Fix dtype and assign* in AutocastVariable.
#1136
Merged
Conversation
This file contains hidden or 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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
github-actions
bot
added
the
technique:pruning
copybara-service
bot
force-pushed
the
test_647135376
branch
6 times, most recently
from
af20f6f to
863e74d
Compare
The `dtype` property would return to true dtype of the variable, instead of the dtype of the value that you get explicitly via `.value()` or implicitly by doing any operation. This would cause seemingly correct things like this to fail with a dtype mismatch: ``` y = variable * tf.cast(x, variable.dtype) ``` Forcing users to write workarounds like: ``` v = variable.value() y = variable * tf.cast(x, v.dtype) ``` Additionally, `assign`, `assign_add`, `assign_sub` expected the value to be of the true dtype, not the cast dtype. This would cause seemingly correct things like this to fail with a dtype mismatch: ``` variable.assign(variable * factor) ``` (This is a common use case for non-trainable variables.) Forcing users to write workarounds like: ``` variable.assign(tf.cast(variable * factor, variable.dtype)) ``` This changes fixes these issues to make autocasting fully transparent: - `dtype` returns the cast dtype if applicable - `assign*` accept the cast dtype for the value if applicable Note that this is consistent with how autocasting works in Keras 3. PiperOrigin-RevId: 650386711
copybara-service
bot
force-pushed
the
test_647135376
branch
from
863e74d to
ed3f017
Compare
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Outdated suggestions cannot be applied.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.
Fix
dtypeandassign*inAutocastVariable.The
dtypeproperty would return to true dtype of the variable, instead of the dtype of the value that you get explicitly via.value()or implicitly by doing any operation.This would cause seemingly correct things like this to fail with a dtype mismatch:
Forcing users to write workarounds like:
Additionally,
assign,assign_add,assign_subexpected the value to be of the true dtype, not the cast dtype.This would cause seemingly correct things like this to fail with a dtype mismatch:
(This is a common use case for non-trainable variables.)
Forcing users to write workarounds like:
This changes fixes these issues to make autocasting fully transparent:
dtypereturns the cast dtype if applicableassign*accept the cast dtype for the value if applicableNote that this is consistent with how autocasting works in Keras 3.