Releases: poets-ai/elegy
Releases · poets-ai/elegy
0.8.6
0.8.5
0.8.4
0.8.3
0.8.2
0.8.1
Elegy is now based on Treex 🎉
Changes
- Remove the
module
,nn
,metrics
, andlosses
from Elegy, instead Elegy reexports these modules from Treex. GeneralizedModule
and friends are gone, to use Flax Modules use theelegy.nn.FlaxModule
wrapper.- Low level API is massively simplified:
States
is removed, sinceModel
is a pytree all parameters are tracked automatically thanks to Treex / Treeo.- All static state arguments (
training
,initializing
) are removed,Module
s can simply useself.training
to pick their training state andself.initializing()
to check whether they are initializing. - Signature for
pred_step
,test_step
, andtrain_step
now simply consists ofinputs
andlabels
, wherelabels
is adict
that can contain additional keys likesample_weight
orclass_weight
as required by the losses and metrics.
- Adds the
DistributedStrategy
class which currently has 3 instancesEager
: Runs model in a single device in eager mode (nojit
)JIT
: Runs model in a single device withjit
DataParallel
: Run the model in multiple devices usingpmap
.
- Adds methods to change the model's distributed strategy:
.distributed(strategy = DataParallel)
: changes the distributed strategy,DataParallel
used by default..local()
: changes the distributed strategy toJIT
..eager()
: changes the distributed strategy toEager
.
- Removes the
.eager
field in favor of the.eager()
method.
0.7.4
0.7.2
0.7.1
0.7.0
Features
init
now only called once internally and required to be called explicitly by the user under certain circumstancessummary
now usesjax.eval_shape
under the hood so its super fast since it doesn't allocate memory or perform any computations on the device.
Merged pull requests: