ab6f9e4c489beea3bec518d60629c0f8af0b6123,opennmt/runner.py,Runner,train,#Runner#Any#,263

Before Change

    report_every = train_config.get("save_summary_steps", 100)
    accum_count = params.get("gradients_accum", 1)

    for i, (source, target) in enumerate(dataset):
      loss, num_words = _step(source, target)
      if i == 0 or (i + 1) % accum_count == 0:
        _apply_gradients()

      for key, value in six.iteritems(num_words):
        value = value.numpy()
        if key not in accum_num_words:
          accum_num_words[key] = value
        else:
          accum_num_words[key] += value

      step = optimizer.iterations.numpy()
      if step % report_every == 0:
        last_report_time = _report_training_status(
            step,
            loss,
            optimizer.learning_rate,
            accum_num_words,
            last_report_time)
      if step % save_checkpoints_steps == 0 or step == train_steps:
        path = checkpoint_manager.save(checkpoint_number=step)
        tf.get_logger().info("Saved checkpoint %s", path)
      if step == train_steps:
        break

    return self._maybe_average_checkpoints()

  def evaluate(self, checkpoint_path=None):
    Runs evaluation.

After Change

      if step == train_steps:
        break

    _save_checkpoint(step)
    return self._maybe_average_checkpoints()

  def evaluate(self, checkpoint_path=None):
    Runs evaluation.

Instances