-
-
Notifications
You must be signed in to change notification settings - Fork 943
2020
Edgar Riba edited this page Oct 18, 2020
·
7 revisions
- Announcements
- Roadmap
- Final discussion
Announcements
- Release v0.4.1 by mid October.
- Pytorch v1.7 release by mid November
- https://github.com/pytorch/pytorch/issues/45592
- Make release ASAP for v0.4.1 and then v0.5 after pytorch v1.7 is out
- descriptors matching API and descriptors benchmark.
- 3d augmentations API and more.
- Next meeting after 0.5 release to discuss 0.7 roadmap.
Roadmap discussion
- Improve docs and generate automatic image examples in docs (mid/long term)
- create ticket: edgar
- Define priorities in tickets
- Create new issue for v0.5 roadmap
- https://github.com/kornia/kornia/issues/700
- start using project tabs for 0.4.1, 0.5.0, 0.6.0 and General Roadmap.
- https://github.com/kornia/kornia/projects
- Take older items and discuss
- Include DSAC [v0.5/6]
- Start with JIT integration in unit tests [v0.6]
- Functions overloading [v0.6/x]
- https://github.com/kornia/kornia/pull/684#discussion_r492263883
- Examples/tutorials reorganization [general Roadmap]
- Check if they crash and test
- Move everything to kornia-examples [general Roadmap]
- Follow pytorch-tutorials pattern
- https://github.com/pytorch/tutorials
- More augmentations [general Roadmap]
- From albumentations
- From tormentor: https://github.com/anguelos/tormentor
- Diamond square
- Meta augmentations - https://github.com/moskomule/dda
- Create kornia.geometry.transform.boxes [v0.6]
- Group several functionalities for bounding boxes manipulation
Final discussion
- Find some tool to find out which part of the tests have to be run when there’s a modification.
- Morphology module: collaboration [v0.6]
- Find more collaborations and start integration to other frameworks: lightning, catalyst, monai, ROS.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
- Association
- Website
- Open Technical Meetings
- Marketing
- Improve tutorials/examples
Association
- Create a legal entity so that we can operate and receive donations from third parties.
- Edgar will check on that in Barcelona: need at least 3 people on the board, can be from any country.
- Edgar and Stefano need to check on Gary to discuss about the relation between kornia and opencv.
Website
- remodel the kornia.org website
- define the web requirements we want before contacting the web developers.
- hire a web developer: budget 200/300E, we want at least to maintain the website, include blog, etc.
Open Technical Meetings
- We are organizing open meeting to get and receive feedback from the community.
- Reach more users to feedback, create forms to for feedback.
- Contact with companies using to self-promoting and collaborate.
- Better define roadmaps and look for collaborators, etc.
Marketing
- Need to coordinate social media presence.
- Get more users, get donations.
- Frequency of social media (plan and set schedule).
Improve tutorials/examples
- Check tutorials and see how we could improve them.
- Create kornia cheat-sheet.
- Contact guy that made tutorials in turkish to translate in English.
- Morphology module/tutorials.
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
- Next release
- Discuss data augmentation API
- Infrastructure, Testing
- Documentation
- Roadmap functionalities
Next release v0.2.2
- v0.2.2 after decopupling in augmentation to still support pytorch 1.4
- v0.3.0 to align with pytorch v1.5 and torchvision v0.6
Discuss data augmentation API
- for v0.3.0 we'll keep stick to the functional structure we have.
We want to let user to override
gen_params
,apply
and possiblygen_transformation
and easily create new ops. - For next iterations, Jian proposes to include
kornia.Sequential
to chain operations. - Implement
kornia.eye_like(n, tensor)
- Target compatibility with albumentations and rising for next release.
- torchvision plans: https://github.com/pytorch/vision/issues/1375
Infrastructure, Testing
- Automate GPU using AWS credits and CircleCI -> Edgar
- Edgar is meeting google collaborator to get TPUs for testing.
- Edgar has nvidia jetson to test half precision
- Will share resources with lightning
- We need to test JIT: start by augmentation module.
- Update testing framework using pytest tools
- Pytest-cov
- Pytest-mypy
- Define minimal testing api (check functions)
check_is_tensor(tensor)
check_is_shape(tensor, tuple) # (2,3,4,4)
-
check_eq_shape(tensor1, tensor2, …)
check_is_dtype(tensor, (torch.float32, torch.float64))
check_is_device(tensor, (‘cpu’, ‘cuda’)) -> tbd
Documentation
- Update to pytorch style -> Edgar (almost done)
- Verify all docs look good. Any tools for that?
- Doctest ? make it incremental.
Next Roadmap functionalities
- Old (0.2.1): https://github.com/kornia/kornia/issues/434
- Luis will refactor warper module
- AffineGridWarper -> AffineWarper. HomographyWarper, PerspectiveWarper (3d)
- Volumetric (rising guys)
- Features matching: kornia.feature.matching (edgar/dmytro)
- Nn, snn, sinkhorn
- homography/fundamental -> edgar
- TBD: morphology
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
- Discuss data augmentation module
- Allow data augmentation with branching -> for text.
- Other frameworks for data augmentation:
- Nvidia DALI: check timigs
- https://github.com/PhoenixDL/rising
- Priorities:
- Slower in CPU compared torchvision
- compute final transformation and apply
- create benchmarks
- empirical test, not to rely on a specific framework
- compare to torchvision using their parameters
- different types of transforms
- spatial transform
- channel, pixel not move at all
- gaussian noise, independent pixels
<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
The Kornia.org team