TensorKit offers a scikit-learn like API to tensor learning in Python.
Currently, we support CP and PARAFAC2 decomposition using alternating least squares and we are working on coupled decompositions.
Note that the code is still under development so the API may be subject to change.
TensorKit is created by Marie Roald and Yngve Mardal Moe.
TensorLy is a good way into providing tensor learning in Python, however, we found that taking a more object-oriented approach was more useful when conducting many experiments. Also, it should be easier to start using TensorKit for people who have experience with scikit-learn.
To install the latest version, run the following commands
git clone https://github.com/marieroald/tensorkit
cd TensorKit
python setup.pyBelow is an example where we create a random Kruskal tensor and decompose it using the CP decomposition.
import numpy as np
from tenkit.decomposition import decompositions
from tenkit.decomposition.cp import CP_ALS
# Generate random tensor
shape = (30, 40, 50)
rank = 4
random_tensor = decompositions.KruskalTensor.random_init(shape, rank)
# Add noise
noise_level = 0.3
tensor = random_tensor.construct_tensor()
noise = np.random.standard_normal(tensor.shape)
noise *= noise_level*np.linalg.norm(tensor)/np.linalg.norm(noise)
noisy_tensor = tensor + noise
# Fit a CP model
cp = CP_ALS(rank)
learned_decomposition = cp.fit_transform(noisy_tensor)
# Evaluate performance
fms, permutation = random_tensor.factor_match_score(learned_decomposition)
print(f'The factor match score is {fms:.3e} and the factor permutation is {permutation}')