Revision 1295ccb09626f89f20d0c0183d618f96b4833bf1 authored by Jean Kossaifi on 08 May 2018, 21:04:53 UTC, committed by Jean Kossaifi on 08 May 2018, 22:15:23 UTC
1 parent c729db7
Raw File
numpy_backend.py
"""
Core tensor operations.
"""

from numpy import testing
import numpy as np
import scipy.linalg
import scipy.sparse.linalg

from numpy import reshape, moveaxis, where, copy, transpose
from numpy import arange, ones, zeros, zeros_like
from numpy import dot, kron, concatenate
from numpy import max, min, maximum, all, mean, sum, sign, abs, prod, sqrt
from numpy.linalg import solve, qr

# Author: Jean Kossaifi

# License: BSD 3 clause


def context(tensor):
    """Returns the context of a tensor

        Creates a dictionary of the parameters characterising the tensor

    Parameters
    ----------
    tensor : tensorly.tensor

    Returns
    -------
    context : dict

    Examples
    --------
    >>> import tensorly as tl
    Using numpy backend.

    Imagine you have an existing tensor `tensor`:

    >>> import numpy as np
    >>> tensor = tl.tensor([0, 1, 2], dtype=np.float32)

    The context, here, will simply be the dtype:

    >>> tl.context(tensor)
    {'dtype': dtype('float32')}
    
    Note that, if you were using, say, PyTorch, the context would also
    include the device (i.e. CPU or GPU) and device ID.

    If you want to create a new tensor in the same context, use this context:

    >>> new_tensor = tl.tensor([1, 2, 3], **tl.context(tensor))
    """
    return {'dtype':tensor.dtype}

def tensor(data, dtype=np.float64):
    """Tensor class
        
        Returns a tensor on the specified context, depending on the backend
    """
    return np.array(data, dtype=dtype)


def to_numpy(tensor):
    """Returns a copy of the tensor as a NumPy array

    Parameters
    ----------
    tensor : tl.tensor

    Returns
    -------
    numpy_tensor : numpy.ndarray
    """
    return np.copy(tensor)

def assert_array_equal(a, b, **kwargs):
    return testing.assert_array_equal(a, b, **kwargs)

def assert_array_almost_equal(a, b, **kwargs):
    testing.assert_array_almost_equal(to_numpy(a), to_numpy(b), **kwargs)

assert_raises = testing.assert_raises
assert_equal = testing.assert_equal
assert_ = testing.assert_

def shape(tensor):
    return tensor.shape

def ndim(tensor):
    return tensor.ndim

def clip(tensor, a_min=None, a_max=None, inplace=False):
    return np.clip(tensor, a_min, a_max)

def norm(tensor, order=2, axis=None):
    """Computes the l-`order` norm of tensor

    Parameters
    ----------
    tensor : ndarray
    order : int
    axis : int or tuple

    Returns
    -------
    float or tensor
        If `axis` is provided returns a tensor.
    """
    # handle difference in default axis notation
    if axis == ():
        axis = None

    if order == 'inf':
        return np.max(np.abs(tensor), axis=axis)
    if order == 1:
        return np.sum(np.abs(tensor), axis=axis)
    elif order == 2:
        return np.sqrt(np.sum(tensor**2, axis=axis))
    else:
        return np.sum(np.abs(tensor)**order, axis=axis)**(1/order)

def kr(matrices):
    """Khatri-Rao product of a list of matrices

        This can be seen as a column-wise kronecker product.

    Parameters
    ----------
    matrices : ndarray list
        list of matrices with the same number of columns, i.e.::

            for i in len(matrices):
                matrices[i].shape = (n_i, m)

    Returns
    -------
    khatri_rao_product: matrix of shape ``(prod(n_i), m)``
        where ``prod(n_i) = prod([m.shape[0] for m in matrices])``
        i.e. the product of the number of rows of all the matrices in the product.

    Notes
    -----
    Mathematically:

    .. math::
         \\text{If every matrix } U_k \\text{ is of size } (I_k \\times R),\\\\
         \\text{Then } \\left(U_1 \\bigodot \\cdots \\bigodot U_n \\right) \\text{ is of size } (\\prod_{k=1}^n I_k \\times R)

    A more intuitive but slower implementation is::

        kr_product = np.zeros((n_rows, n_columns))
        for i in range(n_columns):
            cum_prod = matrices[0][:, i]  # Acuumulates the khatri-rao product of the i-th columns
            for matrix in matrices[1:]:
                cum_prod = np.einsum('i,j->ij', cum_prod, matrix[:, i]).ravel()
            # the i-th column corresponds to the kronecker product of all the i-th columns of all matrices:
            kr_product[:, i] = cum_prod

        return kr_product
    """
    n_columns = matrices[0].shape[1]
    n_factors = len(matrices)

    start = ord('a')
    common_dim = 'z'
    target = ''.join(chr(start + i) for i in range(n_factors))
    source = ','.join(i+common_dim for i in target)
    operation = source+'->'+target+common_dim
    return np.einsum(operation, *matrices).reshape((-1, n_columns))

def partial_svd(matrix, n_eigenvecs=None):
    """Computes a fast partial SVD on `matrix`

        if `n_eigenvecs` is specified, sparse eigendecomposition
        is used on either matrix.dot(matrix.T) or matrix.T.dot(matrix)

    Parameters
    ----------
    matrix : 2D-array
    n_eigenvecs : int, optional, default is None
        if specified, number of eigen[vectors-values] to return

    Returns
    -------
    U : 2D-array
        of shape (matrix.shape[0], n_eigenvecs)
        contains the right singular vectors
    S : 1D-array
        of shape (n_eigenvecs, )
        contains the singular values of `matrix`
    V : 2D-array
        of shape (n_eigenvecs, matrix.shape[1])
        contains the left singular vectors
    """
    # Check that matrix is... a matrix!
    if matrix.ndim != 2:
        raise ValueError('matrix be a matrix. matrix.ndim is {} != 2'.format(
            matrix.ndim))

    # Choose what to do depending on the params
    dim_1, dim_2 = matrix.shape
    if dim_1 <= dim_2:
        min_dim = dim_1
    else:
        min_dim = dim_2

    if n_eigenvecs is None or n_eigenvecs >= min_dim:
        # Default on standard SVD
        U, S, V = scipy.linalg.svd(matrix)
        U, S, V = U[:, :n_eigenvecs], S[:n_eigenvecs], V[:n_eigenvecs, :]
        return U, S, V

    else:
        # We can perform a partial SVD
        # First choose whether to use X * X.T or X.T *X
        if dim_1 < dim_2:
            S, U = scipy.sparse.linalg.eigsh(np.dot(matrix, matrix.T.conj()), k=n_eigenvecs, which='LM')
            S = np.sqrt(S)
            V = np.dot(matrix.T.conj(), U * 1/S[None, :])
        else:
            S, V = scipy.sparse.linalg.eigsh(np.dot(matrix.T.conj(), matrix), k=n_eigenvecs, which='LM')
            S = np.sqrt(S)
            U = np.dot(matrix, V) * 1/S[None, :]

        # WARNING: here, V is still the transpose of what it should be
        U, S, V = U[:, ::-1], S[::-1], V[:, ::-1]
        return U, S, V.T.conj()
back to top