Source code for onnx_extended.tools.einsum.einsum_impl_ext

from typing import Dict, List, Optional, Tuple, Union
import numpy


def numpy_diagonal(m: numpy.ndarray, axis: int, axes: Tuple[int, ...]) -> numpy.ndarray:
    """
    Extracts diagonal coefficients from an array.

    :param m: input array
    :param axis: kept axis among the diagonal ones
    :param axes: diagonal axes (axis must be one of them)
    :return: output

    .. runpython::
        :showcode:

        import numpy
        from onnx_extended.tools.einsum import numpy_diagonal

        mat = numpy.arange(8).reshape((2, 2, 2))
        print(mat)
        diag = numpy_diagonal(mat, 1, [1, 2])
        print(diag)
    """
    if axis not in axes:
        raise RuntimeError(f"axis {axis!r} must be in axes {axes!r}.")
    shape = []
    new_shape = []
    for i, s in enumerate(m.shape):
        if i in axes:
            if i == axis:
                shape.append(s)
                new_shape.append(s)
            else:
                shape.append(1)
        else:
            shape.append(s)
            new_shape.append(s)

    # Extracts coefficients.
    output = numpy.empty(tuple(shape), dtype=m.dtype)
    index_in: List[Union[int, slice]] = [slice(s) for s in m.shape]
    index_out: List[Union[int, slice]] = [slice(s) for s in m.shape]
    for i in range(0, shape[axis]):
        for a in axes:
            index_in[a] = i
            index_out[a] = i if a == axis else 0
        output[tuple(index_out)] = m[tuple(index_in)]

    # Removes axis.
    return output.reshape(tuple(new_shape))


def _numpy_extended_dot_equation(
    m1_dim: int,
    m2_dim: int,
    axes: Tuple[int, ...],
    left: Tuple[int, ...],
    right: Tuple[int, ...],
) -> str:
    """
    Returns the equation equivalent to an extended version
    of an aligned matrix multiplication
    (see :func:`numpy_extended_dot
    <onnx_extended.tools.einsum.einsum_impl_ext.numpy_extended_dot>`).

    :param m1: number of dimensions of the first matrix
    :param m2: number of dimensions of the second matrix
    :param axes: summation axes
    :param axes: summation axes
    :param left: left axes
    :param right: right axes
    :return: equation

    .. runpython::
        :showcode:

        import numpy
        from onnx_extended.tools.einsum.einsum_impl_ext import (
            numpy_extended_dot_python, _numpy_extended_dot_equation)

        a = numpy.arange(6).reshape((3, 2, 1))
        b = numpy.arange(12).reshape((3, 1, 4))

        print(numpy_extended_dot_python(
            a, b, axes=(0, ), left=(1,), right=(2,)))

        # Equivalent einsum equation
        print('equation', _numpy_extended_dot_equation(
            len(a.shape), len(a.shape), axes=(0, ), left=(1,), right=(2,)))

        # Same einsum computation written in a different way.
        print(numpy.einsum('kix,kxj->xij', a, b))
    """
    assert m1_dim == m2_dim, (
        "Matrices m1 and m2 must have the same number of dimensions, "
        "m1=%r, m2=%r." % (m1_dim, m2_dim)
    )
    total = set(axes) | set(left) | set(right)
    assert len(total) <= m1_dim, (
        "Whole set of involved axes should be inferior to the number "
        "of dimensions: %r = {%r} | {%r} | {%r} has more than %d elements"
        "." % (total, axes, left, right, m1_dim)
    )

    def _check_(axs, n):
        for a in axs:
            if a < 0 or a >= n:
                raise ValueError(
                    "One axis %d (in %r) is negative or above the maximum "
                    "dimension %d." % (a, axs, n)
                )

