def test_merge_subtract(self):
    i1 = keras.layers.Input(shape=(4, 5))
    i2 = keras.layers.Input(shape=(4, 5))
    i3 = keras.layers.Input(shape=(4, 5))

    subtract_layer = keras.layers.Subtract()
    o = subtract_layer([i1, i2])
    self.assertListEqual(o.shape.as_list(), [None, 4, 5])
    model = keras.models.Model([i1, i2], o)
    model.run_eagerly = testing_utils.should_run_eagerly()

    x1 = np.random.random((2, 4, 5))
    x2 = np.random.random((2, 4, 5))
    out = model.predict([x1, x2])
    self.assertEqual(out.shape, (2, 4, 5))
    self.assertAllClose(out, x1 - x2, atol=1e-4)

    self.assertEqual(subtract_layer.compute_mask([i1, i2], [None, None]), None)
    self.assertTrue(
        np.all(
            K.eval(
                subtract_layer.compute_mask(
                    [i1, i2], [K.variable(x1), K.variable(x2)]))))

    with self.assertRaisesRegexp(ValueError, "`mask` should be a list."):
      subtract_layer.compute_mask([i1, i2], x1)
    with self.assertRaisesRegexp(ValueError, "`inputs` should be a list."):
      subtract_layer.compute_mask(i1, [None, None])
    with self.assertRaisesRegexp(ValueError,
                                 "layer should be called on exactly 2 inputs"):
      subtract_layer([i1, i2, i3])
    with self.assertRaisesRegexp(ValueError,
                                 "layer should be called on exactly 2 inputs"):
      subtract_layer([i1])

 def test_metrics(self):
   with self.test_session():
     y_a = K.variable(np.random.random((6, 7)))
     y_b = K.variable(np.random.random((6, 7)))
     for metric in [metrics.binary_accuracy, metrics.categorical_accuracy]:
       output = metric(y_a, y_b)
       self.assertEqual(K.eval(output).shape, (6,))

  def test_merge_concatenate(self):
    i1 = keras.layers.Input(shape=(4, 5))
    i2 = keras.layers.Input(shape=(4, 5))
    concat_layer = keras.layers.Concatenate(axis=1)
    o = concat_layer([i1, i2])
    self.assertListEqual(o.shape.as_list(), [None, 8, 5])
    model = keras.models.Model([i1, i2], o)
    model.run_eagerly = testing_utils.should_run_eagerly()

    x1 = np.random.random((2, 4, 5))
    x2 = np.random.random((2, 4, 5))
    out = model.predict([x1, x2])
    self.assertEqual(out.shape, (2, 8, 5))
    self.assertAllClose(out, np.concatenate([x1, x2], axis=1), atol=1e-4)

    self.assertEqual(concat_layer.compute_mask([i1, i2], [None, None]), None)
    self.assertTrue(
        np.all(
            K.eval(
                concat_layer.compute_mask(
                    [i1, i2], [K.variable(x1), K.variable(x2)]))))

    with self.assertRaisesRegexp(ValueError, "`mask` should be a list."):
      concat_layer.compute_mask([i1, i2], x1)
    with self.assertRaisesRegexp(ValueError, "`inputs` should be a list."):
      concat_layer.compute_mask(i1, [None, None])
    with self.assertRaisesRegexp(ValueError, "should have the same length"):
      concat_layer.compute_mask([i1, i2], [None])
    with self.assertRaisesRegexp(ValueError,
                                 "layer should be called on a list of inputs"):
      concat_layer(i1)

  def test_merge_add(self):
    i1 = keras.layers.Input(shape=(4, 5))
    i2 = keras.layers.Input(shape=(4, 5))
    i3 = keras.layers.Input(shape=(4, 5))

    add_layer = keras.layers.Add()
    o = add_layer([i1, i2, i3])
    self.assertListEqual(o.shape.as_list(), [None, 4, 5])
    model = keras.models.Model([i1, i2, i3], o)
    model.run_eagerly = testing_utils.should_run_eagerly()

    x1 = np.random.random((2, 4, 5))
    x2 = np.random.random((2, 4, 5))
    x3 = np.random.random((2, 4, 5))
    out = model.predict([x1, x2, x3])
    self.assertEqual(out.shape, (2, 4, 5))
    self.assertAllClose(out, x1 + x2 + x3, atol=1e-4)

