Loss functions applied in alphabetical order instead of by dictionary keys (since Keras 3.5.0)

[malcolmlett]

Current behavior
When training models with multiple outputs, and using dictionary as loss parameter to model.compile(), the wrong loss functions are applied against the different model outputs. The alphabetical order of the dictionary keys determines the order of loss function against model output.

Expected behavior
Loss functions should be applied against named model outputs.

Workaround
Bug is with Keras 3.5.0 (Google Colab's current default), and can be avoided by downgrading to Keras 3.4.1.

Additional context
Issue has been raised with Keras team: keras-team/keras#20596
Detailed instructions for replication are included there.

This present issue is for awareness of Colab team.

[cperry-goog]

This is a great bug report, thank you! I'll pass on to the Keras folks as well.

[malcolmlett]

Update: I'd mistakenly typed Keras 2.5.0 and 2.4.1, when that should have been 3.5.0 / 3.4.1 -- have updated the description.

Also, Keras folks say that Keras 3.7.0 works ok. What does that mean for Google Colab and its current 3.5.0 default?

[colaboratory-team]

Tracking internally as b/383879003, we will try to get Keras updated