Added `PanopticDeepLabFusion` layer

official/vision/beta/modeling/layers/nn_layers.py

@@ -320,6 +320,136 @@ def pyramid_feature_fusion(inputs, target_level):
 
   return tf.math.add_n(resampled_feats)
 
+class PanopticDeepLabFusion(tf.keras.layers.Layer):
+  """Creates a Panoptic DeepLab feature Fusion layer.
+
+  This implements the feature fusion introduced in the paper:
+  Cheng et al. Panoptic-DeepLab
+  (https://arxiv.org/pdf/1911.10194.pdf)
+  """
+  def __init__(
+      self,
+      level: int,
+      low_level: List[int] = [3, 2],
+      num_projection_filters: List[int] = [64, 32],
+      num_output_filters: int = 256,
+      activation: str = 'relu',
+      use_sync_bn: bool = False,
+      norm_momentum: float = 0.99,
+      norm_epsilon: float = 0.001,
+      kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
+      bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
+      interpolation: str = 'bilinear',
+      **kwargs):
+
+    """Initializes panoptic FPN feature fusion layer.
+
+    Args:
+      level: An `int` level at which the decoder was appled at.
+      low_level: A list of `int` of minimum level to use in feature fusion.
+      num_filters: An `int` number of filters in conv2d layers.
+      activation: A `str` name of the activation function.
+      use_sync_bn: A `bool` that indicates whether to use synchronized batch
+        normalization across different replicas.
+      norm_momentum: A `float` of normalization momentum for the moving average.
+      norm_epsilon: A `float` added to variance to avoid dividing by zero.      
+      kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
+        Conv2D. Default is None.
+      bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
+      interpolation: A `str` interpolation method for upsampling. Defaults to
+        `bilinear`.      
+      **kwargs: Additional keyword arguments to be passed.
+    Returns:
+      A `float` `tf.Tensor` of shape [batch_size, feature_height, feature_width,
+        feature_channel].
+    """
+    super(PanopticDeepLabFusion, self).__init__(**kwargs)
+
+    self._config_dict = {
+        'level': level,
+        'low_level': low_level,
+        'num_projection_filters': num_projection_filters,
+        'num_output_filters': num_output_filters,
+        'activation': activation,
+        'use_sync_bn': use_sync_bn,
+        'norm_momentum': norm_momentum,
+        'norm_epsilon': norm_epsilon,
+        'kernel_regularizer': kernel_regularizer,
+        'bias_regularizer': bias_regularizer,
+        'interpolation': interpolation
+    }
+    if tf.keras.backend.image_data_format() == 'channels_last':
+      self._channel_axis = -1
+    else:
+      self._channel_axis = 1
+    self._activation = tf_utils.get_activation(activation)
+
+  def build(self, input_shape: List[tf.TensorShape]):
+    conv_op = tf.keras.layers.Conv2D
+    conv_kwargs = {
+        'padding': 'same',
+        'use_bias': False,
+        'kernel_initializer': tf.initializers.VarianceScaling(),
+        'kernel_regularizer': self._config_dict['kernel_regularizer'],
+    }
+    bn_op = (tf.keras.layers.experimental.SyncBatchNormalization
+             if self._config_dict['use_sync_bn']
+             else tf.keras.layers.BatchNormalization)
+    bn_kwargs = {
+        'axis': self._channel_axis,
+        'momentum': self._config_dict['norm_momentum'],
+        'epsilon': self._config_dict['norm_epsilon'],
+    }
+
+    self._projection_convs = []
+    self._projection_norms = []
+    self._fusion_convs = []
+    self._fusion_norms = []
+    for i in range(len(self._config_dict['low_level'])):
+      self._projection_convs.append(
+          conv_op(
+              filters=self._config_dict['num_projection_filters'][i],
+              kernel_size=1,
+              **conv_kwargs))
+      self._fusion_convs.append(
+          conv_op(
+              filters=self._config_dict['num_output_filters'],
+              kernel_size=5,
+              **conv_kwargs))
+      self._projection_norms.append(bn_op(**bn_kwargs))
+      self._fusion_norms.append(bn_op(**bn_kwargs))
+
+  def call(self, inputs, training=None):
+    if training is None:
+      training = tf.keras.backend.learning_phase()
+
+    backbone_output = inputs[0]
+    decoder_output = inputs[1][str(self._config_dict['level'])]
+
+    x = decoder_output
+    for i in range(len(self._config_dict['low_level'])):
+      feature = backbone_output[str(self._config_dict['low_level'][i])]
+      feature = self._projection_convs[i](feature)
+      feature = self._projection_norms[i](feature, training=training)
+      feature = self._activation(feature)
+
+      shape = tf.shape(feature)
+      x = tf.image.resize(
+          x, size=[shape[1], shape[2]],
+          method=self._config_dict['interpolation'])
+      x = tf.concat([x, feature], axis=self._channel_axis)
+
+      x = self._fusion_convs[i](x)
+      x = self._fusion_norms[i](x, training=training)
+      x = self._activation(x)
+    return x
+
+  def get_config(self) -> Mapping[str, Any]:
+    return self._config_dict
+
+  @classmethod
+  def from_config(cls, config, custom_objects=None):
+    return cls(**config)
 
 class PanopticFPNFusion(tf.keras.Model):
   """Creates a Panoptic FPN feature Fusion layer.