    self.assertEqual(
        add_layer.compute_mask([i1, i2, i3], [None, None, None]), None)
    self.assertTrue(
        np.all(
            K.eval(
                add_layer.compute_mask(
                    [i1, i2], [K.variable(x1), K.variable(x2)]))))

    with self.assertRaisesRegexp(ValueError, "`mask` should be a list."):
      add_layer.compute_mask([i1, i2, i3], x1)
    with self.assertRaisesRegexp(ValueError, "`inputs` should be a list."):
      add_layer.compute_mask(i1, [None, None, None])
    with self.assertRaisesRegexp(ValueError, " should have the same length."):
      add_layer.compute_mask([i1, i2, i3], [None, None])

  def test_sparse_top_k_categorical_accuracy(self):
    with self.cached_session():
      # Test correctness if the shape of y_true is (num_samples, 1)
      y_pred = K.variable(np.array([[0.3, 0.2, 0.1], [0.1, 0.2, 0.7]]))
      y_true = K.variable(np.array([[1], [0]]))
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=3))
      self.assertEqual(result, 1)
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=2))
      self.assertEqual(result, 0.5)
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=1))
      self.assertEqual(result, 0.)

      # Test correctness if the shape of y_true is (num_samples,)
      y_pred = K.variable(np.array([[0.3, 0.2, 0.1], [0.1, 0.2, 0.7]]))
      y_true = K.variable(np.array([1, 0]))
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=3))
      self.assertEqual(result, 1)
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=2))
      self.assertEqual(result, 0.5)
      result = K.eval(
          metrics.sparse_top_k_categorical_accuracy(y_true, y_pred, k=1))
      self.assertEqual(result, 0.)

def offset_sep_conv2d_eval(depth, padding, x):
    """Perform a separable conv2d on x with a given padding"""
    depthwise_kernel = K.variable(value=np.array([[[[1]] * depth]]),
                                  dtype='float32')
    pointwise_kernel = K.variable(value=np.array([[[[1]] + [[0]] * (depth - 1)]]),
        dtype='float32')
    return K.separable_conv2d(x, depthwise_kernel,
                              pointwise_kernel, strides=(3, 3), padding=padding)

 def __init__(self, lr=0.01, momentum=0., decay=0., nesterov=False, **kwargs):
   super(SGD, self).__init__(**kwargs)
   with K.name_scope(self.__class__.__name__):
     self.iterations = K.variable(0, dtype='int64', name='iterations')
     self.lr = K.variable(lr, name='lr')
     self.momentum = K.variable(momentum, name='momentum')
     self.decay = K.variable(decay, name='decay')
   self.initial_decay = decay
   self.nesterov = nesterov

 def test_top_k_categorical_accuracy(self):
   with self.test_session():
     y_pred = K.variable(np.array([[0.3, 0.2, 0.1], [0.1, 0.2, 0.7]]))
     y_true = K.variable(np.array([[0, 1, 0], [1, 0, 0]]))
     result = K.eval(metrics.top_k_categorical_accuracy(y_true, y_pred, k=3))
     self.assertEqual(result, 1)
     result = K.eval(metrics.top_k_categorical_accuracy(y_true, y_pred, k=2))
     self.assertEqual(result, 0.5)
     result = K.eval(metrics.top_k_categorical_accuracy(y_true, y_pred, k=1))
     self.assertEqual(result, 0.)

 def __init__(self, optimizer, iterations=None):  # pylint: disable=super-init-not-called
   self.optimizer = optimizer
   self._track_checkpointable(optimizer, name='optimizer')
   if iterations is None:
     with K.name_scope(self.__class__.__name__):
       self.iterations = K.variable(0, dtype='int64', name='iterations')
   else:
     self.iterations = iterations
   self._track_checkpointable(self.iterations, name='global_step')

def check_operation_offset(depth, eval_f, padding):
    """Check if backend used an offset while placing the filter
    e.g. during a convolution.
    TensorFlow is inconsistent in doing so depending
    on the type of operation, the used device (CPU/GPU) and the input depth.
    """
    in_arr = np.array([[[[i] * depth for i in range(6)]]])
    input_data = K.variable(value=in_arr, dtype='float32')
    output = eval_f(depth, padding, input_data)
    result = K.eval(output).flatten().tolist()
    assert result in [[0, 3], [1, 4]]
    return result == [1, 4]

  def test_sparse_categorical_accuracy(self):
    with self.cached_session():
      metric = metrics.sparse_categorical_accuracy
      y_true = K.variable(np.random.randint(0, 7, (6,)))
      y_pred = K.variable(np.random.random((6, 7)))
      self.assertEqual(K.eval(metric(y_true, y_pred)).shape, (6,))