    _check_(axes, m1_dim)
    _check_(left, m1_dim)
    _check_(right, m1_dim)

    l1 = [chr(i + 97) for i in range(m1_dim)]
    l2 = [chr(i + 97) for i in range(m1_dim)]
    l3: List[Optional[str]] = [chr(i + 97) for i in range(m1_dim)]
    for a in left:
        l1[a] = l1[a].upper()
        l3[a] = l3[a].upper()
    for a in right:
        l2[a] = l2[a].upper()
        l3[a] = l3[a].upper()
    for a in axes:
        l1[a] = l1[a].lower()
        l2[a] = l2[a].lower()
        if a not in right:
            l3[a] = None
        else:
            l3[a] = l3[a].lower()
    eq = f"{''.join(l1)},{''.join(l2)}->{''.join(s for s in l3 if s)}"
    return eq


def _common_check_numpy_extended_dot(
    m1: numpy.ndarray,
    m2: numpy.ndarray,
    axes: Tuple[int, ...],
    left: Tuple[int, ...],
    right: Tuple[int, ...],
):
    """
    Common verifications for all implementations of
    :func:`numpy_extended_dot
     <onnx_extended.tools.einsum.einsum_impl_ext.numpy_extended_dot>`.
    """
    assert (
        m1.dtype == m2.dtype
    ), f"Both matrices should share the same dtype {m1.dtype!r} != {m2.dtype!r}."
    m1_dim = len(m1.shape)
    m2_dim = len(m2.shape)
    assert m1_dim == m2_dim, (
        "Matrices m1 and m2 must have the same number of dimensions, "
        "m1=%r, m2=%r." % (m1_dim, m2_dim)
    )
    total = set(axes) | set(left) | set(right)
    assert len(total) <= m1_dim, (
        "Whole set of involved axes should be inferior to the number "
        "of dimensions: %r = {%r} | {%r} | {%r} has more than %d elements"
        "." % (total, axes, left, right, m1_dim)
    )


