TensorFlow 2.13 introduces three problems raising errors in unit tests:
Module conv_utils from Keras, required for our convolutional layers, has been moved (again) in a different Keras folder. Depending on TF version, we import this module from different paths.
Serialization/deserialization of h5 models with custom objects (here, custom initializer) is not supported anymore in TF 2.13. The new .keras format is now preferred, instead of h5 format.
One of our unit test was failing (comparison of SVD and power iteration for a Dense kernel). This is not a problem from TF 2.13 but from non-deterministic computation: for some random cases, the power iteration algorithm had difficulties to converge close enough to the true value. This failing test is now deterministic with seeded kernel and initialization u.
TensorFlow 2.13 introduces three problems raising errors in unit tests:
conv_utilsfrom Keras, required for our convolutional layers, has been moved (again) in a different Keras folder. Depending on TF version, we import this module from different paths..kerasformat is now preferred, instead of h5 format.