      # Test correctness if the shape of y_true is (num_samples,)
      y_true = K.variable([1., 0., 0., 0.])
      y_pred = K.variable([[0.8, 0.2], [0.6, 0.4], [0.7, 0.3], [0.9, 0.1]])
      print(K.eval(metric(y_true, y_pred)))
      self.assertAllEqual(K.eval(metric(y_true, y_pred)), [0., 1., 1., 1.])

      # Test correctness if the shape of y_true is (num_samples, 1)
      y_true = K.variable([[1.], [0.], [0.], [0.]])
      y_pred = K.variable([[0.8, 0.2], [0.6, 0.4], [0.7, 0.3], [0.9, 0.1]])
      print(K.eval(metric(y_true, y_pred)))
      self.assertAllEqual(K.eval(metric(y_true, y_pred)), [0., 1., 1., 1.])

 def test_sparse_categorical_accuracy_float(self):
   with self.cached_session():
     metric = metrics.sparse_categorical_accuracy
     y_true = K.variable(np.random.random((6,)))
     y_pred = K.variable(np.random.random((6, 7)))
     self.assertEqual(K.eval(metric(y_true, y_pred)).shape, (6,))

def experimental_tpu_fit_loop(model,
                              dataset,
                              epochs=100,
                              verbose=1,
                              callbacks=None,
                              initial_epoch=0,
                              steps_per_epoch=None,
                              val_dataset=None,
                              validation_steps=None,
                              validation_freq=1):
  """Fit loop for training with TPU DistributionStrategy.

  Arguments:
      model: Keras Model instance.
      dataset: Dataset that returns inputs and targets
      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
      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`.
      val_dataset: Dataset for validation data.
      validation_steps: Number of steps to run validation for
          (only if doing validation from data tensors).
          Ignored with the default value of `None`.
      validation_freq: Only relevant if validation data is provided. Integer or
          `collections.Container` instance (e.g. list, tuple, etc.). If an
          integer, specifies how many training epochs to run before a new
          validation run is performed, e.g. `validation_freq=2` runs
          validation every 2 epochs. If a Container, specifies the epochs on
          which to run validation, e.g. `validation_freq=[1, 2, 10]` runs
          validation at the end of the 1st, 2nd, and 10th epochs.

  Returns:
      Returns `None`.

  Raises:
      ValueError: in case of invalid arguments.
  """
  mode = ModeKeys.TRAIN
  # TODO(fchollet): add support for `steps_per_epoch=None` in TPU loops.
  current_strategy = model._distribution_strategy
  iterator = distributed_training_utils.get_iterator(dataset, current_strategy)
  steps_per_epoch = training_utils.infer_steps_for_dataset(
      dataset, steps_per_epoch, epochs, steps_name='steps_per_epoch')
  if (current_strategy.extended.steps_per_run != 1 and
      steps_per_epoch is None):
    raise ValueError('`steps_per_epoch` should be specified when calling '
                     '`fit` on the model with TPUStrategy when '
                     '`steps_per_run` != 1 .')

  scope = distributed_training_utils.distributed_scope(
      strategy=current_strategy, learning_phase=1)
  scope.__enter__()

  out_labels = model.metrics_names or []

  step_fn = _make_step_fn(model, ModeKeys.TRAIN, current_strategy, out_labels)