[docs] def numpy_extended_dot( m1: numpy.ndarray, m2: numpy.ndarray, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...], verbose: bool = False, ) -> numpy.ndarray: """ Extended version of a matrix multiplication (:func:`numpy.dot`) with two matrices *m1*, *m2* of the same dimensions. Loops over *left* axes for *m1* and *right* axes for *m2*, summation is done over *axes*. Other axes must be empty. This multiplication combines matrix multiplication (dot) and broadcasted multiplication term by term. :param m1: first matrix :param m2: second matrix :param axes: summation axes :param left: left axes :param right: right axes :param verbose: display intermediate information :return: output The dot product is equivalent to: .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot m1 = numpy.arange(4).reshape((2, 2)) m2 = m1 + 10 print("dot product") print(m1 @ m2) dm1 = m1.reshape((2, 2, 1)) dm2 = m2.reshape((1, 2, 2)) dot = numpy_extended_dot(dm1, dm2, axes=[1], left=[0], right=[2], verbose=True) print("extended dot product") print(dot) Empty axes should be squeezed to get identical results. Dot product when the second matrix is transposed. .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot m1 = numpy.arange(4).reshape((2, 2)) m2 = m1 + 10 print("dot product") print(m1 @ m2.T) dm1 = m1.reshape((2, 1, 2)) dm2 = m2.reshape((1, 2, 2)) dot = numpy_extended_dot(dm1, dm2, axes=[2], left=[0], right=[1], verbose=True) print("extended dot product") print(dot) An example when right axes include the summation axis. .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot m1 = numpy.arange(4).reshape((2, 2)) m2 = m1 + 10 dm1 = m1.reshape((2, 2, 1)) dm2 = m2.reshape((1, 2, 2)) dot = numpy_extended_dot(dm1, dm2, axes=[2], left=[0], right=[1, 2], verbose=True) print(dot) Example in higher dimension: .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot m1 = numpy.arange(8).reshape((2, 2, 2)) m2 = m1 + 10 dot = numpy_extended_dot(m1, m2, [1], [0], [2], verbose=True) print(dot) The current implementation still uses :func:`numpy.einsum` but this should be replaced. """ _common_check_numpy_extended_dot(m1, m2, axes, left, right) eq = _numpy_extended_dot_equation(len(m1.shape), len(m2.shape), axes, left, right) if verbose: print(f" [numpy_extended_dot] {eq}: {m1.shape!r} @ {m2.shape!r}") output = numpy.einsum(eq, m1, m2) new_shape = list(output.shape) for a in axes: if a not in right: new_shape.insert(a, 1) if verbose: print(f" [numpy_extended_dot] {output.shape!r} reshaped into {new_shape!r} ") return output.reshape(tuple(new_shape))
def numpy_extended_dot_ouput_shape( m1: numpy.ndarray, m2: numpy.ndarray, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...], ) -> numpy.ndarray: """ Computes the output shape of results produced by function :func:`numpy_extended_dot <onnx_extended.tools.einsum_impl_ext.numpy_extended_dot>` or :func:`numpy_extended_dot_python <onnx_extended.tools.einsum_impl_ext.numpy_extended_dot_python>`. """ _common_check_numpy_extended_dot(m1, m2, axes, left, right) m1_dim = len(m1.shape) new_shape = numpy.full(m1_dim, 1, dtype=numpy.int64) for i in left: new_shape[i] = m1.shape[i] for i in right: if ( i in left and m1.shape[i] != m2.shape[i] and m1.shape[i] != 1 and m2.shape[i] != 1 ): raise RuntimeError( "Matrices should have the same dimension for dimension %d, " "shapes=%r @ %r." % (i, m1.shape, m2.shape) ) new_shape[i] = m2.shape[i] return new_shape def _numpy_extended_dot_python_l1l2l3( m1_dim: int, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...] ) -> Tuple[List[str], List[str], List[Optional[str]]]: l1 = [chr(i + 97) for i in range(m1_dim)] l2 = [chr(i + 97) for i in range(m1_dim)] l3: List[Optional[str]] = [chr(i + 97) for i in range(m1_dim)] for a in left: l1[a] = l1[a].upper() l3[a] = l3[a].upper() for a in right: l2[a] = l2[a].upper() l3[a] = l3[a].upper() for a in axes: l1[a] = l1[a].lower() l2[a] = l2[a].lower() if a not in right: l3[a] = "-" else: l3[a] = l3[a].lower() return l1, l2, l3 def _numpy_extended_dot_python_intermediate( m1_shape: Tuple[int, ...], m2_shape: Tuple[int, ...], l1: List[str], l2: List[str], l3: List[Optional[str]], ) -> Tuple[ List[str], numpy.ndarray, Dict[str, int], List[bool], List[bool], numpy.ndarray ]: names = list(sorted(set(l1 + l2))) kind = numpy.zeros(len(names), dtype=numpy.int64) cols = {} for i, n in enumerate(names): if n in l1: kind[i] += 1 cols[n] = l1.index(n) if n in l2: kind[i] += 2 cols[n] = l2.index(n) if n in l3: kind[i] += 4 pos = numpy.zeros(len(names), dtype=numpy.int64) for j in range(0, pos.shape[0]): pos[j] = cols[names[j]] common = [(kind[i] & 3) == 3 for i in range(len(kind))] broadcast = [ common[i] and m1_shape[pos[i]] != m2_shape[pos[i]] for i in range(len(common)) ] return names, kind, cols, common, broadcast, pos def _numpy_extended_dot_python_update_broadcast( m1: numpy.ndarray, m2: numpy.ndarray, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...], l1: List[str], l2: List[str], l3: List[Optional[str]], names: List[str], broadcast: List[bool], cols: Dict[str, int], kind: numpy.ndarray, common: List[bool], verbose: bool = False, ) -> Tuple[List[str], List[str], List[Optional[str]]]: def dispb(c): return "".join("o" if b else "." for b in c) if verbose: print( "[GENERICDOT] before broadcast %s,%s->%s or %s" % ( "".join(l1), "".join(l2), "".join([(s if s else "?") for s in l3]), _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ), ) ) print( "[GENERICDOT] names=%s kind=%r common=%s broadcast=%s" % ("".join(names), kind.tolist(), dispb(common), dispb(broadcast)) ) for i in range(len(broadcast)): if broadcast[i] and not (kind[i] & 3) == 3: raise RuntimeError( "Broadcast should only happen on common axes, " "axes=%r left=%r right=%r shape1=%r shape2=%r." "" % (axes, left, right, m1.shape, m2.shape) ) if not broadcast[i]: continue # We split letters. p = cols[names[i]] dim = (m1.shape[p], m2.shape[p]) let = [l1[p], l2[p], l3[p]] inp = 1 if dim[0] == 1 else 0 if verbose: print( "[GENERICDOT] name=%s dim=%r let=%r inp=%r p=%r" % (names[i], dim, let, inp, p) ) print(f" B0 l1={l1!r}, l2={l2!r} l3={l3!r}") if (kind[i] & 4) > 0: # Summation axis is part of the output. assert ( let[inp] is not None ), f"Unexpected value for let[{inp}] in let={let}." if let[inp].lower() == let[inp]: let[inp] = let[inp].upper() else: let[inp] = let[inp].lower() l3[p] = let[inp] if inp == 1: l2[p] = let[inp] else: l1[p] = let[inp] if verbose: print(f" B1 l1={l1!r}, l2={l2!r} l3={l3!r}") else: # Summation axis is not part of the output. assert ( let[inp] is not None ), f"Unexpected value for let[{inp}] in let={let}." if let[inp].lower() == let[inp]: let[inp] = let[inp].upper() else: let[inp] = let[inp].lower() if inp == 1: l2[p] = let[inp] else: l1[p] = let[inp] if verbose: print(f" B2 l1={l1!r}, l2={l2!r} l3={l3!r}") return l1, l2, l3
[docs] def numpy_extended_dot_python( m1: numpy.ndarray, m2: numpy.ndarray, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...], verbose: bool = False, ) -> numpy.ndarray: """ Implementation of :func:`numpy_extended_dot <onnx_extended.tools.einsum.einsum_impl_ext.numpy_extended_dot>` in pure python. This implementation is not efficient but shows how to implement this operation without :func:`numpy.einsum`. .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot_python from onnx_extended.tools.einsum.einsum_impl_ext import ( _numpy_extended_dot_equation) a = numpy.arange(6).reshape((3, 2, 1)) b = numpy.arange(12).reshape((3, 1, 4)) print(numpy_extended_dot_python( a, b, axes=(0, ), left=(1,), right=(2,))) # Equivalent einsum equation print('equation', _numpy_extended_dot_equation( len(a.shape), len(a.shape), axes=(0, ), left=(1,), right=(2,))) # Same einsum computation written in a different way. print(numpy.einsum('kix,kxj->xij', a, b)) """ def dispb(c): return "".join("o" if b else "." for b in c) new_shape = numpy_extended_dot_ouput_shape(m1, m2, axes, left, right) m1_dim = len(m1.shape) # output result res = numpy.full(tuple(new_shape), 0, dtype=m1.dtype) # indices l1, l2, l3 = _numpy_extended_dot_python_l1l2l3(m1_dim, axes, left, right) names, kind, cols, common, broadcast, pos = _numpy_extended_dot_python_intermediate( m1.shape, m2.shape, l1, l2, l3 ) if