Serialize data classes based on their fields only #2697
Closed
+21
−0
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D61665461 |
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #2697 +/- ##
=======================================
Coverage 95.28% 95.28%
=======================================
Files 493 493
Lines 47861 47876 +15
=======================================
+ Hits 45605 45620 +15
Misses 2256 2256 ☔ View full report in Codecov by Sentry. |
Summary: Pull Request resolved: facebook#2697 # Context: **I think the unit test is the easiest way to understand this!** There are several cases in Ax where an instance has an attribute that 1) can't or shouldn't be serialized; this is usually subclasses of `torch.nn.Module`, including BoTorch models, but can be other large and complex objects, 2) isn't typically passed at initialization, but rather constructed afterwards, and 3) should be an attribute rather than a property because it's too expensive to construct more than once These classes have thus required custom serialization logic so that such attributes are not serialized. Classes that have this issue include many benchmarking classes that work with surrogates or neural nets, such as `SurrogateRunner`, `PyTorchCNNTorchvisionRunner`, and `PyTorchCNNTorchvisionBenchmarkProblem`, as well as MBM classes. A simpler solution is to use dataclasses, by - specifying features that satisfy (1)-(3) with `InitVar`, and, if they are needed immediately, constructing them in the `__post_init__` - only serializing fields; `InitVar`s are not fields This gives more flexibility in what we serialize without taking any away: If an attribute is constructed in the post-init and *should* be serialized, that is still supported by marking it as a `field` and not an `InitVar`. Attributes that are constructed in the init will be serialized, even if they are modified elsewhere. ## Downside It is not quite the normal usage to use an `InitVar` to define a persistent non-field attribute; instead `InitVar` is intended for something that is needed only for initializing fields. So the usage pattern outlined in the unit test causes Pyre errors, and someone who uses an `InitVar` that way might be surprised that it can't be recovered. However, it might not need to be, since the other fields would be recovered. Also, `InitVar` is a rarely used feature. # This PR: * Changes Ax's JSON serialization for dataclasses to exclude non-fields * Adds a unit test Reviewed By: danielcohenlive Differential Revision: D61665461
|
This pull request was exported from Phabricator. Differential Revision: D61665461 |
esantorella
force-pushed
the
export-D61665461
branch
from
fcb368e to
2345564
Compare
|
This pull request has been merged in 726bb84. |
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.
Summary:
Context:
I think the unit test is the easiest way to understand this!
There are several cases in Ax where an instance has an attribute that
torch.nn.Module, including BoTorch models, but can be other large and complex objects,These classes have thus required custom serialization logic so that such attributes are not serialized. Classes that have this issue include many benchmarking classes that work with surrogates or neural nets, such as
SurrogateRunner,PyTorchCNNTorchvisionRunner, andPyTorchCNNTorchvisionBenchmarkProblem, as well as MBM classes.A simpler solution is to use dataclasses, by
InitVar, and, if they are needed immediately, constructing them in the__post_init__InitVars are not fieldsThis gives more flexibility in what we serialize without taking any away: If an attribute is constructed in the post-init and should be serialized, that is still supported by marking it as a
fieldand not anInitVar. Attributes that are constructed in the init will be serialized, even if they are modified elsewhere.Downside
It is not quite the normal usage to use an
InitVarto define a persistent non-field attribute; insteadInitVaris intended for something that is needed only for initializing fields. So the usage pattern outlined in the unit test causes Pyre errors, and someone who uses anInitVarthat way might be surprised that it can't be recovered. However, it might not need to be, since the other fields would be recovered. Also,InitVaris a rarely used feature.This PR:
Differential Revision: D61665461