  # Add initial dummy values for loss and other metric tensors.
  initial_loop_values = {}
  initial_loop_values['loss'] = constant_op.constant(1e7)
  for name in model.metrics_names[1:]:
    tensor = model._all_stateful_metrics_tensors[name]
    initial_loop_values[name] = array_ops.zeros(tensor.shape, tensor.dtype)

  use_steps = steps_per_epoch is not None
  if use_steps:
    iteration_value = min(steps_per_epoch,
                          current_strategy.extended.steps_per_run)
  else:
    iteration_value = current_strategy.extended.steps_per_run

  steps_per_run = K.variable(
      value=iteration_value,
      dtype='int32',
      name='steps_per_run')
  ctx = current_strategy.extended.experimental_run_steps_on_iterator(
      step_fn, iterator, iterations=steps_per_run,
      initial_loop_values=initial_loop_values)
  train_op = ctx.run_op
  output_tensors = ctx.last_step_outputs

  do_validation = bool(validation_steps)

  if model._compile_distribution:
    distributed_training_utils._copy_weights_to_distributed_model(model, mode)

  callbacks = cbks.configure_callbacks(
      callbacks,
      model,
      do_validation=do_validation,
      epochs=epochs,
      steps_per_epoch=steps_per_epoch,
      verbose=verbose,
      count_mode='steps',
      mode=mode)

#.........这里部分代码省略.........

  def test_dynamic_loss_scaling(self, strategy_fn, cloning=True):
    strategy = strategy_fn()
    initial_loss_scale = 2.
    batch_size = 4
    expected_gradient = backend.variable([initial_loss_scale / batch_size],
                                         dtype=dtypes.float16)
    # If this variable is set to True, the model below will have NaN gradients
    have_nan_gradients = backend.variable(False, dtype=dtypes.bool)
    with strategy.scope():
      with policy.policy_scope(policy.Policy('infer_float32_vars')):
        x = layers.Input(shape=(1,), batch_size=batch_size,
                         dtype=dtypes.float16)
        layer = AddLayer(assert_type=dtypes.float16)
        y = layer(x)
        identity_with_nan_grads = (
            mp_test_util.create_identity_with_nan_gradients_fn(
                have_nan_gradients))
        y = core.Lambda(identity_with_nan_grads)(y)
        identity_with_grad_check_fn = (
            mp_test_util.create_identity_with_grad_check_fn(
                expected_dtype=dtypes.float16,
                expected_gradient=expected_gradient))
        y = core.Lambda(identity_with_grad_check_fn)(y)
        y = math_ops.cast(y, dtypes.float32)
        model = models.Model(inputs=x, outputs=y)

        def loss_fn(y_true, y_pred):
          del y_true
          return math_ops.reduce_mean(y_pred)

        opt = gradient_descent.SGD(1.)
        loss_scale = loss_scale_module.DynamicLossScale(
            initial_loss_scale=initial_loss_scale, increment_period=2)
        opt = loss_scale_optimizer.LossScaleOptimizer(opt, loss_scale)
        model.compile(opt, loss=loss_fn, cloning=cloning)

    self.assertEqual(backend.eval(layer.v), 1)
    x = np.ones((batch_size, 1))
    y = np.ones((batch_size, 1))
    dataset = dataset_ops.Dataset.from_tensor_slices((x, y)).batch(batch_size)
    model.fit(dataset)
    # The variables starts with 1 and has a gradient of 1, so will go down by 1
    # each step.
    self.assertEqual(backend.eval(layer.v), 0)

    model.fit(dataset)
    self.assertEqual(backend.eval(layer.v), -1)

    # There have been two steps without NaNs, so the loss scale will double
    backend.set_value(expected_gradient,
                      backend.get_value(expected_gradient * 2))
    model.fit(dataset)
    self.assertEqual(backend.eval(layer.v), -2)

    # Next test with NaN gradients.
    backend.set_value(have_nan_gradients, True)
    model.fit(dataset)
    # Variable should not be updated
    self.assertEqual(backend.eval(layer.v), -2)

    # Test with finite gradients again
    backend.set_value(have_nan_gradients, False)
    # The loss scale will be halved due to the NaNs, so the gradient will also
    # be halved
    backend.set_value(expected_gradient,
                      backend.get_value(expected_gradient / 2))
    model.fit(dataset)
    self.assertEqual(backend.eval(layer.v), -3)

