keras-team/keras · error · ValueError

When setting values directly, `mean` and `variance` must hav

Error message

When setting values directly, `mean` and `variance` must have the same shape. Received: mean shape {np.shape(mean)} and variance shape {np.shape(variance)}

What it means

Normalization broadcasts mean and variance across the non-normalized axes, so the two arrays must have identical shapes. __init__ compares np.shape(mean) vs np.shape(variance) and raises on mismatch.

Source

Thrown at keras/src/layers/preprocessing/normalization.py:140

        self.axis = axis

        self.input_mean = mean
        self.input_variance = variance
        self.invert = invert
        self.supports_masking = True
        self._build_input_shape = None
        self.mean = None

        # Set `mean` and `variance` if passed.
        if (mean is not None) != (variance is not None):
            raise ValueError(
                "When setting values directly, both `mean` and `variance` "
                f"must be set. Received: mean={mean} and variance={variance}"
            )
        if mean is not None:
            # Verify mean and variance have the same shape.
            if np.shape(mean) != np.shape(variance):
                raise ValueError(
                    "When setting values directly, `mean` and `variance` "
                    "must have the same shape. Received: "
                    f"mean shape {np.shape(mean)} and "
                    f"variance shape {np.shape(variance)}"
                )
            # Verify mean rank <= number of axes.
            if len(np.shape(mean)) > len(self.axis):
                raise ValueError(
                    "The rank of `mean` must be less than or equal to the "
                    f"number of axes ({len(self.axis)}). Received: "
                    f"mean shape {np.shape(mean)} for axis {self.axis}"
                )

    def build(self, input_shape):
        if input_shape is None:
            return

        ndim = len(input_shape)

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Reshape both to the same shape, typically via .ravel() or reshape to the kept-axis dims
2. Recompute both statistics from the same dataset pass
3. Verify: assert np.shape(mean) == np.shape(variance)

Example fix

// before
layer = Normalization(axis=-1, mean=mu, variance=var)  # (768,) vs (1,768)
// after
import numpy as np
layer = Normalization(axis=-1, mean=np.ravel(mu), variance=np.ravel(var))

import numpy as np
if np.shape(mean) != np.shape(variance):
    mean, variance = np.ravel(mean), np.ravel(variance)
assert np.shape(mean) == np.shape(variance)

When it happens

Trigger: Normalizing axis=-1 with mean of shape (768,) but variance of shape (1,768); statistics exported from different sources or with squeeze/reshape applied inconsistently.

Common situations: Loading channel statistics from separate files (mean.npy, std.npy then squaring); mixing per-feature stats computed at different data versions.

Related errors

• The layer was built with input_shape={self._build_input_shap
• The `weights` argument should be either `None` (random initi
• Expected mode to be one of `caffe`, `tf` or `torch`. Receive
• TF-IDF data must be a 1-index array. Received: type(idf_weig
• When using `output_mode={self.output_mode}` and `pad_to_max_

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/9a3575b06f40d0d7. Report an issue: GitHub.