tensorflow/models · error · ValueError

Unsupported feature_transform. The supported feature_transfo

Error message

Unsupported feature_transform. The supported feature_transform are %s. Got '%s'.

What it means

Error "Unsupported feature_transform. The supported feature_transform are %s. Got '%s'." thrown in tensorflow/models.

Source

Thrown at official/nlp/modeling/layers/kernel_attention.py:601

        the dot product based on key length. Set as log_512^(n) to stablize
        attention entropy against length. Refer to
        https://kexue.fm/archives/8823 for details.
      use_causal_windowed: If true perform windowed causal attention. See
        causal_windowed_performer_attention function docstring for more details.
      causal_chunk_length: Length of each chunk in tokens.
      causal_window_length: Length of attention window in chunks.
      causal_window_decay: Float window decay factor or `None`. If set,
        exponentially decay past attention window values by this factor before
        summation.
      causal_padding: Pad the query, value and key input tensors across the axis
        from either left or right if padding is set to "left" or "right"; apply
        no padding if padding is set to None. In the latter case, the axis
        dimension of the query, value and key input tensors must be divisible by
        the chunk_length.
      **kwargs: The same arguments `MultiHeadAttention` layer.
    """
    if feature_transform not in _TRANSFORM_MAP:
      raise ValueError("Unsupported feature_transform. The supported "
                       "feature_transform are %s. "
                       "Got '%s'." % (_TRANSFORM_MAP.keys(), feature_transform))
    if num_random_features <= 0 and redraw:
      raise ValueError(
          "There is nothing to redraw when num_random_features <= 0.")
    self._feature_transform = feature_transform
    self._num_random_features = num_random_features
    self._redraw = redraw
    self._is_short_seq = is_short_seq
    self._begin_kernel = begin_kernel
    self._scale_by_length = scale_by_length
    # We use the seed for two scenarios:
    # 1. inference
    # 2. no redraw
    self._seed = seed
    super().__init__(**kwargs)
    if scale is None:
      self._scale = 1.0 / math.sqrt(float(self._key_dim))

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/nlp/modeling/layers/kernel_attention.py:601 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of tensorflow/models@e006f5f0d5 (2026-08-24). Data as JSON: /api/errors/df9a94fa4057e457. Report an issue: GitHub.