#.........这里部分代码省略.........
         current_strategy, grouped_inputs, grouped_outputs,
         grouped_updates, grouped_session_args)
    combined_fn = K.function(
        all_inputs,
        all_outputs,
        updates=all_updates,
        name='distributed_fit_function',
        **all_session_args)

    for label, output in zip(out_labels, combined_fn.outputs):
      if label == 'loss':
        reduce_op = ds_reduce_util.ReduceOp.SUM
      else:
        # We reduce all other metrics using mean for now. This is temporary
        # workaround until new metrics are in place.
        reduce_op = ds_reduce_util.ReduceOp.MEAN
      ctx.set_last_step_output(label, output, reduce_op)

    # TODO(priyag, sourabhbajaj): Ignoring these things from the combined_fn:
    # feed_dict, session kwargs, run options, run_metadata for now. These should
    # be handled appropriately
    return combined_fn.updates_op

  # Add initial dummy values for loss and other metric tensors.
  initial_loop_values = {}
  initial_loop_values['loss'] = constant_op.constant(1e7)
  for name in model.metrics_names[1:]:
    tensor = model._all_stateful_metrics_tensors[name]
    initial_loop_values[name] = array_ops.zeros(tensor.shape, tensor.dtype)

  if steps_per_epoch is None:
    raise ValueError('`steps_per_epoch` should be specified when calling '
                     '`fit` on the model.')
  steps_per_run = K.variable(
      value=min(steps_per_epoch, current_strategy.extended.steps_per_run),
      dtype='int32',
      name='steps_per_run')

  ctx = current_strategy.extended.experimental_run_steps_on_iterator(
      step_fn, iterator, iterations=steps_per_run,
      initial_loop_values=initial_loop_values)

  train_op = ctx.run_op
  output_tensors = ctx.last_step_outputs

  do_validation = bool(validation_steps)

  if model._compile_distribution:
    distributed_training_utils._copy_weights_to_distributed_model(model, mode)

  callbacks = cbks.configure_callbacks(
      callbacks,
      model,
      do_validation=do_validation,
      epochs=epochs,
      steps_per_epoch=steps_per_epoch,
      verbose=verbose,
      count_mode='steps',
      mode=mode)

  # Calculate the steps each time on the device.
  steps_to_run = [current_strategy.extended.steps_per_run] * (
      steps_per_epoch // current_strategy.extended.steps_per_run)
  if steps_per_epoch % current_strategy.extended.steps_per_run:
    steps_to_run.append(
        steps_per_epoch % current_strategy.extended.steps_per_run)

 def __init__(self, name='true_positives', **kwargs):
   super(BinaryTruePositives, self).__init__(name=name, **kwargs)
   self.true_positives = K.variable(value=0, dtype='int32')
   self.stateful = True

def get_locallyconnected_mask(input_shape,
                              kernel_shape,
                              strides,
                              padding,
                              data_format,
                              dtype):
  """Return a mask representing connectivity of a locally-connected operation.

  This method returns a masking tensor of 0s and 1s (of type `dtype`) that,
  when element-wise multiplied with a fully-connected weight tensor, masks out
  the weights between disconnected input-output pairs and thus implements local
  connectivity through a sparse fully-connected weight tensor.

  Assume an unshared convolution with given parameters is applied to an input
  having N spatial dimensions with `input_shape = (d_in1, ..., d_inN)`
  to produce an output with spatial shape `(d_out1, ..., d_outN)` (determined
  by layer parameters such as `strides`).

  This method returns a mask which can be broadcast-multiplied (element-wise)
  with a 2*(N+1)-D weight matrix (equivalent to a fully-connected layer between
  (N+1)-D activations (N spatial + 1 channel dimensions for input and output)
  to make it perform an unshared convolution with given `kernel_shape`,
  `strides`, `padding` and `data_format`.

  Arguments:
    input_shape: tuple of size N: `(d_in1, ..., d_inN)`
                 spatial shape of the input.
    kernel_shape: tuple of size N, spatial shape of the convolutional kernel
                  / receptive field.
    strides: tuple of size N, strides along each spatial dimension.
    padding: type of padding, string `"same"` or `"valid"`.
    data_format: a string, `"channels_first"` or `"channels_last"`.
    dtype: type of the layer operation, e.g. `tf.float64`.

