Haiku 0.0.5
- Added support for mixed precision training (dba1fd9) via jmp
- Added
hk.with_empty_state(..)
. - Added
hk.multi_transform(..)
(#137), supporting transforming multiple functions that share parameters. - Added
hk.data_structures.is_subset(..)
to test whether parameters are a subset of another. - Minimum Python version is now 3.7.
- Multiple changes in preparation for a future version of Haiku changing to plain
dict
s. hk.next_rng_keys(..)
now returns a stacked array rather than a collection.hk.MultiHeadAttention
now supports distinct sequence lengths in query and key/value.hk.LayerNorm
now optionally supports faster (but less stable) variance computation.hk.nets.MLP
now has an output_shape property.hk.nets.ResNet
now supports changing strides.UnexpectedTracerError
inside a Haiku transform now has a more useful error message.hk.{lift,custom_creator,custom_getter}
are no longer experimental.- Haiku now supports JAX's pluggable RNGs.
- We have made multiple improvements to our docs an error messages.
Any many other small fixes and improvements.