keras-team/keras · error · NotImplementedError

Unknown `interpolation` {interpolation}. Expected of one {se

Error message

Unknown `interpolation` {interpolation}. Expected of one {self._SUPPORTED_INTERPOLATION}.

What it means

Raised by RandomTranslation's constructor when `interpolation` is not one of "nearest" or "bilinear". Only these two resampling methods are implemented for translating images.

Source

Thrown at keras/src/layers/preprocessing/image_preprocessing/random_translation.py:134

        **kwargs,
    ):
        super().__init__(data_format=data_format, **kwargs)
        self.height_factor = height_factor
        self.height_lower, self.height_upper = self._set_factor(
            height_factor, "height_factor"
        )
        self.width_factor = width_factor
        self.width_lower, self.width_upper = self._set_factor(
            width_factor, "width_factor"
        )

        if fill_mode not in self._SUPPORTED_FILL_MODE:
            raise NotImplementedError(
                f"Unknown `fill_mode` {fill_mode}. Expected of one "
                f"{self._SUPPORTED_FILL_MODE}."
            )
        if interpolation not in self._SUPPORTED_INTERPOLATION:
            raise NotImplementedError(
                f"Unknown `interpolation` {interpolation}. Expected of one "
                f"{self._SUPPORTED_INTERPOLATION}."
            )

        self.fill_mode = fill_mode
        self.fill_value = fill_value
        self.interpolation = interpolation
        self.seed = seed
        self.generator = SeedGenerator(seed)
        self.supports_jit = False

    def _set_factor(self, factor, factor_name):
        if isinstance(factor, (tuple, list)):
            if len(factor) != 2:
                raise ValueError(
                    self._FACTOR_VALIDATION_ERROR
                    + f"Received: {factor_name}={factor}"
                )

View on GitHub (pinned to 7a34a03db6)

Solutions

Use "bilinear" for smooth results or "nearest" for speed
If you need bicubic-quality augmentation, resize before/after the layer or use RandomRotation/RandomZoom if they support it in your Keras version

Example fix

// before
layer = RandomTranslation(0.1, 0.1, interpolation="bicubic")
// after
layer = RandomTranslation(0.1, 0.1, interpolation="bilinear")

Defensive patterns

Strategy: validation

Validate before calling

assert interpolation in {"nearest", "bilinear"}, 'RandomTranslation supports only nearest/bilinear'

Type guard

def is_valid_interp(v):
    return v in {"nearest", "bilinear"}

Try / catch

try:
    layer = RandomTranslation(0.1, 0.1, interpolation=interp)
except NotImplementedError:
    interp = "bilinear"
    layer = RandomTranslation(0.1, 0.1, interpolation=interp)

Prevention

Don't assume tf.image.resize interpolation names transfer to preprocessing layers
Default to "bilinear" unless speed requires "nearest"

When it happens

Trigger: Calling RandomTranslation(..., interpolation="bicubic") or interpolation="lanczos".

Common situations: Reusing interpolation strings from tf.image.resize (which supports 'bicubic', 'lanczos3', etc.) or from other Keras preprocessing layers like RandomRotation which accept "bicubic" in newer versions.

Related errors

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/a933a60fd0b86f0f. Report an issue: GitHub.