Read the Docs site: http://tntorch.readthedocs.io/

Welcome to tntorch, a PyTorch-powered modeling and learning library using tensor networks. Such networks are unique in that they use multilinear neural units (instead of non-linear activation units). Features include:

• Basic and fancy indexing of tensors, broadcasting, assignment, etc.
• Tensor decomposition and reconstruction
• Element-wise and tensor-tensor arithmetics
• Building tensors from black-box functions using cross-approximation
• Finding global maxima and minima from tensors
• Statistics and sensitivity analysis
• Optimization using autodifferentiation
• Misc. operations on tensors: stacking, unfolding, sampling, derivating, etc.
• Batch operations (work in progress)

If you use this package, please cite our paper:

@article{UBS:22,
  author  = {Mikhail Usvyatsov and Rafael Ballester-Ripoll and Konrad Schindler},
  title   = {tntorch: Tensor Network Learning with {PyTorch}},
  journal = {Journal of Machine Learning Research},
  year    = {2022},
  volume  = {23},
  number  = {208},
  pages   = {1--6},
  url     = {http://jmlr.org/papers/v23/21-1197.html}
}

Available tensor formats include:

• CANDECOMP/PARAFAC (CP)
• Tucker (implemented as TT with increasing ranks, which has equal expressive power. Tucker factors are unconstrained matrices, unlike unitary/orthogonal matrices in some implementations.)
• Tensor train (TT)
• Hybrids: CP-Tucker, TT-Tucker, etc.
• Partial support for other decompositions such as INDSCAL, CANDELINC, DEDICOM, PARATUCK2, and custom formats

For example, the following networks both represent a 4D tensor (i.e. a real function that can take I1 x I2 x I3 x I4 possible values) in the TT and TT-Tucker formats:

In tntorch, all tensor decompositions share the same interface. You can handle them in a transparent form, as if they were plain NumPy arrays or PyTorch tensors:

> import tntorch as tn
> t = tn.randn(32, 32, 32, 32, ranks_tt=5)  # Random 4D TT tensor of shape 32 x 32 x 32 x 32 and TT-rank 5
> print(t)

4D TT tensor:

 32  32  32  32
  |   |   |   |
 (0) (1) (2) (3)
 / \ / \ / \ / \
1   5   5   5   1

> print(tn.mean(t))

tensor(8.0388)

> print(tn.norm(t))

tensor(9632.3726)

Decompressing tensors is easy:

> print(t.torch().shape)
torch.Size([32, 32, 32, 32])

Thanks to PyTorch's automatic differentiation, you can easily define all sorts of loss functions on tensors:

def loss(t):
    return torch.norm(t[:, 0, 10:, [3, 4]].torch())  # NumPy-like "fancy indexing" for arrays

Most importantly, loss functions can be defined on compressed tensors as well:

def loss(t):
    return tn.norm(t[:3, :3, :3, :3] - t[-3:, -3:, -3:, -3:])

Check out the introductory notebook for all the details on the basics.

• Introduction
• Active subspaces
• ANOVA decomposition
• Boolean logic
• Classification
• Cross-approximation
• Differentiable cross-approximation
• Differentiation
• Discrete/weighted finite automata
• Exponential machines
• Main tensor formats available
• Other custom formats
• Polynomial chaos expansions
• Tensor arithmetics
• Tensor completion and regression
• Tensor decomposition
• Sensitivity analysis
• Vector field data

You can install tntorch using pip:

pip install tntorch

Alternatively, you can install from the source:

git clone https://github.com/rballester/tntorch.git
cd tntorch
pip install .

For functions that use cross-approximation, the optional package maxvolpy is required (it can be installed via pip install maxvolpy).

We use pytest. Simply run:

cd tests/
pytest

Pull requests are welcome!

Besides using the issue tracker, feel also free to contact me at rafael.ballester@ie.edu.