tensorflow/models · error · ValueError

Skipping eval during pretraining without supervised head.

Error message

Skipping eval during pretraining without supervised head.

What it means

Error "Skipping eval during pretraining without supervised head." thrown in tensorflow/models.

Source

Thrown at official/projects/simclr/tasks/simclr.py:339

    for var in tvars:
      logging.info(var.name)
    grads = tape.gradient(scaled_loss, tvars)
    # Scales back gradient when LossScaleOptimizer is used.
    if isinstance(optimizer, tf_keras.mixed_precision.LossScaleOptimizer):
      grads = optimizer.get_unscaled_gradients(grads)
    optimizer.apply_gradients(list(zip(grads, tvars)))

    logs = {self.loss: losses['total_loss']}

    for m in metrics:  # pyrefly: ignore[not-iterable]
      m.update_state(losses[m.name])
      logs.update({m.name: m.result()})

    return logs

  def validation_step(self, inputs, model, metrics=None):
    if self.task_config.model.supervised_head is None:
      raise ValueError(
          'Skipping eval during pretraining without supervised head.')

    features, labels = inputs
    if self.task_config.evaluation.one_hot:
      num_classes = self.task_config.model.supervised_head.num_classes
      labels = tf.one_hot(labels, num_classes)

    outputs = model(
        features, training=False)[simclr_model.SUPERVISED_OUTPUT_KEY]
    outputs = tf.nest.map_structure(lambda x: tf.cast(x, tf.float32), outputs)

    logs = {self.loss: 0}

    if metrics:
      self.process_metrics(metrics, labels, outputs)
      logs.update({m.name: m.result() for m in metrics})
    elif model.compiled_metrics:
      self.process_compiled_metrics(model.compiled_metrics, labels, outputs)

View on GitHub (pinned to e006f5f0d5)

When it happens

Trigger: Thrown at official/projects/simclr/tasks/simclr.py:339 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/0e3d52661c7048e9. Report an issue: GitHub.