Himalaya
implements machine learning linear models in Python, focusing on
computational efficiency for large numbers of targets.
Use himalaya
if you need a library that:
- estimates linear models on large numbers of targets,
- runs on CPU and GPU hardware,
- provides estimators compatible with
scikit-learn
's API.
Himalaya
is stable (with particular care for backward compatibility) and
open for public use (give it a star!).
import numpy as np
n_samples, n_features, n_targets = 10, 5, 4
np.random.seed(0)
X = np.random.randn(n_samples, n_features)
Y = np.random.randn(n_samples, n_targets)
from himalaya.ridge import RidgeCV
model = RidgeCV(alphas=[1, 10, 100])
model.fit(X, Y)
print(model.best_alphas_) # [ 10. 100. 10. 100.]
- The model
RidgeCV
uses the same API asscikit-learn
estimators, with methods such asfit
,predict
,score
, etc. - The model is able to efficiently fit a large number of targets (routinely used with 100k targets).
- The model selects the best hyperparameter
alpha
for each target independently.
Check more examples of use of himalaya
in the gallery of examples.
Himalaya
was designed primarily for functional magnetic resonance imaging
(fMRI) encoding models. In depth tutorials about using himalaya
for fMRI
encoding models can be found at gallantlab/voxelwise_tutorials.
Himalaya
implements the following models:
- Ridge
- RidgeCV
- GroupRidgeCV
- KernelRidge
- KernelRidgeCV
- WeightedKernelRidge
- MultipleKernelRidgeCV
- SparseGroupLassoCV
See the model descriptions in the documentation website.
Himalaya
can be used seamlessly with different backends.
The available backends are numpy
(default), cupy
, and pytorch
.
To change the backend (e.g. to cupy
), call:
from himalaya.backend import set_backend
backend = set_backend("cupy")
and give cupy
arrays inputs to the himalaya
solvers. For convenience,
estimators implementing scikit-learn
's API can cast arrays to the correct
input type.
To run himalaya
on a graphics processing unit (GPU), you can use both
cupy
or pytorch
backends.
To use the cupy
backend, call:
from himalaya.backend import set_backend
backend = set_backend("cupy")
data = backend.asarray(data)
To use the pytorch
backend, call:
from himalaya.backend import set_backend
backend = set_backend("torch_cuda")
# "torch" uses pytorch on CPU, "torch_cuda" uses pytorch on GPU
data = backend.asarray(data)
Himalaya
requires:
- Python 3
- Numpy
- Scikit-learn
- PyTorch (optional GPU backend) (1.9+ preferred)
- Cupy (optional GPU backend)
- Matplotlib (optional, for visualization only)
- Pytest (optional, for testing only)
You may install the latest version of himalaya
using the package manager
pip
, which will automatically download himalaya
from the Python Package
Index (PyPI):
pip install himalaya
To install himalaya
from the latest source (main
branch), you may
call:
pip install git+https://github.com/gallantlab/himalaya.git
Developers can also install himalaya
in editable mode via:
git clone https://github.com/gallantlab/himalaya
cd himalaya
pip install --editable .
If you use himalaya
in your work, please give it a star and cite our
(future) publication:
[1] | Dupré La Tour, T., Eickenberg, M., & Gallant, J. L. (2021). Feature-space selection with banded ridge regression. In preparation. |