def _prepare_feed_values(model, inputs, targets, sample_weights, mode):
  """Prepare feed values to the model execution function.

  Arguments:
    model: Model to prepare feed values for.
    inputs: List or dict of model inputs.
    targets: Optional list of model targets.
    sample_weights: Optional list of sample weight arrays.
    mode: One of 'train'/'test'/'predict'.

  Returns:
    Feed values for the model in the given mode.
  """
  strategy = model._distribution_strategy
  inputs, targets, sample_weights = _get_input_from_iterator(inputs, model)
  inputs = distributed_training_utils.flatten_perdevice_values(strategy, inputs)
  targets = distributed_training_utils.flatten_perdevice_values(
      strategy, targets)
  if mode == 'predict':
    sample_weights = []
    targets = []
  else:
    sample_weights = [
        None for _ in range(len(model.outputs) * strategy.num_replicas_in_sync)
    ]
  ins = inputs + targets + sample_weights
  if mode == 'train' and not isinstance(K.learning_phase(), int):
    ins += [True]
  return ins

  def on_epoch_end(self, epoch, logs=None):
    """Checks if summary ops should run next epoch, logs scalar summaries."""

    # don't output batch_size and
    # batch number as Tensorboard summaries
    logs = {('epoch_' + k): v
            for k, v in logs.items()
            if k not in ['batch', 'size']}
    self._write_custom_summaries(epoch, logs)

    # pop the histogram summary op after each epoch
    if self.histogram_freq:
      if self.merged in self.model.test_function.fetches:
        self.model.test_function.fetches.remove(self.merged)
      if self.merged in self.model.test_function.fetch_callbacks:
        self.model.test_function.fetch_callbacks.pop(self.merged)

    if self.embeddings_data is None and self.embeddings_freq:
      raise ValueError('To visualize embeddings, embeddings_data must '
                       'be provided.')

    if self.embeddings_freq and self.embeddings_data is not None:
      if epoch % self.embeddings_freq == 0:
        # We need a second forward-pass here because we're passing
        # the `embeddings_data` explicitly. This design allows to pass
        # arbitrary data as `embeddings_data` and results from the fact
        # that we need to know the size of the `tf.Variable`s which
        # hold the embeddings in `set_model`. At this point, however,
        # the `validation_data` is not yet set.

        embeddings_data = self.embeddings_data
        n_samples = embeddings_data[0].shape[0]
        i = 0
        while i < n_samples:
          step = min(self.batch_size, n_samples - i)
          batch = slice(i, i + step)

          if isinstance(self.model.input, list):
            feed_dict = {
                model_input: embeddings_data[idx][batch]
                for idx, model_input in enumerate(self.model.input)
            }
          else:
            feed_dict = {self.model.input: embeddings_data[0][batch]}

          feed_dict.update({self.batch_id: i, self.step: step})

          if self.model.uses_learning_phase:
            feed_dict[K.learning_phase()] = False

          self.sess.run(self.assign_embeddings, feed_dict=feed_dict)
          self.saver.save(self.sess,
                          os.path.join(self.log_dir, 'keras_embedding.ckpt'),
                          epoch)

          i += self.batch_size

  def on_epoch_end(self, epoch, logs=None):
    logs = logs or {}

    if not self.validation_data and self.histogram_freq:
      raise ValueError('If printing histograms, validation_data must be '
                       'provided, and cannot be a generator.')
    if self.validation_data and self.histogram_freq:
      if epoch % self.histogram_freq == 0:

        val_data = self.validation_data
        tensors = (
            self.model.inputs + self.model.targets + self.model.sample_weights)

        if self.model.uses_learning_phase:
          tensors += [K.learning_phase()]

        assert len(val_data) == len(tensors)
        val_size = val_data[0].shape[0]
        i = 0
        while i < val_size:
          step = min(self.batch_size, val_size - i)
          batch_val = []
          batch_val.append(val_data[0][i:i + step]
                           if val_data[0] is not None else None)
          batch_val.append(val_data[1][i:i + step]
                           if val_data[1] is not None else None)
          batch_val.append(val_data[2][i:i + step]
                           if val_data[2] is not None else None)
          if self.model.uses_learning_phase:
            # do not slice the learning phase
            batch_val = [x[i:i + step] if x is not None else None
                         for x in val_data[:-1]]
            batch_val.append(val_data[-1])
          else:
            batch_val = [x[i:i + step] if x is not None else None
                         for x in val_data]
          feed_dict = {}
          for key, val in zip(tensors, batch_val):
            if val is not None:
              feed_dict[key] = val
          result = self.sess.run([self.merged], feed_dict=feed_dict)
          summary_str = result[0]
          self.writer.add_summary(summary_str, epoch)
          i += self.batch_size

