keras-team/keras · error · ValueError

Tensor {tensor} from operation '{operation.name}' is part of

Error message

Tensor {tensor} from operation '{operation.name}' is part of a cycle.

What it means

While recursively mapping the symbolic graph, _build_map_helper tracks nodes currently being traversed. Encountering a node that is already in progress means the tensor depends on itself — a cycle — which a feed-forward function graph cannot represent, so construction aborts.

Source

Thrown at keras/src/ops/function.py:462

        return

    node = operation._inbound_nodes[node_index]

    # Don't repeat work for shared subgraphs
    if node in finished_nodes:
        return

    # If this tensor is one of the declared inputs and its producing
    # operation is not an InputLayer, stop traversal here. The operation
    # that produced this tensor is outside the Function's graph.
    flat_inputs = tree.flatten(inputs)
    if not node.is_input and tensor in flat_inputs:
        finished_nodes.add(node)
        return

    # Prevent cycles.
    if node in nodes_in_progress:
        raise ValueError(
            f"Tensor {tensor} from operation '{operation.name}' is part of a "
            "cycle."
        )

    # Store the traversal order for operation sorting.
    if operation not in operation_indices:
        operation_indices[operation] = len(operation_indices)

    # Propagate to all previous tensors connected to this node.
    nodes_in_progress.add(node)
    if not node.is_input:
        for input_tensor in node.input_tensors:
            _build_map_helper(
                inputs,
                input_tensor,
                finished_nodes,
                nodes_in_progress,
                nodes_in_decreasing_depth,

View on GitHub (pinned to 7a34a03db6)

Solutions

Inspect the operation named in the message and break the feedback loop: its input must come from the function inputs or earlier nodes only
For recurrent behavior, use keras.layers.RNN with a cell, or a keras.ops.custom_op/python function with scan-style loops instead of the symbolic graph
Check for accidental self-assignment/shadowing of the tensor variable during graph construction

Example fix

# before
out = block(x)
out = layers.Add()([out, out_prev])  # out_prev derived from out -> cycle

# after
out = block(x)
out = layers.Add()([out, skip_from_input])  # skip comes from inputs/earlier node

Defensive patterns

Strategy: validation

Prevention

Never wire a block's output back into its own input chain in the functional API
Use keras.layers.RNN for recurrence instead of manual feedback
Review variable assignments in long build functions for accidental self-dependency

When it happens

Trigger: Wiring a layer's output back into its own input chain, e.g. y = layer(layer(x) ... y), or building a recurrent/feedback structure with plain functional ops instead of a proper RNN cell; aliasing bugs where a variable is overwritten and feeds itself.

Common situations: Attempting custom recurrence with the functional API; variable shadowing in long build functions (out = f(out) before out is defined from inputs); copy-paste wiring that connects a block's output to its own input.

Related errors

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/458633a382eca5d9. Report an issue: GitHub.