Force ICNN to adopt default initialization of its own layers #551
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #551 +/- ##
==========================================
+ Coverage 91.38% 91.39% +0.01%
==========================================
Files 69 69
Lines 7242 7242
Branches 1019 1018 -1
==========================================
+ Hits 6618 6619 +1
Misses 472 472
+ Partials 152 151 -1
|
michalk8
approved these changes
Jun 25, 2024
michalk8
pushed a commit
that referenced
this pull request
Jun 27, 2024
michalk8
added a commit
that referenced
this pull request
Oct 16, 2024
* Start batched vmap
* Initial `batched_vmap` impl
* Nicer formatting
* Fix getting shape
* Remove private API usage
* Fix new args
* Add a TODO
* Canonicalize axes
* Add `batched_vmap` to docs
* Removed batched transport functions
* Remove `_norm_{x,y}` from `CostFn`
* Implement `apply_lse_kernel`
* Implememt `apply_kernel`
* Implement `apply_cost`
* Remove old functions
* Make function private
* Refactor `apply_cost` to have consistent shapes
* Use `_apply_cost_to_vec` in `PointCloud`
* Remoeve TODO
* Formatting
* Simplify `_apply_sqeucl_cost`
* Fix `RecusionError`
* Remove docstring of a private method
* Fix `apply_lse_kernel`
* Squeeze only 1 axis of the cost
* Add TODO
* Rename function, make a property
* Remove unused helper function
* Compute mean summary online
* Compute mean online
* Compute max cost matrix
* Update error message
* Remove TODO
* Flatten out axes
* Fix missing cross terms in the costs
* Fix geom tests
* Fix dtype
* Start implementing transport functions
* Implement online transport functions
* Fix solver tests
* Fix Bures test
* Don't use `pairwise` in tests
* Update notebook that uses `norm`
* Fix bug in `UnbalancedBures`
* Rename `pairwise -> __call__`
* Remove old shape code
* Always instantiate the cost for online
* Remove old TODO
* Extract `_apply_cost_to_vec_fast`
* Update max cost in LRCGeom
* Fix test, use more `multi_dot`
* Remove `batch_size` from `LRCGeometry`
* Add better warning error
* Reorder properties
* Add docs to `batched_vmap`
* Start adding tests
* Reorder functions in test
* Fix axes, add a test
* Update test fn
* Move out assert
* Dont canon out_axes
* Check max traces
* Test memory of batched vmap
* Install `typing_extensions`
* Remove `.` from description
* Add more `out_axes` tests
* Add `in_axes` test
* Fix negative axes
* Increase memory limit in the test
* Add in_axes pytree test
* Remove old warnings filters
* Update fixtures
* Update SqEucl cost.
* Update docstrings
* Remove unused imports from the docs
* Revert test pre-commits
* Fix ICNN init notebook
Was broken by #551
* Improve error message
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The ICNN used to rely on initialisation with
normal matrices. Now, it fallbacks to the behavior of the layers, i.elecun_normal, which scales the standard deviation of the weights with1/sqrt(fan_in). Thisstdis much smaller for wide networks.