Releases: reservoirpy/reservoirpy
Releases · reservoirpy/reservoirpy
ReservoirPy v0.3.11
Major changes
- Fixed
numpy.memmap
concurrent accesses. Which fixes many issues with parallelization. The memmap name is now unique for each buffer.
Fixes #141, fixes #112, fixes #57 - Sparse matrix initializers in
mat_gen
can now take adegree
argument instead ofconnectivity
. Delay
node, which returns its input with a specified delay.- Complete reimplementation of the
ScikitLearnNode
introduced in ReservoirPy v0.3.10. This has breaking changes, as the previous implementation was not consistent with the library node API. The tutorial has subsequently been updated. - Dropped support for Python 3.6. Python 3.7 should still work though it reached end-of-life.
Minor changes
dataset.narma
now takes au
parameter so that the input timeseries can be accessed. Fixes #142- Fix : The ESN node now has consistent results across backends
- Creating a Reservoir node with incorrect argument now raises a
ValueError
. Fixes #138 - Fix: the default
Node.partial_backward
method of offline nodes didn't concatenate output timeseries correctly. - Many small corrections in documentation. In particular:
ReservoirPy v0.3.10
Major changes
- Addition of a ScikitLearnNode : scikit-learn linear models can now be used as ReservoirPy's nodes. This can be used in particular for classification.
A detailed tutorial on how to use this new node can be found here.
This fixes #82
(by @Deepayan137 )
Minor changes
ReservoirPy v0.3.9.post1
Major Fix
ESN.run
had the same problem asESN.fit
when usingmultiprocessing
as joblib backend.
Full Changelog: v0.3.9...v0.3.9.post1
ReservoirPy v0.3.9
Major fixes
- Deterministic behavior of Concat node: fix a major issue with Concat nodes being fed inputs in random order (#115 #114 by @PAUL-BERNARD)
japanese_vowels
link was dead, the dataset could not be loaded (#113).- Using
multiprocessing
as joblib backend was failing because of a locale function inESN.fit
.
Minor Fixes
- Fixed rsquare averaging y_pred instead of y_true (#110 by @HugoChateauLaurent).
- Improved robustness of spectral radius computation for reproducibility (#116 by @PAUL-BERNARD) .
dataset
module coherence and precision: Add**kwargs
everywhere, fix some issues with timeseries length (#118 #117)- Typos in Reservoir equations (#101)
New Contributors
- @PAUL-BERNARD made their first contribution in #116
- @HugoChateauLaurent made their first contribution in #110
Full Changelog: v0.3.8...v0.3.9
ReservoirPy v0.3.8
Major changes
- Leak rate
lr
in reservoir nodes can now be an array or list of float values, one per neurons.
What's Changed
- V0.3.8 by @nTrouvain in #108
Full Changelog: v0.3.7...v0.3.8
ReservoirPy v0.3.7
Minor fixes
- Fix #97: noise is now consistently generated.
- Fix #94:
h
parameter can now be changed indatasets.mackey_glass
. - Fix #98: adding a
noise_kwargs
parameters to reservoir nodes (Reservoir and IPReservoir) to change noise distribution parameters.noise - Fix infinite loop of AttributeError on nodes.
Full Changelog: v0.3.6...v0.3.7
ReservoirPy v0.3.6
Minor fixes
Node.partial_fit
was not passingkwargs
to thepartial_backward
function (omitting thread lock when usingESN.fit
with several workers)ESN.fit
method missed a warmup argument.
ReservoirPy v0.3.5
Minor fixes
- ReservoirPy was changing the default tempfile directory for the entire environement.
ReservoirPy v0.3.4
New features
- Add Lorenz96, Rössler and Kuramoto-Sivashinsky attractors/oscillators in
datasets
!
Minor changes
FORCE
class is now split in two:RLS
andLMS
classes, to avoid confusion between FORCE algorithm and RLS/LMS learning rules.FORCE
is still available but deprecated.- Typo in documentation
What's Changed
- Patch 0.3.2.post1 by @nTrouvain in #66
- v0.3.2 -> v0.3.3 by @nTrouvain in #69
- observables.py: add shape verification by @AlArths in #71
- Dev by @nTrouvain in #72
New Contributors
Full Changelog: v0.3.2...v0.3.4
ReservoirPy v0.3.3
New features
- Add
japanese_vowels
todatasets
! You can now dowload the Japanese vowels dataset to try Reservoir Computing on a well known pattern recognition task.
Major fixes
- Incorrect import of
ArpackNoConvergence
Exception (#67) - Models with multiple inputs and complex branches were not created properly (
Concat
nodes were not created at the right place) (#68) Model.run
can now be called on multiple series (using lists of arrays or arrays with ndim > 2)
Minor fixes
Lock
was used too soon inRidge.partial_fit
. Dot product is now performed in parallel.- Better data handling with complex models.
- New
Unsupervised
node subclass can be used to mark a Node as unsupervised learner. - Update random state generator to
numpy.random.Generator
inhyper.research
. - Other minor fixes in examples and typos.
What's Changed
- Patch 0.3.2.post1 by @nTrouvain in #66
Full Changelog: v0.3.2...v0.3.3