any(broadcast): l1, l2, l3 = _numpy_extended_dot_python_update_broadcast( m1, m2, axes, left, right, l1, l2, l3, names, broadcast, cols, kind, common, verbose=verbose, ) ( names, kind, cols, common, broadcast, pos, ) = _numpy_extended_dot_python_intermediate(m1.shape, m2.shape, l1, l2, l3) indices = numpy.array([0 for n in names], dtype=numpy.int64) pl1 = numpy.array([names.index(c) for c in l1], dtype=numpy.int64) pl2 = numpy.array([names.index(c) for c in l2], dtype=numpy.int64) limits = numpy.array( [ m1.shape[pos[n]] if (kind[n] & 1) == 1 else m2.shape[pos[n]] for n in range(len(names)) ], dtype=numpy.int64, ) plo = numpy.array( [-1 if c not in names else names.index(c) for c in l3], dtype=numpy.int64 ) if verbose: print( "[GENERICDOT] %s,%s->%s or %s" % ( "".join(l1), "".join(l2), "".join([(s if s else "?") for s in l3]), _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ), ) ) print( "[GENERICDOT] shape1=%r shape2=%r shape=%r" % (m1.shape, m2.shape, res.shape) ) print(f"[GENERICDOT] axes={axes!r} left={left!r} right={right!r}") print(f"[GENERICDOT] pl1={pl1!r} pl2={pl2!r} plo={plo!r}") print( "[GENERICDOT] names=%s kind=%r common=%s broadcast=%s" % ("".join(names), kind.tolist(), dispb(common), dispb(broadcast)) ) print(f"[GENERICDOT] pos={pos.tolist()!r}") print(f"[GENERICDOT] cols={cols!r}") print(f"[GENERICDOT] limits={limits!r}") while indices[0] < limits[0]: # The function spends most of its time is these three lines. t1 = tuple(indices[n] for n in pl1) t2 = tuple(indices[n] for n in pl2) to = tuple(0 if n == -1 else indices[n] for n in plo) c = m1[t1] * m2[t2] if verbose: print(f" {t1!r} x {t2!r} -> {to!r} v={c!r} I={indices!r}") res[to] += c last = len(indices) - 1 indices[last] += 1 for i in range(last, 0, -1): if indices[i] < limits[i]: break indices[i] = 0 if i > 0: indices[i - 1] += 1 return res
[docs] def numpy_extended_dot_matrix( m1: numpy.ndarray, m2: numpy.ndarray, axes: Tuple[int, ...], left: Tuple[int, ...], right: Tuple[int, ...], verbose: bool = False, ) -> numpy.ndarray: """ Implementation of :func:`numpy_extended_dot <onnx_extended.tools.einsum.einsum_impl_ext.numpy_extended_dot>` using dot product, multiplication, transpose and reduction but not a custom python implementation like :func:`numpy_extended_dot_python <onnx_extended.tools.einsum.einsum_impl_ext.numpy_extended_dot_python>`. .. runpython:: :showcode: import numpy from onnx_extended.tools.einsum import numpy_extended_dot_matrix from onnx_extended.tools.einsum.einsum_impl_ext import ( _numpy_extended_dot_equation) a = numpy.arange(6).reshape((3, 2, 1)) b = numpy.arange(12).reshape((3, 1, 4)) print(numpy_extended_dot_matrix( a, b, axes=(0, ), left=(1,), right=(2,))) # Equivalent einsum equation print('equation', _numpy_extended_dot_equation( len(a.shape), len(a.shape), axes=(0, ), left=(1,), right=(2,))) # Same einsum computation written in a different way. print(numpy.einsum('kix,kxj->xij', a, b)) """ _common_check_numpy_extended_dot(m1, m2, axes, left, right) if verbose: print( "[GENERICDOT] shape1=%r shape2=%r axes=%r " "left=%r right=%r -- %s" % ( m1.shape, m2.shape, axes, left, right, _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ), ) ) if len(axes) == 0 and not (set(left) & set(right)): # Simple multiplication res = m1 * m2 if verbose: print(f"[GENERICDOT] Mul {m1.shape!r} @ {m2.shape!r} -> {res.shape!r}") return res if len(set(axes) & set(left)) == 0 and len(set(axes) & set(right)) == 0: # No intersection between axes and right: matrix multiplication # ReduceSum right_no_left = set(right) - (set(right) & (set(left) | set(axes))) if right_no_left: red1 = m1.sum(axis=tuple(sorted(right_no_left)), keepdims=True) if verbose: print( "[GENERICDOT] reducesumL=%r, %r -> %r" % (right_no_left, m1.shape, red1.shape) ) else: red1 = m1 left_no_right = set(left) - (set(left) & (set(right) | set(axes))) if left_no_right: red2 = m2.sum(axis=tuple(sorted(left_no_right)), keepdims=True) if verbose: print( "[GENERICDOT] reducesumR=%r, %r -> %r" % (left_no_right, m2.shape, red2.shape) ) else: red2 = m2 # Transpose common_axes = sorted(set(left) & set(right)) i_axes = [ (-1 if i in common_axes else (1 if i in axes else 0), i) for i in range(len(m1.shape)) ] i_axes.sort() perm = [_[1] for _ in i_axes] trm1 = numpy.transpose(red1, axes=perm) trm2 = numpy.transpose(red2, axes=perm) if verbose: print(f"[GENERICDOT] transposeL={perm!r}, {red1.shape!r} -> {trm1.shape!r}") print(f"[GENERICDOT] transposeR={perm!r}, {red2.shape!r} -> {trm2.shape!r}") final_shape = numpy_extended_dot_ouput_shape(m1, m2, axes, left, right) perm_left = [i for i in range(len(perm)) if perm[i] in left] perm_right = [i for i in range(len(perm)) if perm[i] in right] perm_common_axes = [i for i in range(len(perm)) if perm[i] in common_axes] if verbose: print( "[GENERICDOT] MatMul %r @ %r -> %r -- %s" % ( m1.shape, m2.shape, final_shape, _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ), ) ) print(f"[GENERICDOT] axes={axes!r} left={left!r} right={right!r}") print( "[GENERICDOT] perm=%r perm_left=%r " "perm_right=%r perm_common_axes=%r" % (perm, perm_left, perm_right, perm_common_axes) ) # Reshape dim0 = int(numpy.prod([trm1.shape[i] for i in perm_common_axes])) dim0b = int(numpy.prod([trm2.shape[i] for i in perm_common_axes])) if len(axes) > 0: all_axes = list(range(0, len(m1.shape))) new_axes = all_axes[-len(axes) :] else: new_axes = [] dim1 = int(numpy.prod([trm1.shape[i] for i in new_axes])) dim2 = int(numpy.prod([trm2.shape[i] for i in new_axes])) if dim1 != dim2: raise RuntimeError( "Summation axis do not have the same length %d != %d, " "trshape1=%r trshape2=%r " "p_axes=%r p_left=%r p_right=%r p_common=%r" "." % ( dim1, dim2, trm1.shape, trm2.shape, new_axes, perm_left, perm_right, perm_common_axes, ) ) else: shm1 = trm1.reshape((dim0, -1, dim1)) shm2 = trm2.reshape((dim0b, -1, dim2)) if verbose: print( "[GENERICDOT] Reshape %r @ %r -> %r @ %r" % ((dim0, -1, dim1), (dim0, -1, dim2), shm1.shape, shm2.shape) ) print("[GENERICDOT] matmul") # Multiplication (this should be done in a different way. res = shm1 @ numpy.transpose(shm2, axes=(0, 2, 1)) if verbose: print(f"[GENERICDOT] Shape after multiplication {res.shape}") # Transpose again not_in_both = [] for i in range(0, len(m1.shape)): if i not in left and i not in right: not_in_both.append(i) ordered_axes = ( common_axes + list(i for i in left if i not in right) + list(i for i in right if i not in left) + not_in_both ) perm_not_in_both = [i for i in range(len(perm)) if perm[i] in not_in_both] current_shape = ( [max(trm1.shape[i], trm2.shape[i]) for i in sorted(perm_common_axes)] + [trm1.shape[i] for i in sorted(perm_left) if i not in perm_common_axes] + [trm2.shape[i] for i in sorted(perm_right) if i not in perm_common_axes] + [1 for i in perm_not_in_both] ) if verbose: print( "[GENERICDOT] current_shape=%r final_shape=%r " "last_shape=%r" % (current_shape, final_shape, res.shape) ) assert len(current_shape) == len(final_shape), ( "Shapes mismatch %r > %r, " "shape1=%r shape2=%r axes=%r left=%r right=%r." % (current_shape, final_shape, m1.shape, m2.shape, axes, left, right) ) res = res.reshape(current_shape) perm = [(a, i) for i, a in enumerate(ordered_axes)] perm.sort() perm = [p[1] for p in perm] if verbose: print(f"[GENERICDOT] ordered_axes={ordered_axes!r} perm={perm!r}") return numpy.transpose(res, axes=perm) else: # Multiplication and Matrix multiplication at the same time. l_axes = set(left) & set(axes) r_axes = set(right) & set(axes) if r_axes and not l_axes: new_axes = list(a for a in axes if a not in right) new_left = list(sorted(set(left) | r_axes)) if verbose: eq1 = _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ) eq2 = _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), tuple(new_axes), tuple(new_left), tuple(right), ) print( "[GENERICDOT] replace left %r by %r axes %r by %r, " "eq %r by %r" % (left, new_left, axes, new_axes, eq1, eq2) ) return numpy_extended_dot_matrix( m1, m2, tuple(new_axes), tuple(new_left), tuple(right), verbose=verbose ) raise RuntimeError( "shape1=%r shape2=%r axes=%r left=%r right=%r eq=%s." % ( m1.shape, m2.shape, axes, left, right, _numpy_extended_dot_equation( len(m1.shape), len(m1.shape), axes, left, right ), ) )