missing ?
For you guys information, currently there is an issue using the mmdeploy to the v1. that's related with the |
Summary
This PR enables mmdeploy export to OTX2.0. Currently, detection and instance segmentation need mmdeploy for export.
This PR includes
mmdeploy
OVInstanceSegmentationModel
TBD
Difference between mmdeploy export in OTX 1.X and OTX 2.0
When exporting a model w/ mmdeploy in OTX, model is converted to ONNX and ONNX is converted to OV.
In OTX 1.X, mmdeploy is used for both conversion but OTX 2.0 uses OpenVINO to convert ONNX model to OV.
It's for not falling behind latest update in OV because current latest mmdeploy is still using OpenVINO 22.03 but current latest version is 23.02.
I also removed unused code as mmdeploy ONNX -> OV is dropped.
Update integration test.
Export model score can be unstable if a model isn't fully trained.
To prevent it, some detection models get more epoch.
maskrcnn_swint
isn't trained at all, so skip it in the test.python format mmdeploy config
Because config files given by mmdeploy has python format as default,
I leave it as it is not to give user additional burden to get mmdeploy configs for new models in OTX.
P.S. I bring some code unused now but for future (e.g. export tilling, configurable parameter related setting)
How to test
Checklist
License
Feel free to contact the maintainers if that's a concern.