keras-team/keras · error · ValueError

Invalid einsum equation '{equation}'. Equations must be in t

Error message

Invalid einsum equation '{equation}'. Equations must be in the form [X],[Y]->[Z], ...[X],[Y]->...[Z], or [X]...,[Y]->[Z]....

What it means

EinsumDense parses its equation with a regex covering 'ab,bc->ac', ellipsis forms like '...a,ab->...b', and split forms like 'aab,bc->acd' (where repeated left-side letters define extra output dims). Any equation matching none of these patterns raises this ValueError at build or compute_output_shape time.

Source

Thrown at keras/src/layers/core/einsum_dense.py:1759

    # This is the case where ellipses are present on the left.
    split_string = re.match(
        "0([a-zA-Z]+),([a-zA-Z]+)->0([a-zA-Z]+)", dot_replaced_string
    )
    if split_string:
        return _analyze_split_string(
            split_string, bias_axes, input_shape, output_shape, left_elided=True
        )

    # This is the case where ellipses are present on the right.
    split_string = re.match(
        "([a-zA-Z]{2,})0,([a-zA-Z]+)->([a-zA-Z]+)0", dot_replaced_string
    )
    if split_string:
        return _analyze_split_string(
            split_string, bias_axes, input_shape, output_shape
        )

    raise ValueError(
        f"Invalid einsum equation '{equation}'. Equations must be in the form "
        "[X],[Y]->[Z], ...[X],[Y]->...[Z], or [X]...,[Y]->[Z]...."
    )


def _analyze_split_string(
    split_string, bias_axes, input_shape, output_shape, left_elided=False
):
    """Computes kernel and bias shapes from a parsed einsum equation.

    This function takes the components of an einsum equation, validates them,
    and calculates the required shapes for the kernel and bias weights.

    Args:
        split_string: A regex match object containing the input, weight, and
            output specifications.
        bias_axes: A string indicating which output axes to apply a bias to.
        input_shape: The shape of the input tensor.

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Restrict the equation to exactly two operands joined by ',' with a '->' output, e.g. 'ab,bc->ac'
2. Place ellipsis only at the start of operands as in '...a,ab->...b'
3. For split equations, ensure the output is a strict superset of the shared letters of the two inputs
4. Print-check the equation string for stray spaces or wrong arrow characters before constructing the layer

Example fix

# before
layer = keras.layers.EinsumDense('a,b,c->d', ...)
# after
layer = keras.layers.EinsumDense('ab,bc->ac', ...)

# ellipsis form
layer = keras.layers.EinsumDense('...a,ab->...b', ...)

import re
EINSUM_RE = re.compile(r'^[a-z.]*,[a-z.]*->[a-z.]*$')
def valid_two_operand(eq):
    eq = eq.replace(' ', '')
    if not EINSUM_RE.match(eq):
        return False
    left, _ = eq.split('->')
    return len(left.split(',')) == 2

def is_valid_einsum_equation(eq):
    eq = eq.replace(' ', '')
    return '->' in eq and eq.count(',') == 1 and valid_two_operand(eq)

try:
    layer = keras.layers.EinsumDense(eq, output_shape=shape)
    layer.compute_output_shape(input_shape)
except ValueError as e:
    if 'Invalid einsum equation' in str(e):
        raise ValueError('Bad equation: ' + str(e)) from e
    raise

When it happens

Trigger: Passing an equation with three operands ('a,b,c->d'), malformed separators (missing '->', spaces inside subscripts), a wrong arrow, or an unparseable split equation to keras.layers.EinsumDense, then calling build() or compute_output_shape().

Common situations: Copy-pasting numpy.einsum equations with three operands; typos like 'ab,bc=>ac' or 'ab bc->ac'; using an ellipsis in an unsupported position.

Related errors

• Dimension '{dim}' was specified in the output '{output_spec}
• Weight dimension '{dim}' did not have a match in either the
• Bias dimension '{char}' was requested, but is not part of th
• If using `weights="imagenet"` as true, `classes` should be 1
• The number of repeats in `EfficientNet` must be > 0. Receive

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/f00cd16a74f3f0a3. Report an issue: GitHub.