Python tensorflow.python.ops.math_ops.reduce_mean() Examples

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.cached_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_a(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_a(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEqual(predictions.get_shape().as_list(), [eval_batch_size]) 

def loss(self, data, labels):
    """The loss to minimize while training."""

    if self.is_regression:
      diff = self.training_inference_graph(data) - math_ops.to_float(labels)
      mean_squared_error = math_ops.reduce_mean(diff * diff)
      root_mean_squared_error = math_ops.sqrt(mean_squared_error, name="loss")
      loss = root_mean_squared_error
    else:
      loss = math_ops.reduce_mean(
          nn_ops.sparse_softmax_cross_entropy_with_logits(
              self.training_inference_graph(data),
              array_ops.squeeze(math_ops.to_int32(labels))),
          name="loss")
    if self.regularizer:
      loss += layers.apply_regularization(self.regularizer,
                                          variables.trainable_variables())
    return loss 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 300, 400
    num_classes = 1000
    with self.test_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = alexnet.alexnet_v2(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = alexnet.alexnet_v2(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 4, 7, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 231, 231
    eval_height, eval_width = 281, 281
    num_classes = 1000
    with self.test_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = overfeat.overfeat(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = overfeat.overfeat(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def loss(self, data, labels):
    """The loss to minimize while training."""

    if self.is_regression:
      diff = self.training_inference_graph(data) - math_ops.to_float(labels)
      mean_squared_error = math_ops.reduce_mean(diff * diff)
      root_mean_squared_error = math_ops.sqrt(mean_squared_error, name="loss")
      loss = root_mean_squared_error
    else:
      loss = math_ops.reduce_mean(
          nn_ops.sparse_softmax_cross_entropy_with_logits(
              labels=array_ops.squeeze(math_ops.to_int32(labels)),
              logits=self.training_inference_graph(data)),
          name="loss")
    if self.regularizer:
      loss += layers.apply_regularization(self.regularizer,
                                          variables.trainable_variables())
    return loss 

def _rescale_eval_loss(loss, weights):
  """Rescales evaluation loss according to the given weights.

  The rescaling is needed because in the training loss weights are not
  considered in the denominator, whereas  for the evaluation loss we should
  divide by the sum of weights.

  The rescaling factor is:
    R = sum_{i} 1 / sum_{i} w_{i}

  Args:
    loss: the scalar weighted loss.
    weights: weight coefficients. Either a scalar, or a `Tensor` of shape
      [batch_size].

  Returns:
    The given loss multiplied by the rescaling factor.
  """
  rescaling_factor = math_ops.reduce_mean(weights)
  return math_ops.div(loss, rescaling_factor) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.test_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_16(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_16(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.cached_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_19(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_19(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEqual(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.cached_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_16(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_16(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEqual(predictions.get_shape().as_list(), [eval_batch_size]) 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 300, 400
    num_classes = 1000
    with self.cached_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = alexnet.alexnet_v2(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = alexnet.alexnet_v2(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 4, 7, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEqual(predictions.get_shape().as_list(), [eval_batch_size]) 

def loss(self, data, labels):
    """The loss to minimize while training."""

    if self.is_regression:
      diff = self.training_inference_graph(data) - math_ops.to_float(labels)
      mean_squared_error = math_ops.reduce_mean(diff * diff)
      root_mean_squared_error = math_ops.sqrt(mean_squared_error, name="loss")
      loss = root_mean_squared_error
    else:
      loss = math_ops.reduce_mean(
          nn_ops.sparse_softmax_cross_entropy_with_logits(
              labels=array_ops.squeeze(math_ops.to_int32(labels)),
              logits=self.training_inference_graph(data)),
          name="loss")
    if self.regularizer:
      loss += layers.apply_regularization(self.regularizer,
                                          variables.trainable_variables())
    return loss 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 231, 231
    eval_height, eval_width = 281, 281
    num_classes = 1000
    with self.cached_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = overfeat.overfeat(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = overfeat.overfeat(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEqual(predictions.get_shape().as_list(), [eval_batch_size]) 

def var(x, axis=None, keepdims=False):
  """Variance of a tensor, alongside the specified axis.