    for name, value in logs.items():
      if name in ['batch', 'size']:
        continue
      summary = tf_summary.Summary()
      summary_value = summary.value.add()
      summary_value.simple_value = value.item()
      summary_value.tag = name
      self.writer.add_summary(summary, epoch)
    self.writer.flush()

  def call(self, inputs, training=None):
    if training is None:
      training = K.learning_phase()

    def dropped_inputs():
      return nn.dropout(inputs, 1  - self.rate,
                        noise_shape=self._get_noise_shape(inputs),
                        seed=self.seed)
    output = tf_utils.smart_cond(training,
                                 dropped_inputs,
                                 lambda: array_ops.identity(inputs))
    return output

  def call(self, inputs, training=None):
    original_training_value = training
    if training is None:
      training = K.learning_phase()

    def dropped_inputs():
      return nn.dropout(inputs, 1  - self.rate,
                        noise_shape=self._get_noise_shape(inputs),
                        seed=self.seed)
    output = tf_utils.smart_cond(training,
                                 dropped_inputs,
                                 lambda: array_ops.identity(inputs))
    # EagerTensor object has no attribute _uses_learning_phase
    if not context.executing_eagerly() and original_training_value is None:
      output._uses_learning_phase = True  # pylint: disable=protected-access
    return output

  def call(self, inputs, training=None):
    if training is None:
      training = K.learning_phase()

    if self.virtual_batch_size is not None:
      # Virtual batches (aka ghost batches) can be simulated by reshaping the
      # Tensor and reusing the existing batch norm implementation
      original_shape = [-1] + inputs.shape.as_list()[1:]
      expanded_shape = [self.virtual_batch_size, -1] + original_shape[1:]

      # Will cause errors if virtual_batch_size does not divide the batch size
      inputs = array_ops.reshape(inputs, expanded_shape)

      def undo_virtual_batching(outputs):
        outputs = array_ops.reshape(outputs, original_shape)
        return outputs

    if self.fused:
      outputs = self._fused_batch_norm(inputs, training=training)
      if self.virtual_batch_size is not None:
        # Currently never reaches here since fused_batch_norm does not support
        # virtual batching
        outputs = undo_virtual_batching(outputs)
      return outputs

    # Compute the axes along which to reduce the mean / variance
    input_shape = inputs.shape
    ndims = len(input_shape)
    reduction_axes = [i for i in range(ndims) if i not in self.axis]
    if self.virtual_batch_size is not None:
      del reduction_axes[1]     # Do not reduce along virtual batch dim

    # Broadcasting only necessary for single-axis batch norm where the axis is
    # not the last dimension
    broadcast_shape = [1] * ndims
    broadcast_shape[self.axis[0]] = input_shape.dims[self.axis[0]].value
    def _broadcast(v):
      if (v is not None and len(v.shape) != ndims and
          reduction_axes != list(range(ndims - 1))):
        return array_ops.reshape(v, broadcast_shape)
      return v

    scale, offset = _broadcast(self.gamma), _broadcast(self.beta)

    def _compose_transforms(scale, offset, then_scale, then_offset):
      if then_scale is not None:
        scale *= then_scale
        offset *= then_scale
      if then_offset is not None:
        offset += then_offset
      return (scale, offset)

    # Determine a boolean value for `training`: could be True, False, or None.
    training_value = tf_utils.constant_value(training)
    if training_value is not False:
      if self.adjustment:
        adj_scale, adj_bias = self.adjustment(array_ops.shape(inputs))
        # Adjust only during training.
        adj_scale = tf_utils.smart_cond(training,
                                        lambda: adj_scale,
                                        lambda: array_ops.ones_like(adj_scale))
        adj_bias = tf_utils.smart_cond(training,
                                       lambda: adj_bias,
                                       lambda: array_ops.zeros_like(adj_bias))
        scale, offset = _compose_transforms(adj_scale, adj_bias, scale, offset)

      # Some of the computations here are not necessary when training==False
      # but not a constant. However, this makes the code simpler.
      keep_dims = self.virtual_batch_size is not None or len(self.axis) > 1
      mean, variance = self._moments(
          math_ops.cast(inputs, self._param_dtype),
          reduction_axes,
          keep_dims=keep_dims)

      moving_mean = self.moving_mean
      moving_variance = self.moving_variance

      mean = tf_utils.smart_cond(training,
                                 lambda: mean,
                                 lambda: moving_mean)
      variance = tf_utils.smart_cond(training,
                                     lambda: variance,
                                     lambda: moving_variance)

