tensorflow/models · error · ValueError

Unknown `output` value "%s". `output` can be either "logits"

Error message

Unknown `output` value "%s". `output` can be either "logits" or "predictions"

What it means

Error "Unknown `output` value "%s". `output` can be either "logits" or "predictions"" thrown in tensorflow/models.

Source

Thrown at official/nlp/modeling/networks/classification.py:73

    logits = tf_keras.layers.Dense(
        num_classes,
        activation=None,
        kernel_initializer=initializer,
        name='predictions/transform/logits')(
            cls_output)

    if output == 'logits':
      output_tensors = logits
    elif output == 'predictions':
      policy = tf_keras.mixed_precision.global_policy()
      if policy.name == 'mixed_bfloat16':
        # b/158514794: bf16 is not stable with post-softmax cross-entropy.
        policy = tf.float32
      output_tensors = tf_keras.layers.Activation(
          tf.nn.log_softmax, dtype=policy)(
              logits)
    else:
      raise ValueError(
          ('Unknown `output` value "%s". `output` can be either "logits" or '
           '"predictions"') % output)

    super().__init__(
        inputs=[cls_output], outputs=output_tensors, **kwargs)

    # b/164516224
    # Once we've created the network using the Functional API, we call
    # super().__init__ as though we were invoking the Functional API Model
    # constructor, resulting in this object having all the properties of a model
    # created using the Functional API. Once super().__init__ is called, we
    # can assign attributes to `self` - note that all `self` assignments are
    # below this line.
    config_dict = {
        'input_width': input_width,
        'num_classes': num_classes,
        'initializer': initializer,
        'output': output,

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/nlp/modeling/networks/classification.py:73 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/404b4e474d522845. Report an issue: GitHub.