  Arguments:
      x: A tensor or variable.
      axis: An integer, the axis to compute the variance.
      keepdims: A boolean, whether to keep the dimensions or not.
          If `keepdims` is `False`, the rank of the tensor is reduced
          by 1. If `keepdims` is `True`,
          the reduced dimension is retained with length 1.

  Returns:
      A tensor with the variance of elements of `x`.
  """
  axis = _normalize_axis(axis, ndim(x))
  if x.dtype.base_dtype == dtypes_module.bool:
    x = math_ops.cast(x, floatx())
  m = math_ops.reduce_mean(x, reduction_indices=axis, keep_dims=True)
  devs_squared = math_ops.square(x - m)
  return math_ops.reduce_mean(
      devs_squared, reduction_indices=axis, keep_dims=keepdims) 

def _covariance(x, diag):
  """Defines the covariance operation of a matrix.

  Args:
    x: a matrix Tensor. Dimension 0 should contain the number of examples.
    diag: if True, it computes the diagonal covariance.

  Returns:
    A Tensor representing the covariance of x. In the case of
  diagonal matrix just the diagonal is returned.
  """
  num_points = math_ops.to_float(array_ops.shape(x)[0])
  x -= math_ops.reduce_mean(x, 0, keep_dims=True)
  if diag:
    cov = math_ops.reduce_sum(
        math_ops.square(x), 0, keep_dims=True) / (num_points - 1)
  else:
    cov = math_ops.matmul(x, x, transpose_a=True) / (num_points - 1)
  return cov 

def _covariance(x, diag):
  """Defines the covariance operation of a matrix.

  Args:
    x: a matrix Tensor. Dimension 0 should contain the number of examples.
    diag: if True, it computes the diagonal covariance.

  Returns:
    A Tensor representing the covariance of x. In the case of
  diagonal matrix just the diagonal is returned.
  """
  num_points = math_ops.to_float(array_ops.shape(x)[0])
  x -= math_ops.reduce_mean(x, 0, keep_dims=True)
  if diag:
    cov = math_ops.reduce_sum(
        math_ops.square(x), 0, keep_dims=True) / (num_points - 1)
  else:
    cov = math_ops.matmul(x, x, transpose_a=True) / (num_points - 1)
  return cov 

def testTrainEvalWithReuse(self):
    train_batch_size = 2
    eval_batch_size = 1
    train_height, train_width = 224, 224
    eval_height, eval_width = 256, 256
    num_classes = 1000
    with self.test_session():
      train_inputs = random_ops.random_uniform(
          (train_batch_size, train_height, train_width, 3))
      logits, _ = vgg.vgg_19(train_inputs)
      self.assertListEqual(logits.get_shape().as_list(),
                           [train_batch_size, num_classes])
      variable_scope.get_variable_scope().reuse_variables()
      eval_inputs = random_ops.random_uniform(
          (eval_batch_size, eval_height, eval_width, 3))
      logits, _ = vgg.vgg_19(
          eval_inputs, is_training=False, spatial_squeeze=False)
      self.assertListEqual(logits.get_shape().as_list(),
                           [eval_batch_size, 2, 2, num_classes])
      logits = math_ops.reduce_mean(logits, [1, 2])
      predictions = math_ops.argmax(logits, 1)
      self.assertEquals(predictions.get_shape().as_list(), [eval_batch_size]) 

def training_loss(self, logits, target, features, name="training_loss"):
    """Returns training loss tensor for this head.

    Training loss is different from the loss reported on the tensorboard as we
    should respect the example weights when computing the gradient.

      L = sum_{i} w_{i} * l_{i} / B

    where B is the number of examples in the batch, l_{i}, w_{i} are individual
    losses, and example weight.

    Args:
      logits: logits, a float tensor.
      target: either a tensor for labels or in multihead case, a dict of string
        to target tensor.
      features: features dict.
      name: Op name.