      if self.virtual_batch_size is not None:
        # This isn't strictly correct since in ghost batch norm, you are
        # supposed to sequentially update the moving_mean and moving_variance
        # with each sub-batch. However, since the moving statistics are only
        # used during evaluation, it is more efficient to just update in one
        # step and should not make a significant difference in the result.
        new_mean = math_ops.reduce_mean(mean, axis=1, keepdims=True)
        new_variance = math_ops.reduce_mean(variance, axis=1, keepdims=True)
      else:
        new_mean, new_variance = mean, variance

      if self.renorm:
        r, d, new_mean, new_variance = self._renorm_correction_and_moments(
            new_mean, new_variance, training)
        # When training, the normalized values (say, x) will be transformed as
        # x * gamma + beta without renorm, and (x * r + d) * gamma + beta
#.........这里部分代码省略.........

def predict_loop(model, iterator, verbose=0, steps=None):
  """Predict loop for predicting with DistributionStrategy.

  Arguments:
      model: Keras Model instance.
      iterator: Iterator for input data.
      verbose: Integer, Verbosity mode 0 or 1.
      steps: Total number of steps (batches of samples)
          before declaring `_predict_loop` finished.
          Ignored with the default value of `None`.

  Returns:
      Array of predictions (if the model has a single output)
      or list of arrays of predictions
      (if the model has multiple outputs).
  """
  current_strategy = model._distribution_strategy

  # TODO(priyag, sourabhbajaj): Remove this when the codepaths are merged.
  if current_strategy.__class__.__name__ == 'TPUStrategy':
    return _experimental_predict_loop(model, iterator, verbose, steps)

  if not model._grouped_model:
    clone_model_on_replicas(model, current_strategy)

  def _per_device_predict_function(model):
    model._make_predict_function()
    return (model.predict_function.inputs,
            model.predict_function.outputs,
            model.predict_function.updates_op,
            model.predict_function.session_kwargs)

  inputs, _, _ = _get_input_from_iterator(iterator, model)
  with current_strategy.scope():
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_replica(
         _per_device_predict_function, args=(model._grouped_model,))

    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs, grouped_updates,
         grouped_session_args)

    dataset_inputs = distributed_training_utils.flatten_perdevice_values(
        current_strategy, inputs)

    distributed_predict_function = K.function(
        all_inputs, all_outputs,
        updates=all_updates,
        name='distributed_predict_function',
        **all_session_args)

    if not isinstance(K.learning_phase(), int):
      ins = dataset_inputs + [0]
    else:
      ins = dataset_inputs

    if verbose == 1:
      progbar = Progbar(target=steps)

    # Copy the weights from the original model to each of the replicated models.
    orig_model_weights = model.get_weights()
    distributed_model = current_strategy.unwrap(model._grouped_model)[0]
    distributed_training_utils.set_weights(
        current_strategy, distributed_model, orig_model_weights)

    num_replicas = current_strategy.num_replicas_in_sync
    # Since we do not know how many samples we will see, we cannot
    # pre-allocate the returned Numpy arrays. Instead, we store one array per
    # batch seen and concatenate them upon returning.
    unconcatenated_outs = []
    assert steps is not None
    for step in range(steps):
      batch_outs = distributed_predict_function(ins)
      if not isinstance(batch_outs, list):
        batch_outs = [batch_outs]
      if step == 0:
        # batch_outs gives you the number of model outputs. In the distributed
        # case this will be number of model_outputs * num_replicas.
        for _ in range(len(model.outputs)):
          unconcatenated_outs.append([])
      for i in range(len(model.outputs)):
        nested_outs = batch_outs[i * num_replicas:
                                 i * num_replicas + num_replicas]
        outs = nest.flatten(nested_outs)
        unconcatenated_outs[i].extend(outs)
      if verbose >= 1:
        progbar.update(step + 1)
    if len(unconcatenated_outs) == 1:
      return np.concatenate(unconcatenated_outs[0], axis=0)
    return [
        np.concatenate(unconcatenated_outs[i], axis=0)
        for i in range(len(unconcatenated_outs))
    ]

def fit_loop(
    model,
    iterator,
    epochs=100,
    verbose=1,
    callbacks=None,
    val_iterator=None,
    initial_epoch=0,
    steps_per_epoch=None,
    validation_steps=None):
  """Fit loop for training with DistributionStrategy.

  Arguments:
      model: Keras Model instance.
      iterator: Iterator for input data.
      epochs: Number of times to iterate over the data
      verbose: Integer, Verbosity mode, 0, 1 or 2
      callbacks: List of callbacks to be called during training
      val_iterator: Iterator for validation data.
      initial_epoch: Epoch at which to start training
          (useful for resuming a previous training run)
      steps_per_epoch: Total number of steps (batches of samples)
          before declaring one epoch finished and starting the
          next epoch. Ignored with the default value of `None`.
      validation_steps: Number of steps to run validation for
          (only if doing validation from data tensors).
          Ignored with the default value of `None`.

