tensorflow/models · error · ValueError

Number of embedding features (num_pos_features) must be even

Error message

Number of embedding features (num_pos_features) must be even when column and row embeddings are concatenated.

What it means

Error "Number of embedding features (num_pos_features) must be even when column and row embeddings are concatenated." thrown in tensorflow/models.

Source

Thrown at official/projects/mae/modeling/utils.py:49

    attention_mask: a `bool` Tensor specifying the size of the input image to
      the Transformer and which elements are padded, of size [batch_size,
      height, width]
    num_pos_features: a `int` specifying the number of positional features,
      should be equal to the hidden size of the Transformer network
    temperature: a `float` specifying the temperature of the positional
      embedding. Any type that is converted to a `float` can also be accepted.
    normalize: a `bool` determining whether the positional embeddings should be
      normalized between [0, scale] before application of the sine and cos
      functions.
    scale: a `float` if normalize is True specifying the scale embeddings before
      application of the embedding function.

  Returns:
    embeddings: a `float` tensor of the same shape as input_tensor specifying
      the positional embeddings based on sine features.
  """
  if num_pos_features % 2 != 0:
    raise ValueError(
        "Number of embedding features (num_pos_features) must be even when "
        "column and row embeddings are concatenated.")
  num_pos_features = num_pos_features // 2

  # Produce row and column embeddings based on total size of the image
  # <tf.float>[batch_size, height, width]
  attention_mask = tf.cast(attention_mask, tf.float32)
  row_embedding = tf.cumsum(attention_mask, 1)
  col_embedding = tf.cumsum(attention_mask, 2)

  if normalize:
    eps = 1e-6
    row_embedding = row_embedding / (row_embedding[:, -1:, :] + eps) * scale
    col_embedding = col_embedding / (col_embedding[:, :, -1:] + eps) * scale

  dim_t = tf.range(num_pos_features, dtype=row_embedding.dtype)
  dim_t = tf.pow(temperature, 2 * (dim_t // 2) / num_pos_features)

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/projects/mae/modeling/utils.py:49 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/91adc7ee40f3a828. Report an issue: GitHub.