    Returns:
      Loss tensor.
    """
    target = target[self.name] if isinstance(target, dict) else target
    loss_unweighted = self._loss_fn(logits, target)

    weight_tensor = self.get_weight_tensor(features)
    if weight_tensor is None:
      return math_ops.reduce_mean(loss_unweighted, name=name)
    loss_weighted = self._weighted_loss(loss_unweighted, weight_tensor)
    return math_ops.reduce_mean(loss_weighted, name=name) 

def _centered_bias_step(logits_dimension, weight_collection, labels,
                        train_loss_fn):
  """Creates and returns training op for centered bias."""
  centered_bias = ops.get_collection(weight_collection)
  batch_size = array_ops.shape(labels)[0]
  logits = array_ops.reshape(
      array_ops.tile(centered_bias[0], [batch_size]),
      [batch_size, logits_dimension])
  with ops.name_scope(None, "centered_bias", (labels, logits)):
    centered_bias_loss = math_ops.reduce_mean(
        train_loss_fn(logits, labels), name="training_loss")
  # Learn central bias by an optimizer. 0.1 is a convervative lr for a
  # single variable.
  return training.AdagradOptimizer(0.1).minimize(
      centered_bias_loss, var_list=centered_bias) 

def loss(self, logits, target, features):
    """Returns loss tensor for this head.

    The loss returned is the weighted average.

      L = sum_{i} w_{i} * l_{i} / sum_{i} w_{i}

    Args:
      logits: logits, a float tensor.
      target: either a tensor for labels or in multihead case, a dict of string
        to target tensor.
      features: features dict.

    Returns:
      Loss tensor.
    """
    target = target[self.name] if isinstance(target, dict) else target
    loss_unweighted = self._loss_fn(logits, target)

    weight_tensor = self.get_weight_tensor(features)
    if weight_tensor is None:
      return math_ops.reduce_mean(loss_unweighted, name="loss")
    loss_weighted = self._weighted_loss(loss_unweighted, weight_tensor)
    return math_ops.div(
        math_ops.reduce_sum(loss_weighted),
        math_ops.to_float(math_ops.reduce_sum(weight_tensor)),
        name="loss") 

def testCreateFCFollowByDropout(self):
    height, width = 3, 3
    with self.cached_session() as sess:
      images = random_ops.random_uniform(
          (5, height, width, 3), seed=1, name='images')
      output = _layers.fully_connected(images, 50)
      num_elem_initial = math_ops.reduce_mean(
          math_ops.cast(output > 0, dtypes.float32))
      output = _layers.dropout(output)
      num_elem = math_ops.reduce_mean(math_ops.cast(output > 0, dtypes.float32))
      sess.run(variables_lib.global_variables_initializer())
      num_elem, num_elem_initial = sess.run([num_elem, num_elem_initial])
      self.assertLess(num_elem, num_elem_initial / 2 + 0.1)
      self.assertGreater(num_elem, num_elem_initial / 2 - 0.1) 

def _sample_mean(values):
  """Mean over sample indices.  In this module this is always [0]."""
  return math_ops.reduce_mean(values, reduction_indices=[0]) 

def _loss(loss_unweighted, weight, name):
  """Returns loss."""
  if weight is None:
    loss = math_ops.reduce_mean(loss_unweighted, name=name)
    return loss, loss
  else:
    loss_weighted = _weighted_loss(loss_unweighted, weight)
    weighted_average_loss = math_ops.div(
        math_ops.reduce_sum(loss_weighted),
        math_ops.to_float(math_ops.reduce_sum(weight)),
        name="weighted_average_loss")
    loss = math_ops.reduce_mean(loss_weighted, name=name)
    return loss, weighted_average_loss 

def testNpairsMultiLabel(self):
    with self.cached_session():
      num_data = 15
      feat_dim = 6
      num_classes = 10
      reg_lambda = 0.02

      embeddings_anchor = np.random.rand(num_data, feat_dim).astype(np.float32)
      embeddings_positive = np.random.rand(num_data, feat_dim).astype(
          np.float32)

      labels = np.random.randint(0, 2, (num_data, num_classes))
      # set entire column to one so that each row has at least one bit set.
      labels[:, -1] = 1