  Returns:
      `History` object.

  Raises:
      ValueError: in case of invalid arguments.
  """
  current_strategy = model._distribution_strategy

  # TODO(priyag, sourabhbajaj): Remove this when the codepaths are merged.
  if current_strategy.__class__.__name__ == 'TPUStrategy':
    return _experimental_fit_loop(
        model, iterator, epochs, verbose, callbacks, initial_epoch,
        steps_per_epoch, val_iterator, validation_steps)

  if not model._grouped_model:
    clone_model_on_replicas(model, current_strategy, make_callback_model=True)

  def _per_device_fit_function(model):
    model._make_fit_function()
    return (model._fit_function.inputs, model._fit_function.outputs,
            model._fit_function.updates_op, model._fit_function.session_kwargs)

  inputs, targets, sample_weights = _get_input_from_iterator(iterator, model)
  with current_strategy.scope():
    # Create train ops on each of the devices when we call
    # `_per_device_fit_function`.
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_replica(
         _per_device_fit_function, args=(model._grouped_model,))
    # Unwrap all the per device values returned from `call_for_each_replica`.
    # Unwrapping per device values gives you a list of values that can be
    # used to construct a new train function that is composed of update ops on
    # all the devices over which the model is distributed.
    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs,
         grouped_updates, grouped_session_args, with_loss_tensor=True)

    # Dataset inputs and targets are also per devices values that need to be
    # unwrapped.
    dataset_inputs = distributed_training_utils.flatten_perdevice_values(
        current_strategy, inputs)
    dataset_targets = distributed_training_utils.flatten_perdevice_values(
        current_strategy, targets)

    # Create a train function that is composed of all the parameters above.
    distributed_fit_function = K.function(
        all_inputs,
        all_outputs,
        updates=all_updates,
        name='distributed_fit_function',
        **all_session_args)

    # We need to set sample_weights to None since there are sample weight
    # placeholders that are created with default values.
    sample_weights = [None for _ in range(
        len(model.outputs) * current_strategy.num_replicas_in_sync)]
    if not isinstance(K.learning_phase(), int):
      ins = dataset_inputs + dataset_targets + sample_weights + [1]
    else:
      ins = dataset_inputs + dataset_targets

    do_validation = False
    if validation_steps:
      do_validation = True

    # Copy the weights from the original model to each of the replicated models.
    orig_model_weights = model.get_weights()
    distributed_model = current_strategy.unwrap(model._grouped_model)[0]
    distributed_training_utils.set_weights(
        current_strategy, distributed_model, orig_model_weights)

#.........这里部分代码省略.........

def test_loop(model, iterator, verbose=0, steps=None):
  """Test loop for evaluating with DistributionStrategy.

  Arguments:
      model: Keras Model instance.
      iterator: Iterator for input data.
      verbose: Integer, Verbosity mode 0 or 1.
      steps: Total number of steps (batches of samples)
          before declaring predictions finished.
          Ignored with the default value of `None`.

  Returns:
      Scalar loss (if the model has a single output and no metrics)
      or list of scalars (if the model has multiple outputs
      and/or metrics). The attribute `model.metrics_names` will give you
      the display labels for the outputs.
  """
  current_strategy = model._distribution_strategy

  # TODO(priyag, sourabhbajaj): Remove this when the codepaths are merged.
  if current_strategy.__class__.__name__ == 'TPUStrategy':
    return _experimental_test_loop(model, iterator, verbose, steps)

  if not model._grouped_model:
    clone_model_on_replicas(model, current_strategy)

  def _per_device_eval_function(model):
    model._make_eval_function()
    return (model._eval_function.inputs, model._eval_function.outputs,
            model._eval_function.updates_op,
            model._eval_function.session_kwargs)

  inputs, targets, sample_weights = _get_input_from_iterator(iterator, model)
  with current_strategy.scope():
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_replica(
         _per_device_eval_function, args=(model._grouped_model,))

    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs, grouped_updates,
         grouped_session_args, with_loss_tensor=True)

    dataset_inputs = distributed_training_utils.flatten_perdevice_values(
        current_strategy, inputs)
    dataset_targets = distributed_training_utils.flatten_perdevice_values(
        current_strategy, targets)

    distributed_test_function = K.function(
        all_inputs, all_outputs,
        updates=all_updates,
        name='distributed_test_function',
        **all_session_args)

