tucker.py
from .base import unfold, tensor_to_vec
from .tenalg import multi_mode_dot
from .tenalg import kronecker
# Author: Jean Kossaifi <jean.kossaifi+tensors@gmail.com>
def tucker_to_tensor(core, factors, skip_factor=None, transpose_factors=False):
"""Converts the Tucker tensor into a full tensor
Parameters
----------
core : ndarray
core tensor
factors : ndarray list
list of matrices of shape (s_i, core.shape[i])
skip_factor : None or int, optional, default is None
if not None, index of a matrix to skip
Note that in any case, `modes`, if provided, should have a lengh of `tensor.ndim`
transpose_factors : bool, optional, default is False
if True, the matrices or vectors in in the list are transposed
Returns
-------
2D-array
full tensor of shape `(U[0].shape[0], ..., U[-1].shape[0])`
Notes
-----
This implementation is equivalent to:
>>> def tucker_to_tensor(G, U):
... for i, matrix in enumerate(U):
... if not i:
... res = mode_dot(G, matrix, i)
... else:
... res = mode_dot(res, matrix, i)
... return res
"""
return multi_mode_dot(core, factors, skip=skip_factor, transpose=transpose_factors)
def tucker_to_unfolded(core, factors, mode=0, skip_factor=None, transpose_factors=False):
"""Converts the Tucker decomposition into an unfolded tensor (i.e. a matrix)
Parameters
----------
G : ndarray
core tensor
U : ndarray list
list of matrices
mode : None or int list, optional, default is None
skip_factor : None or int, optional, default is None
if not None, index of a matrix to skip
Note that in any case, `modes`, if provided, should have a lengh of `tensor.ndim`
transpose_factors : bool, optional, default is False
if True, the matrices or vectors in in the list are transposed
Returns
-------
2D-array
unfolded tensor
"""
return unfold(tucker_to_tensor(core, factors, skip_factor=skip_factor, transpose_factors=transpose_factors), mode)
def tucker_to_vec(core, factors, skip_factor=None, transpose_factors=False):
"""Converts a Tucker decomposition into a vectorised tensor
Parameters
----------
core : ndarray
core tensor
factors : ndarray list
list of factor matrices
skip_factor : None or int, optional, default is None
if not None, index of a matrix to skip
Note that in any case, `modes`, if provided, should have a lengh of `tensor.ndim`
transpose_factors : bool, optional, default is False
if True, the matrices or vectors in in the list are transposed
Returns
-------
1D-array
vectorised tensor
Notes
-----
Mathematically equivalent but much slower,
you can obtain the same result using:
>>> def tucker_to_vec(core, factors):
... return kronecker(factors).dot(tensor_to_vec(core))
In this implementation we:
1* take the n-mode product of the core with all the factors
2* vectorize the result
(rather than computing potentially large kronecker product of factors)
"""
return tensor_to_vec(tucker_to_tensor(core, factors, skip_factor=skip_factor, transpose_factors=transpose_factors))