https://github.com/tensorly/tensorly
Raw File
plugins.py
import tensorly as tl

# Author: Jean Kossaifi


PREVIOUS_EINSUM = None
OPT_EINSUM_PATH_CACHE = dict()
CUQUANTUM_PATH_CACHE = dict()
CUQUANTUM_HANDLE = None


def use_default_einsum():
    """Revert to the original einsum for the current backend"""
    global PREVIOUS_EINSUM

    if PREVIOUS_EINSUM is not None:
        tl.backend.BackendManager.register_backend_method("einsum", PREVIOUS_EINSUM)
        PREVIOUS_EINSUM = None


def use_opt_einsum(optimize="auto-hq"):
    """Plugin to use opt-einsum [1]_ to precompute (and cache) a better contraction path

    Examples
    --------
    >>> import tensorly as tl

    Use your favourite backend, here PyTorch:
    >>> tl.set_backend('pytorch')

    Use the convenient backend system to automatically dispatch all tenalg operations to einsum

    >>> from tensorly import tenalg
    >>> tenalg.set_backend('einsum')

    Now you can transparently cache the optimal contraction path, transparently:

    >>> from tensorly import plugins
    >>> plugins.use_opt_einsum()

    That's it! You can revert to the original einsum just as easily:

    >>> plugings.use_default_einsum()

    Revert to the original tensor algebra backend:

    >>> tenalg.set_backend('core')

    References
    ----------
    .. [1] Daniel G. A. Smith and Johnnie Gray, opt_einsum,
           A Python package for optimizing contraction order for einsum-like expressions.
           Journal of Open Source Software, 2018, 3(26), 753
    """
    global PREVIOUS_EINSUM

    try:
        import opt_einsum as oe
    except ImportError as error:
        message = (
            "Impossible to import opt-einsum.\n"
            "First install it:\n"
            "conda install opt_einsum -c conda-forge\n"
            " or pip install opt_einsum"
        )
        raise ImportError(message) from error

    def cached_einsum(equation, *args):
        shapes = [tl.shape(arg) for arg in args]
        key = equation + "," + ",".join(f"{s}" for s in shapes)
        try:
            expression = OPT_EINSUM_PATH_CACHE[key]
        except KeyError:
            expression = oe.contract_expression(equation, *shapes, optimize=optimize)
            OPT_EINSUM_PATH_CACHE[key] = expression

        return expression(*args)

    if PREVIOUS_EINSUM is None:
        PREVIOUS_EINSUM = tl.backend.current_backend().einsum

    tl.backend.BackendManager.register_backend_method("einsum", cached_einsum)


def use_cuquantum(optimize="auto-hq"):
    """Plugin to use `cuQuantum <https://developer.nvidia.com/cuquantum-sdk>`_ to precompute (and cache) a better contraction path

    Examples
    --------
    >>> import tensorly as tl

    Use your favourite backend, here PyTorch:

    >>> tl.set_backend('pytorch')

    Use the convenient backend system to automatically dispatch all tenalg operations to einsum

    >>> from tensorly import tenalg
    >>> tenalg.set_backend('einsum')

    Now you can transparently cache the optimal contraction path, transparently:

    >>> from tensorly import plugins
    >>> plugins.use_cuquantum()

    That's it! Now opt-einsum will be used for finding an (near) optimal contraction path
    and cuQuantum will be used to actually perform the tensor contractions!

    You can revert to the original einsum just as easily:

    >>> plugings.use_default_einsum()

    Revert to the original tensor algebra backend:

    >>> tenalg.set_backend('core')
    """
    global PREVIOUS_EINSUM
    global CUQUANTUM_HANDLE

    # Import opt-einsum for the contraction path
    try:
        import opt_einsum as oe
    except ImportError as error:
        message = (
            "Impossible to import opt-einsum.\n"
            "First install it:\n"
            "conda install opt_einsum -c conda-forge\n"
            " or pip install opt_einsum"
        )
        raise ImportError(message) from error

    # Import cuQuantum for the actual contraction
    try:
        import cuquantum
    except ImportError as error:
        message = (
            "Impossible to import cuquantum.\n"
            "First install it:\n"
            "conda install -c conda-forge cuquantum-python\n"
            " or pip install cuquantum-python"
        )
        raise ImportError(message) from error

    if CUQUANTUM_HANDLE is None:
        CUQUANTUM_HANDLE = cuquantum.cutensornet.create()

    def cached_einsum(equation, *args):
        shapes = [tl.shape(arg) for arg in args]
        key = equation + "," + ",".join(f"{s}" for s in shapes)
        try:
            path = CUQUANTUM_PATH_CACHE[key]
        except KeyError:
            path, _ = oe.contract_path(equation, *args, optimize=optimize)
            CUQUANTUM_PATH_CACHE[key] = path

        network = cuquantum.Network(
            equation, *args, options={"handle": CUQUANTUM_HANDLE}
        )
        network.contract_path(optimize={"path": path})
        return network.contract()

        # return cuquantum.contract(equation, *args, optimize={'path': path})

    if PREVIOUS_EINSUM is None:
        PREVIOUS_EINSUM = tl.backend.current_backend().einsum

    tl.backend.BackendManager.register_backend_method("einsum", cached_einsum)
back to top