    # We need to set sample_weights to None since there are sample weight
    # placeholders that are created with default values.
    sample_weights = [None for _ in range(
        len(model.outputs) * current_strategy.num_replicas_in_sync)]
    if not isinstance(K.learning_phase(), int):
      ins = dataset_inputs + dataset_targets + sample_weights + [0]
    else:
      ins = dataset_inputs + dataset_targets

    for m in model.stateful_metric_functions:
      m.reset_states()

    outs = []
    if verbose == 1:
      progbar = Progbar(target=steps)

    # Copy the weights from the original model to each of the replicated models.
    orig_model_weights = model.get_weights()
    distributed_model = current_strategy.unwrap(model._grouped_model)[0]
    distributed_training_utils.set_weights(
        current_strategy, distributed_model, orig_model_weights)

    assert steps is not None
    for step in range(steps):
      batch_outs = distributed_test_function(ins)
      if isinstance(batch_outs, list):
        if step == 0:
          outs = [0.] * len(batch_outs)
        outs[0] += batch_outs[0]  # index 0 = 'loss'
        outs[1:] = batch_outs[1:]
      else:
        if step == 0:
          outs.append(0.)
        outs[0] += batch_outs  # index 0 = 'loss'
      if verbose >= 1:
        progbar.update(step + 1)
    outs[0] /= steps  # index 0 = 'loss'

    if len(outs) == 1:
      return outs[0]
    return outs

def fit_loop(model,
             inputs,
             targets,
             sample_weights=None,
             batch_size=None,
             epochs=100,
             verbose=1,
             callbacks=None,
             val_inputs=None,
             val_targets=None,
             val_sample_weights=None,
             shuffle=True,
             initial_epoch=0,
             steps_per_epoch=None,
             validation_steps=None):
  """Abstract fit function for arrays of data.

  Arguments:
      model: Keras Model instance.
      inputs: List of input arrays.
      targets: List of target arrays.
      sample_weights: Optional list of sample weight arrays.
      batch_size: Integer batch size or None if unknown.
      epochs: Number of times to iterate over the data
      verbose: Verbosity mode, 0, 1 or 2
      callbacks: List of callbacks to be called during training
      val_inputs: List of input arrays.
      val_targets: List of target arrays.
      val_sample_weights: Optional list of sample weight arrays.
      shuffle: Whether to shuffle the data at the beginning of each epoch
          concatenation of list the display names of the outputs of
           `f` and the list of display names of the outputs of `f_val`.
      initial_epoch: Epoch at which to start training
          (useful for resuming a previous training run)
      steps_per_epoch: Total number of steps (batches of samples)
          before declaring one epoch finished and starting the
          next epoch. Ignored with the default value of `None`.
      validation_steps: Number of steps to run validation for
          (only if doing validation from data tensors).
          Ignored with the default value of `None`.

  Returns:
      `History` object.

  Raises:
      ValueError: in case of invalid arguments.
  """
  model._make_train_function()
  f = model.train_function

  sample_weights = sample_weights or []
  val_sample_weights = val_sample_weights or []
  if model.uses_learning_phase and not isinstance(K.learning_phase(), int):
    ins = inputs + targets + sample_weights + [1]
  else:
    ins = inputs + targets + sample_weights

  do_validation = False
  if val_inputs:
    do_validation = True
    if (steps_per_epoch is None and verbose and inputs and
        hasattr(inputs[0], 'shape') and hasattr(val_inputs[0], 'shape')):
      print('Train on %d samples, validate on %d samples' %
            (inputs[0].shape[0], val_inputs[0].shape[0]))
  if validation_steps:
    do_validation = True
    if steps_per_epoch is None:
      raise ValueError('Can only use `validation_steps` '
                       'when doing step-wise '
                       'training, i.e. `steps_per_epoch` '
                       'must be set.')

  num_train_samples = training_utils.check_num_samples(
      ins, batch_size, steps_per_epoch, 'steps_per_epoch')
  count_mode = 'steps' if steps_per_epoch else 'samples'
  callbacks = cbks.configure_callbacks(
      callbacks,
      model,
      do_validation=do_validation,
      val_inputs=val_inputs,
      val_targets=val_targets,
      val_sample_weights=val_sample_weights,
      batch_size=batch_size,
      epochs=epochs,
      steps_per_epoch=steps_per_epoch,
      samples=num_train_samples,
      validation_steps=validation_steps,
      verbose=verbose,
      count_mode=count_mode)

  if num_train_samples is not None:
    index_array = np.arange(num_train_samples)

  # To prevent a slowdown, we find beforehand the arrays that need conversion.
  feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights
  indices_for_conversion_to_dense = []
  for i in range(len(feed)):
    if issparse is not None and issparse(ins[i]) and not K.is_sparse(feed[i]):
      indices_for_conversion_to_dense.append(i)

#.........这里部分代码省略.........

def test_loop(model,
              inputs,
              targets,
              sample_weights=None,
              batch_size=None,
              verbose=0,
              steps=None):
  """Abstract method to loop over some data in batches.