      # Compute the loss in NP
      reg_term = np.mean(np.sum(np.square(embeddings_anchor), 1))
      reg_term += np.mean(np.sum(np.square(embeddings_positive), 1))
      reg_term *= 0.25 * reg_lambda

      similarity_matrix = np.matmul(embeddings_anchor, embeddings_positive.T)

      labels_remapped = np.dot(labels, labels.T).astype(np.float)
      labels_remapped /= np.sum(labels_remapped, 1, keepdims=True)

      xent_loss = math_ops.reduce_mean(nn.softmax_cross_entropy_with_logits(
          logits=ops.convert_to_tensor(similarity_matrix),
          labels=ops.convert_to_tensor(labels_remapped))).eval()
      loss_np = xent_loss + reg_term

      # Compute the loss in TF
      loss_tf = metric_learning.npairs_loss_multilabel(
          sparse_labels=convert_to_list_of_sparse_tensor(labels),
          embeddings_anchor=ops.convert_to_tensor(embeddings_anchor),
          embeddings_positive=ops.convert_to_tensor(embeddings_positive),
          reg_lambda=reg_lambda)
      loss_tf = loss_tf.eval()

      self.assertAllClose(loss_np, loss_tf) 

def testCreateFCWithDropout(self):
    height, width = 3, 3
    with self.cached_session() as sess:
      images = random_ops.random_uniform(
          (5, height, width, 3), seed=1, name='images')
      output = _layers.fully_connected(
          images, 50, normalizer_fn=_layers.dropout)
      num_elem = math_ops.reduce_mean(math_ops.cast(output > 0, dtypes.float32))
      sess.run(variables_lib.global_variables_initializer())
      num_elem = sess.run(num_elem)
      self.assertLess(num_elem, 0.5)
      self.assertGreater(num_elem, 0.1) 

def testNoneLossCollection(self):
    logits = constant_op.constant([[1.2, 0.4, -1.0, -1.1]])
    labels = constant_op.constant([[1.0, 0.0, 0.0, 1.0]])
    losses = loss_ops.hinge_loss(logits, labels)
    self.assertFalse(loss_ops.get_losses())
    loss_ops.add_loss(math_ops.reduce_mean(losses), loss_collection=None)
    self.assertFalse(loss_ops.get_losses())
    with self.cached_session():
      self.assertAllClose(losses.eval(), [[0.0, 1.4, 0.0, 2.1]], atol=1e-3) 

def testCreateDropoutNoTraining(self):
    height, width = 3, 3
    with self.cached_session() as sess:
      images = random_ops.random_uniform(
          (5, height, width, 3), seed=1, name='images')
      num_elem_initial = math_ops.reduce_mean(
          math_ops.cast(images > 0, dtypes.float32))
      output = _layers.dropout(images, is_training=False)
      num_elem = math_ops.reduce_mean(math_ops.cast(output > 0, dtypes.float32))
      num_elem, num_elem_initial = sess.run([num_elem, num_elem_initial])
      self.assertEqual(num_elem, num_elem_initial)
      outputs, inputs = sess.run([output, images])
      self.assertAllClose(outputs, inputs) 

def testDropout(self):
    height, width = 10, 10
    with self.cached_session() as sess:
      images = random_ops.random_uniform(
          (5, height, width, 3), seed=1, name='images')
      num_elem_initial = math_ops.reduce_mean(
          math_ops.cast(images > 0, dtypes.float32))
      output = _layers.dropout(images)
      num_elem = math_ops.reduce_mean(math_ops.cast(output > 0, dtypes.float32))
      num_elem, num_elem_initial = sess.run([num_elem, num_elem_initial])
      self.assertLess(num_elem, num_elem_initial / 2 + 0.1)
      self.assertGreater(num_elem, num_elem_initial / 2 - 0.1) 

def testReuse(self):

