Cherrypick Sequential serialization bug fix for r2.13 #18258
Conversation
|
FYI, original change in a78f714. |
google-ml-butler
bot
added
kokoro:force-run
ready to pull
|
Will let the test run again for the release branch before merge. Thanks for the quick fix. |
|
Will fix tensorflow/tensorflow#61085 |
|
@nkovela1, seems that the presubmit test failed, but it was passing on head. Could it be any change that is between the current head and the release branch also need to be cherrypick into the release branch? To reproduce the test failure, you can checkout the rc branch and this pr and run the blaze test locally. |
|
Tested locally with Neel for this PR. The test error is related caused by a inconsistent version of TF used in the config. We suppose to test the RC build against the tf 2.13 RC version, but the test was using the latest tf-nightly. The delta between them was causing these 4 tests to fail. We have verified those 4 tests locally, and they are passing with tf.2.13 rc2. Merging this PR now. |
google-ml-butler
bot
removed
the
ready to pull
5 participants
This PR is a fix based on the serialization bug reported by @seanpmorgan.
It fixes the serialization format of Sequential models saved with HDF5 and is important for ensuring backwards compatibility with the serialization format switch in TF 2.13.