  Arguments:
      model: Keras Model instance.
      inputs: List of input arrays.
      targets: List of target arrays.
      sample_weights: Optional list of sample weight arrays.
      batch_size: integer batch size or `None`.
      verbose: verbosity mode.
      steps: Total number of steps (batches of samples)
          before declaring predictions finished.
          Ignored with the default value of `None`.

  Returns:
      Scalar loss (if the model has a single output and no metrics)
      or list of scalars (if the model has multiple outputs
      and/or metrics). The attribute `model.metrics_names` will give you
      the display labels for the scalar outputs.
  """
  model._make_test_function()
  f = model.test_function

  sample_weights = sample_weights or []
  if model.uses_learning_phase and not isinstance(K.learning_phase(), int):
    ins = inputs + targets + sample_weights + [0]
  else:
    ins = inputs + targets + sample_weights

  if hasattr(model, 'metrics'):
    for m in model.stateful_metric_functions:
      m.reset_states()
    stateful_metric_indices = [
        i for i, name in enumerate(model.metrics_names)
        if str(name) in model.stateful_metric_names
    ]
  else:
    stateful_metric_indices = []

  num_samples = training_utils.check_num_samples(
      ins, batch_size, steps, 'steps')
  outs = []
  if verbose == 1:
    if steps is not None:
      progbar = Progbar(target=steps)
    else:
      progbar = Progbar(target=num_samples)

  # To prevent a slowdown, we find beforehand the arrays that need conversion.
  feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights
  indices_for_conversion_to_dense = []
  for i in range(len(feed)):
    if issparse is not None and issparse(ins[i]) and not K.is_sparse(feed[i]):
      indices_for_conversion_to_dense.append(i)

  if steps is not None:
    for step in range(steps):
      batch_outs = f(ins)
      if isinstance(batch_outs, list):
        if step == 0:
          for _ in enumerate(batch_outs):
            outs.append(0.)
        for i, batch_out in enumerate(batch_outs):
          if i in stateful_metric_indices:
            outs[i] = batch_out
          else:
            outs[i] += batch_out
      else:
        if step == 0:
          outs.append(0.)
        outs[0] += batch_outs
      if verbose == 1:
        progbar.update(step + 1)
    for i in range(len(outs)):
      if i not in stateful_metric_indices:
        outs[i] /= steps
  else:
    batches = make_batches(num_samples, batch_size)
    index_array = np.arange(num_samples)
    for batch_index, (batch_start, batch_end) in enumerate(batches):
      batch_ids = index_array[batch_start:batch_end]
      if isinstance(ins[-1], int):
        # Do not slice the training phase flag.
        ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
      else:
        ins_batch = slice_arrays(ins, batch_ids)
      for i in indices_for_conversion_to_dense:
        ins_batch[i] = ins_batch[i].toarray()

      batch_outs = f(ins_batch)

      if isinstance(batch_outs, list):
        if batch_index == 0:
          outs.extend([0.] * len(batch_outs))
#.........这里部分代码省略.........

def model_iteration(model,
                    data,
                    steps_per_epoch=None,
                    epochs=1,
                    verbose=1,
                    callbacks=None,
                    validation_data=None,
                    validation_steps=None,
                    class_weight=None,
                    max_queue_size=10,
                    workers=1,
                    use_multiprocessing=False,
                    shuffle=True,
                    initial_epoch=0,
                    mode='train',
                    batch_size=None,
                    **kwargs):
  """Loop function for arrays of data with modes 'train'/'test'/'predict'.

  Arguments:
      model: Keras Model instance.
      data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or `(x, y)` or
        `(x, y, sample_weights)`) or a generator or
        `keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
      steps_per_epoch: Total number of steps (batches of samples) before
        declaring one epoch finished and starting the next epoch. Ignored with
        the default value of `None`.
      epochs: Number of times to iterate over the data.
      verbose: Verbosity mode, 0, 1 or 2.
      callbacks: List of callbacks to be called during training.
      validation_data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or
        `(x, y)` or `(x, y, sample_weights)`) or a generator or
        `keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
      validation_steps: Total number of steps (batches of samples) before
        declaring validation finished.
      class_weight: Dictionary mapping class indices to a weight for the class.
      max_queue_size: Integer. Maximum size for the generator queue. If
        unspecified, `max_queue_size` will default to 10.
      workers: Integer. Maximum number of processes to spin up when using
        process-based threading. If unspecified, `workers` will default to 1. If
        0, will execute the generator on the main thread.
      use_multiprocessing: Boolean. If `True`, use process-based threading. If
        unspecified, `use_multiprocessing` will default to `False`. Note that
        because this implementation relies on multiprocessing, you should not
        pass non-picklable arguments to the generator as they can't be passed
        easily to children processes.
      shuffle: Boolean. Whether to shuffle the order of the batches at the
        beginning of each epoch. Only used with instances of `Sequence`
        (`keras.utils.Sequence`). Has no effect when `steps_per_epoch` is not
        `None`.
      initial_epoch: Epoch at which to start training (useful for resuming a
        previous training run).
      mode: One of 'train'/'test'/'predict'.
      batch_size: Integer batch size or None if unknown. Will only be used if
        `data` is in NumPy/Tensor format.
      **kwargs: Additional arguments for backwards compatibility. `steps` is
        accepted as an alias for `steps_per_epoch`.