    def f(x):
      return core_layers.dense(x, self.CHANNELS // 2)

    def g(x):
      return core_layers.dense(x, self.CHANNELS // 2)

    x = random_ops.random_uniform(
        [self.BATCH_SIZE, self.CHANNELS], dtype=dtypes.float32)
    x1, x2 = array_ops.split(x, 2, axis=-1)

    with variable_scope.variable_scope("test"):
      y1, y2 = rev_block_lib.rev_block(x1, x2, f, g, num_layers=self.NUM_LAYERS)

    num_vars_before = len(variables.global_variables())

    with variable_scope.variable_scope("test", reuse=True):
      y1, y2 = rev_block_lib.rev_block(x1, x2, f, g, num_layers=self.NUM_LAYERS)

    num_vars_after = len(variables.global_variables())
    self.assertEqual(num_vars_before, num_vars_after)

    loss = math_ops.reduce_mean(y1 + y2)
    _ = gradients_impl.gradients(loss,
                                 [x] + variables.trainable_variables())

    with variable_scope.variable_scope("test", reuse=True):
      y1, y2 = rev_block_lib.rev_block(x1, x2, f, g, num_layers=self.NUM_LAYERS)

    num_vars_after = len(variables.global_variables())
    self.assertEqual(num_vars_before, num_vars_after) 

def ssim(img1, img2, max_val):
    _, _, checks = _verify_compatible_image_shapes(img1, img2)
    with ops.control_dependencies(checks):
        img1 = array_ops.identity(img1)

    # Need to convert the images to float32.  Scale max_val accordingly so that
    # SSIM is computed correctly.
    max_val = math_ops.cast(max_val, img1.dtype)
    max_val = tf.image.convert_image_dtype(max_val, dtypes.float32)
    img1 = tf.image.convert_image_dtype(img1, dtypes.float32)
    img2 = tf.image.convert_image_dtype(img2, dtypes.float32)
    ssim_per_channel, cs, luminance = _ssim_per_channel(img1, img2, max_val)
    # Compute average over color channels.
    return math_ops.reduce_mean(ssim_per_channel, [-1]), math_ops.reduce_mean(cs, [-1]), math_ops.reduce_mean(luminance, [-1]) 

def _ssim_per_channel(img1, img2, max_val=1.0):
    filter_size = constant_op.constant(11, dtype=dtypes.int32)
    filter_sigma = constant_op.constant(1.5, dtype=img1.dtype)

    shape1, shape2 = array_ops.shape_n([img1, img2])
    checks = [
      control_flow_ops.Assert(math_ops.reduce_all(math_ops.greater_equal(
          shape1[-3:-1], filter_size)), [shape1, filter_size], summarize=8),
      control_flow_ops.Assert(math_ops.reduce_all(math_ops.greater_equal(
          shape2[-3:-1], filter_size)), [shape2, filter_size], summarize=8)]

    # Enforce the check to run before computation.
    with ops.control_dependencies(checks):
        img1 = array_ops.identity(img1)

    kernel = _fspecial_gauss(filter_size, filter_sigma)
    kernel = array_ops.tile(kernel, multiples=[1, 1, shape1[-1], 1])

    compensation = 1.0

    def reducer(x):
        shape = array_ops.shape(x)
        x = array_ops.reshape(x, shape=array_ops.concat([[-1], shape[-3:]], 0))
        y = nn.depthwise_conv2d(x, kernel, strides=[1, 1, 1, 1], padding='VALID')
        return array_ops.reshape(y, array_ops.concat([shape[:-3],
                                                      array_ops.shape(y)[1:]], 0))

    luminance, cs = _ssim_helper(img1, img2, reducer, max_val, compensation)

    # Average over the second and the third from the last: height, width.
    axes = constant_op.constant([-3, -2], dtype=dtypes.int32)
    ssim_val = math_ops.reduce_mean(luminance * cs, axes)
    cs = math_ops.reduce_mean(cs, axes)
    luminance = math_ops.reduce_mean(luminance, axes)
    return ssim_val, cs, luminance