François CholletVerified account

If a third party writes buggy code and it involves a Keras model, that's not "a bug in Keras", for Christ's sake

Just for the record, trainable weight tracking in custom layers works fine and has always worked fine. You can test it yourself, it takes 7 lines of code. Custom layers with trainable weights are used in countless workflows. Can't believe I have to say this

That said, it's generally good practice to write unit tests for your layers, that test basic assumptions: number of weights, number of trainable weights, output shape, test that the layer produces the same output after deserialization, etc. It enables you to catch common mistakes

So do, in fact, check it yourself, via a unit test -- it's never wasted time. For any codebase that's actively maintained, you should always aim for 100% coverage