  Returns:
      - In 'train' mode: `History` object.
      - In 'test' mode: Evaluation metrics.
      - In 'predict' mode: Outputs of the Model called on inputs.

  Raises:
      ValueError: in case of invalid arguments.
  """
  if 'steps' in kwargs:
    steps_per_epoch = kwargs['steps']

  # Convert to a format that supports `next(generator)`.
  generator, steps_per_epoch = convert_to_generator_like(
      data,
      steps_per_epoch=steps_per_epoch,
      batch_size=batch_size,
      epochs=epochs - initial_epoch,
      shuffle=shuffle)

  do_validation = validation_data is not None
  should_set_learning_phase = context.executing_eagerly() and model.run_eagerly
  is_sequence = isinstance(generator, data_utils.Sequence)
  _validate_arguments(is_sequence, use_multiprocessing, workers,
                      steps_per_epoch, validation_data, validation_steps, mode,
                      kwargs)

  batch_function = _make_execution_function(
      model, mode, class_weight=class_weight)

  # Create the queue for the generator.
  output_generator, enqueuer = _make_enqueued_generator(
      generator,
      workers=workers,
      use_multiprocessing=use_multiprocessing,
      max_queue_size=max_queue_size,
      shuffle=shuffle)

  num_samples_or_steps, use_steps = _get_num_samples_or_steps(
      data, steps_per_epoch)

  count_mode = 'steps' if use_steps else 'samples'
  callbacks = cbks.configure_callbacks(
#.........这里部分代码省略.........

def fit_loop(
    model,
    inputs,
    targets,
    epochs=100,
    verbose=1,
    callbacks=None,
    val_inputs=None,
    val_targets=None,
    initial_epoch=0,
    steps_per_epoch=None,
    validation_steps=None):
  """fit function when using DistributionStrategy for training.

  Arguments:
      model: Keras Model instance.
      inputs: List of input arrays.
      targets: List of target arrays.
      epochs: Number of times to iterate over the data
      verbose: Verbosity mode, 0, 1 or 2
      callbacks: List of callbacks to be called during training
      val_inputs: List of input arrays.
      val_targets: List of target arrays.
      initial_epoch: Epoch at which to start training
          (useful for resuming a previous training run)
      steps_per_epoch: Total number of steps (batches of samples)
          before declaring one epoch finished and starting the
          next epoch. Ignored with the default value of `None`.
      validation_steps: Number of steps to run validation for
          (only if doing validation from data tensors).
          Ignored with the default value of `None`.

  Returns:
      `History` object.

  Raises:
      ValueError: in case of invalid arguments.
  """
  current_strategy = model._distribution_strategy
  def _per_device_train_function(model):
    model._make_train_function()
    return (model.train_function.inputs,
            model.train_function.outputs,
            model.train_function.updates_op,
            model.train_function.session_kwargs)

  with current_strategy.scope():
    # Create train ops on each of the devices when we call
    # `_per_device_train_function`.
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_tower(
         _per_device_train_function, model._grouped_model)
    # Unwrap all the per device values returned from `call_for_each_tower`.
    # Unwrapping per device values gives you a list of values that can be
    # used to construct a new train function that is composed of update ops on
    # all the devices over which the model is distributed.
    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs,
         grouped_updates, grouped_session_args, with_loss_tensor=True)

    # Dataset inputs and targets are also per devices values that need to be
    # unwrapped.
    dataset_inputs = distributed_training_utils.flatten_perdevice_values(
        current_strategy, inputs)
    dataset_targets = distributed_training_utils.flatten_perdevice_values(
        current_strategy, targets)

  # Create a train function that is composed of all the parameters above.
  distributed_train_function = K.Function(
      all_inputs, all_outputs,
      updates=all_updates,
      name='distributed_train_function',
      **all_session_args)

  # We need to set sample_weights to None since there are sample weight
  # placeholders that are created with default values.
  sample_weights = [None for _ in range(len(model.outputs) *
                                        current_strategy.num_towers)]
  if model.uses_learning_phase and not isinstance(K.learning_phase(), int):
    ins = dataset_inputs + dataset_targets + sample_weights + [1]
  else:
    ins = dataset_inputs + dataset_targets

  do_validation = False
  if validation_steps:
    do_validation = True
    if steps_per_epoch is None:
      raise ValueError('Can only use `validation_steps` '
                       'when doing step-wise '
                       'training, i.e. `steps_per_epoch` '
                       'must be set.')

