Update ilastik with the new bioimageio v0.5 spec #2919
Conversation
thodkatz
changed the title
k-dominik
changed the title
disconnect UI updates when closing the project.
also updated grpcio explicitly
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
737127f to
c444612
Compare
fixed shape calculation: * same factor for all axes * fixed for shapes that are bigger than minimum shape
make sure the interface is compatible with our code usage. (`tifffile.TiffFile` constructor)
Adapt classifier tests after tiktorch refactor (very shallow ModelSession determining shape now happens via bioimageio.
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
768ff8b to
8c5a5aa
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2919 +/- ##
==========================================
+ Coverage 56.51% 56.56% +0.04%
==========================================
Files 535 536 +1
Lines 62322 62459 +137
Branches 7714 7727 +13
==========================================
+ Hits 35220 35328 +108
- Misses 25338 25356 +18
- Partials 1764 1775 +11 ☔ View full report in Codecov by Sentry. |
Upgrading bioimageio libs to v0.5 spec broke running with previous saved projects. * Added a test to document that breakage * Added new project file with a v0.5 model
the next version (0.1.1) makes the object classification test time out: TestObjectClassificationGui.test_03_select_object_features
permission errors on appveyor when removing the env. prefer not to deal with manual removal when we can cache only the folder that we care about. Should also speed up our builds a bit
we relied on bioimageio functionality to expose the collection.json previously. This functionality has been removed.
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
ed2c7d5 to
c275088
Compare
same is done on deserialization and also on the tiktorch side with `format_version="latest"`.
|
Note, there's something weird going on with the GH actions on windows - the post run takes ages (not sure about the root cause, didn't have time to investigate, but I guessed this: conda-incubator/setup-miniconda#380). |
ilastik/utility/bioimageio_utils.py
Outdated
| # if all sizes are fixed, factors might be empty | ||
| min_factor = min(factors) if factors else None | ||
| for k, f in size_funcs.items(): | ||
| sized_axes[k] = f(min_factor) |
There was a problem hiding this comment.
Very cool idea with the partials!
There was a problem hiding this comment.
I had concerns about long term maintainability of this code (alternative would be to iterate twice), but if you like it I'll keep it.
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
f496734 to
f0dca68
Compare
There was a problem hiding this comment.
The *Utils and *Validator classes make me think that maybe we shouldn't be that tied to the spec in runtime after all. We could, say, just grab the spec and produce something like a IlastikPredictionPipeline instance that has all the logic and semantics it needs, without needing to keep going back to the spec. As a bonus, this would also work for spec v4 and a potential spec v6 in the future, as they might express the same concepts in different ways; all we'd need is another factory method to read the v4/v6 spec and then we'd be back to only the runtime class.
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
1d9c62e to
714de73
Compare
encountered incompatibilities with conda and boa on win conda-build command seems not available -> conda build
cleanup there takes hours, or even times out ref: conda-incubator/setup-miniconda#380
Improved typing in `bioimageio_utils.py`, which also lead to some code improvements (like missed cases...). Co-authored-by: Tomaz Vieira <[email protected]>
Co-authored-by: Theodoros Katzalis <[email protected]>
k-dominik
force-pushed
the
update-tiktorch-with-v5
branch
from
2e57b40 to
32b5c52
Compare
This is a work in progress to update ilastik with the spec v5.
It uses a non released version of tiktorch. You can find it on this PR ilastik/tiktorch#222.
To test this, you need to manually install the PR version of tiktorch.
The only use case that I have currently tested and it works with our first v5 model is the following:
Known issues:
get_resource_package_content()given weights priority, downloads the weights bioimage-io/spec-bioimage-io#638TODO: