Creating factorized tensors¶

You can create a new factorized tensor easily:

The signature is:

factorized_tensor = FactorizedTensor.new(shape, rank, factorization)

For instance, to create a tensor in Tucker form, that has half the parameters of a dense (non-factorized) tensor of the same shape, you would simply write:

tucker_tensor = FactorizedTensor.new(shape, rank=0.5, factorization='tucker')

Since TensorLy-Torch builds on top of TensorLy, it also comes with tensor decomposition out-of-the-box. To initialize a factorized tensor in CP (Canonical-Polyadic) form, also known as Parafac, or Kruskal tensor, with 1/10th of the parameters, you can simply write:

cp_tensor = FactorizedTensor.new(dense_tensor, rank=0.1, factorization='CP')

Manipulating factorized tensors¶

The first thing you want to do, if you created a new tensor from scratch (by using the new method), is to initialize it, e.g. so that the element of the reconstruction approximately follow a Gaussian distribution:

cp_tensor.normal_(mean=0, std=0.02)

You can even use PyTorch’s functions! This works:

from torch.nn import init

init.kaiming_normal(cp_tensor)

Finally, you can index tensors directly in factorized form, which will return another factorized tensor, whenever possible!

>>> cp_tensor[:2, :2]
CPTensor(shape=(2, 2, 2), rank=2)

If not possible, a dense tensor will be returned:

>>> cp_tensor[2, 3, 1]
tensor(0.0250, grad_fn=<SumBackward0>)

Note how, above, indexing tracks gradients as well!

Creation¶

You can create one in TensorLy-Torch, from a matrix, just as easily as a regular tensor, using the tltorch.TensorizedTensor class, with the following signature:

TensorizedTensor.from_matrix(matrix, tensorized_row_shape, tensorized_column_shape, rank)

where tensorized_row_shape and tensorized_column_shape indicate the shape to which to tensorize the row and column size of the given matrix. For instance, if you have a matrix of size 16x21, you could use tensorized_row_shape=(4, 4) and tensorized_column_shape=(3, 7).

In general, you can tensorize any tensor, not just matrices, even with batched modes (dimensions)!

tensorized_tensor = TensorizedTensor.new(tensorized_shape, rank, factorization)

tensorized_shape is a nested tuple, in which an int represents a batched mode, and a tuple a tensorized mode.

For instance, a batch of 5 matrices of size 16x21 could be tensorized into a batch of 5 tensorized matrices of size (4x4)x(3x7), in the BlockTT form. In code, you would do this using

tensorized_tensor = TensorizedTensor.from_tensor(tensor, (5, (4, 4), (3, 7)), rank=0.7, factorization='BlockTT')

You can of course tensorize any size tensors, e.g. a batch of 5 matrices of size 8x27 can be tensorized into:

>>> ftt = tltorch.TensorizedTensor.new((5, (2, 2, 2), (3, 3, 3)), rank=0.5, factorization='BlockTT')

This returns a tensorized tensor, stored in decomposed form: >>> ftt BlockTT(shape=[5, 8, 27], tensorized_shape=(5, (2, 2, 2), (3, 3, 3)), rank=[1, 20, 20, 1])

Manipulation¶

As for factorized tensors, you can directly index them:

>>> ftt[2]
BlockTT(shape=[8, 27], tensorized_shape=[(2, 2, 2), (3, 3, 3)], rank=[1, 20, 20, 1])

>>> ftt[0, :2, :2]
tensor([[-0.0009,  0.0004],
        [ 0.0007,  0.0003]], grad_fn=<SqueezeBackward0>)

Again, notice that gradients are tracked and all operations on factorized and tensorized tensors are back-propagatable!