keras-team/keras · error · ValueError

Variable is already initialized.

Error message

Variable {self.path} is already initialized.

What it means

Raised by Variable._deferred_initialize when deferred (lazy) initialization is triggered but the variable already holds a concrete value. This typically happens when a variable created under a stateless scope (e.g. inside a functional call or jax.jit trace) was already initialized by a concrete forward pass, and the deferred init path then tries to initialize it again. The NNX-enabled code path above returns early in exactly this situation; with NNX disabled, Keras treats re-initialization as a bug and raises.

Solutions

1. Avoid triggering variable creation/initialization paths twice on the same variable; create variables once (typically in build()) and reuse them.
2. If the variable was already initialized by a concrete call, skip manual re-initialization instead of forcing _deferred_initialize again.
3. On the JAX backend, enabling NNX (keras.config.set_nnx_enabled(True)) makes this case return early instead of raising.
4. Do not call private initialization APIs (_initialize, _deferred_initialize) on variables managed by Keras.

When it happens

Trigger: Thrown at keras/src/backend/common/variables.py:223 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/707fb571bcc5fe32. Report an issue: GitHub.

Appendix: source

Thrown at keras/src/backend/common/variables.py:223

                    "(with a `shape` and `initializer` argument)."
                )
        else:
            if callable(initializer):
                self._shape = self._validate_shape(shape)
                self._initialize_with_initializer(initializer)
            else:
                self._initialize(initializer)
                self._shape = self._validate_shape(self._value.shape)
        self._ndim = len(self._shape)

    def _deferred_initialize(self):
        if self._value is not None:
            # If NNX is enabled, it's possible the variable was already
            # initialized by a concrete call. In this case, _deferred_initialize
            # returns early and does not raise an error.
            if config.is_nnx_enabled():
                return
            raise ValueError(f"Variable {self.path} is already initialized.")

        if in_stateless_scope():
            raise ValueError(
                "You are attempting to initialize a variable "
                "while in a stateless scope. This is disallowed. "
                "Make sure that all variables are initialized "
                "before you start using your layer/model objects."
            )
        self._initialize_with_initializer(self._initializer)
        self._initializer = None

    def _validate_shape(self, shape):
        shape = standardize_shape(shape)
        if None in shape:
            raise ValueError(
                "Shapes used to initialize variables must be "
                "fully-defined (no `None` dimensions). Received: "
                f"shape={shape} for variable path='{self.path}'"

View on GitHub (pinned to 7a34a03db6)