Potential bug in legacy h5 weights loading. about keras

[SuryanarayanaY]

Hi @torzdf ,

Thanks for reporting. It would be appreciated if you can submit a code snippet for reproducing the error.

[torzdf]

Hi @torzdf ,

Thanks for reporting. It would be appreciated if you can submit a code snippet for reproducing the error.

Not straightforward, as I discovered this by injecting print statements at the above linked locations.

Either way, I have demonstrated where the code diverges, The difference is clear to see from the different variables that are used to collect the symbolic weights.

This was discovered by loading a model from Keras 2 into Keras 3 which adds further challenges providing reproducible code. I do not know if this issue exists for legacy models generated in Keras 3 as I resolved the issue my end by not using the by_name variant.

However, I raised this issue as I thought it may be a bug. so thought you may wish to be aware. It also may not be a bug.

At this point, what you do with this information is up to you.