  # Copy the weights from the original model to each of the replicated models.
  orig_model_weights = model.get_weights()
  with current_strategy.scope():
    distributed_model = current_strategy.unwrap(model._grouped_model)[0]
    distributed_training_utils.set_weights(
        current_strategy, distributed_model, orig_model_weights)

#.........这里部分代码省略.........

def predict_loop(model, inputs, verbose=0, steps=None):
  """Abstract method to loop over some data in batches.

  Arguments:
      model: Keras Model instance.
      inputs: list of tensors to be fed to `f`.
      verbose: verbosity mode.
      steps: Total number of steps (batches of samples)
          before declaring `_predict_loop` finished.
          Ignored with the default value of `None`.

  Returns:
      Array of predictions (if the model has a single output)
      or list of arrays of predictions
      (if the model has multiple outputs).
  """
  current_strategy = model._distribution_strategy
  def _per_device_predict_function(model):
    model._make_predict_function()
    return (model.predict_function.inputs,
            model.predict_function.outputs,
            model.predict_function.updates_op,
            model.predict_function.session_kwargs)

  with current_strategy.scope():
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_tower(
         _per_device_predict_function, model._grouped_model)

    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs, grouped_updates,
         grouped_session_args)

    dataset_inputs = distributed_training_utils.flatten_perdevice_values(
        current_strategy, inputs)

  distributed_predict_function = K.Function(
      all_inputs, all_outputs,
      updates=all_updates,
      name='distributed_predict_function',
      **all_session_args)

  if model.uses_learning_phase and not isinstance(K.learning_phase(), int):
    ins = dataset_inputs + [0]
  else:
    ins = dataset_inputs

  if verbose == 1:
    progbar = Progbar(target=steps)

  # Copy the weights from the original model to each of the replicated models.
  orig_model_weights = model.get_weights()
  with current_strategy.scope():
    distributed_model = current_strategy.unwrap(model._grouped_model)[0]
    distributed_training_utils.set_weights(
        current_strategy, distributed_model, orig_model_weights)

  if steps is not None:
    # Since we do not know how many samples we will see, we cannot pre-allocate
    # the returned Numpy arrays. Instead, we store one array per batch seen
    # and concatenate them upon returning.
    unconcatenated_outs = []
    for step in range(steps):
      batch_outs = distributed_predict_function(ins)
      if not isinstance(batch_outs, list):
        batch_outs = [batch_outs]
      if step == 0:
        for _ in batch_outs:
          unconcatenated_outs.append([])
      for i, batch_out in enumerate(batch_outs):
        unconcatenated_outs[i].append(batch_out)
      if verbose == 1:
        progbar.update(step + 1)
    if len(unconcatenated_outs) == 1:
      return np.concatenate(unconcatenated_outs[0], axis=0)
    return [
        np.concatenate(unconcatenated_outs[i], axis=0)
        for i in range(len(unconcatenated_outs))
    ]

def test_loop(model, inputs, targets, verbose=0, steps=None):
  """evaluate method to validate a model that uses DistributionStrategy.

  Arguments:
      model: Keras Model instance.
      inputs: List of input arrays.
      targets: List of target arrays.
      verbose: verbosity mode.
      steps: Total number of steps (batches of samples)
          before declaring predictions finished.
          Ignored with the default value of `None`.

  Returns:
      Scalar loss (if the model has a single output and no metrics)
      or list of scalars (if the model has multiple outputs
      and/or metrics). The attribute `model.metrics_names` will give you
      the display labels for the scalar outputs.
  """
  current_strategy = model._distribution_strategy
  def _per_device_test_function(model):
    model._make_test_function()
    return (model.test_function.inputs,
            model.test_function.outputs,
            model.test_function.updates_op,
            model.test_function.session_kwargs)

  with current_strategy.scope():
    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_tower(
         _per_device_test_function, model._grouped_model)

    (all_inputs, all_outputs, all_updates,
     all_session_args) = distributed_training_utils.unwrap_values(
         current_strategy, grouped_inputs, grouped_outputs, grou