tensorflow/models · error · ValueError

User should set optimizer attribute to model inside…

Error message

User should set optimizer attribute to model inside `model_fn`.

What it means

Guard inside DistributedExecutor.train after model_fn runs in the strategy scope: the produced model has no optimizer attribute, so no gradients can be applied. The model_fn contract requires attaching the optimizer to the returned model (created inside the strategy scope) before training.

Solutions

1. Inside model_fn, set model.optimizer = tf_keras.optimizers.SGD(learning_rate=..., ...) within strategy.scope()
2. Ensure model_fn returns the model object with the optimizer attached, not a bare network

When it happens

Trigger: Thrown at official/legacy/detection/executor/distributed_executor.py:421 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/364e3ccde99ddf26. Report an issue: GitHub.

Appendix: source

Thrown at official/legacy/detection/executor/distributed_executor.py:421

      save_freq = FLAGS.save_checkpoint_freq
    else:
      save_freq = iterations_per_loop

    params = self._params
    strategy = self._strategy
    # To reduce unnecessary send/receive input pipeline operation, we place
    # input pipeline ops in worker task.
    train_iterator = self._get_input_iterator(train_input_fn, strategy)
    train_loss = None
    train_metric_result = None
    eval_metric_result = None
    tf_keras.backend.set_learning_phase(1)
    with strategy.scope():
      # To correctly place the model weights on accelerators,
      # model and optimizer should be created in scope.
      model = self.model_fn(params.as_dict())
      if not hasattr(model, 'optimizer'):
        raise ValueError('User should set optimizer attribute to model '
                         'inside `model_fn`.')
      optimizer = model.optimizer

      # Training loop starts here.
      checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
      latest_checkpoint_file = tf.train.latest_checkpoint(model_dir)
      initial_step = 0
      if latest_checkpoint_file:
        logging.info(
            'Checkpoint file %s found and restoring from '
            'checkpoint', latest_checkpoint_file)
        checkpoint.restore(latest_checkpoint_file)
        initial_step = optimizer.iterations.numpy()
        logging.info('Loading from checkpoint file completed. Init step %d',
                     initial_step)
      elif init_checkpoint:
        logging.info('Restoring from init checkpoint function')
        init_checkpoint(model)

View on GitHub (pinned to e006f5f0d5)