  Returns:
    a `dtype`-tensor of shape
    `(1, d_in1, ..., d_inN, 1, d_out1, ..., d_outN)`
    if `data_format == `"channels_first"`, or
    `(d_in1, ..., d_inN, 1, d_out1, ..., d_outN, 1)`
    if `data_format == "channels_last"`.

  Raises:
    ValueError: if `data_format` is neither `"channels_first"` nor
                `"channels_last"`.
  """
  mask = conv_utils.conv_kernel_mask(
      input_shape=input_shape,
      kernel_shape=kernel_shape,
      strides=strides,
      padding=padding
  )

  ndims = int(mask.ndim / 2)
  mask = K.variable(mask, dtype)

  if data_format == 'channels_first':
    mask = K.expand_dims(mask, 0)
    mask = K.expand_dims(mask, - ndims - 1)

  elif data_format == 'channels_last':
    mask = K.expand_dims(mask, ndims)
    mask = K.expand_dims(mask, -1)

  else:
    raise ValueError('Unrecognized data_format: ' + str(data_format))

  return mask

 def test_sparse_categorical_accuracy(self):
   with self.test_session():
     metric = metrics.sparse_categorical_accuracy
     y_a = K.variable(np.random.randint(0, 7, (6,)))
     y_b = K.variable(np.random.random((6, 7)))
     self.assertEqual(K.eval(metric(y_a, y_b)).shape, (6,))

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

  Arguments:
      model: Keras Model instance.
      iterator: Iterator that returns inputs and targets
      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
      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`.
      val_iterator: Iterator for validation data.
      validation_steps: Number of steps to run validation for
          (only if doing validation from data tensors).
          Ignored with the default value of `None`.

  Returns:
      Returns `None`.

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

  K.get_session().run(current_strategy.initialize())

  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)

  # TODO(priyag, sourabhbajaj): This should likely not be hardcoded here.
  K.set_learning_phase(1)
  out_labels = model.metrics_names or []

  def step_fn(ctx, inputs, targets):
    """Clones the model and calls make_fit_function."""
    # TODO(priyag, sourabhbajaj): The model gets cloned every time
    # fit/test/predict is called. We should look into caching this keyed on
    # input shapes.
    clone_model_on_replicas(
        model,
        current_strategy,
        make_callback_model=True,
        inputs=inputs,
        targets=targets,
        mode=_Mode.TRAIN)

    (grouped_inputs, grouped_outputs, grouped_updates,
     grouped_session_args) = current_strategy.call_for_each_replica(
         _per_device_fit_function, args=(model._grouped_model_train,))
    (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)
    combined_fn = K.function(
        all_inputs,
        all_outputs,
        updates=all_updates,
        name='distributed_fit_function',
        **all_session_args)

    for label, output in zip(out_labels, combined_fn.outputs):
      if label == 'loss':
        aggregation = distribute_lib.get_loss_reduction()
      else:
        # We aggregate all other metrics using mean for now. This is temporary
        # workaround until new metrics are in place.
        aggregation = variable_scope.VariableAggregation.MEAN
      ctx.set_last_step_output(label, output, aggregation)

    # TODO(priyag, sourabhbajaj): Ignoring these things from the combined_fn:
    # feed_dict, session kwargs, run options, run_metadata for now. These should
    # be handled appropriately
    return combined_fn.updates_op

  # Add initial dummy values for loss and other metric tensors.
  initial_loop_values = {}
  initial_loop_values['loss'] = constant_op.constant(1e7)
  for name, tensor in zip(model.metrics_names[1:], model.metrics_tensors):
    initial_loop_values[name] = array_ops.zeros(tensor.shape, tensor.dtype)

  if steps_per_epoch is None:
    raise ValueError('`steps_per_epoch` should be specified when calling '
                     '`fit` on the model.')
  steps_per_run = K.variable(
      value=min(steps_per_epoch, current_strategy.steps_per_run),
      dtype='int32',
#.........这里部分代码省略.........

def offset_conv2d_eval(depth, padding, x):
    """Perform a conv2d on x with a given padding"""
    kernel = K.variable(value=np.array([[[[1]] + [[0]] * (depth - 1)]]),
        dtype='float32')
    return K.conv2d(x, kernel, strides=(3, 3), padding=padding)