While this provided some performance benefits, it also produced some surprising behavior for users in eager mode.
We plan to switch on the new code path by default in Tf.random.Generator in keras backend, which will be the new backend forĪll the RNG in Keras. Added 3 new APIs for enable/disable/check the usage of.
tf.random.Generator for keras initializers and all RNG code.The output type of any preprocessing layer can be controlled individuallyīy passing a dtype argument to the layer. Output_mode="int" in which case output will be tf.int64. Tf.keras.mixed_precision.Policy, unless constructed with All preprocessing layer output will follow the compute dtype of a.AllĬategorical preprocessing layers now support output_mode. Layers with the same semantics as other preprocessing layers. Added an output_mode argument to the Discretization and Hashing.split="character" will split on every unicode character.standardize="string_punctuation" will remove all puncuation.standardize="lower" will lowercase inputs.Added additional standardize and split modes to TextVectorization.Tf.sparse.cross/ tf.ragged.cross directly. Should migrate to the HashedCrossing layer or use This provides a stateless way to try adding feature crosses Layer which applies the hashing trick to the concatenation of crossed