Merge branch 'tensorflow:master' into master

official/vision/beta/projects/yolo/train.py

@@ -18,7 +18,7 @@ from absl import app
 from absl import flags
 
 from official.common import flags as tfm_flags
-from official.vision.beta import train
+from official.vision import train
 from official.vision.beta.projects.yolo.common import registry_imports  # pylint: disable=unused-import
 
 FLAGS = flags.FLAGS

official/vision/data/create_coco_tf_record.py

@@ -112,7 +112,7 @@ def generate_coco_panoptics_masks(segments_info, mask_path,
       represent "stuff" and "things" classes respectively.
 
   Returns:
-    A dict with with keys: [u'semantic_segmentation_mask', u'category_mask',
+    A dict with keys: [u'semantic_segmentation_mask', u'category_mask',
       u'instance_mask']. The dict contains 'category_mask' and 'instance_mask'
       only if `include_panoptic_eval_masks` is set to True.
   """

official/vision/examples/starter/example_task.py

@@ -138,7 +138,7 @@ class ExampleTask(base_task.Task):
     between output from Parser and input used here.
 
     Args:
-      inputs: A tuple of of input tensors of (features, labels).
+      inputs: A tuple of input tensors of (features, labels).
       model: A tf.keras.Model instance.
       optimizer: The optimizer for this training step.
       metrics: A nested structure of metrics objects.
@@ -186,7 +186,7 @@ class ExampleTask(base_task.Task):
     """Runs validatation step.
 
     Args:
-      inputs: A tuple of of input tensors of (features, labels).
+      inputs: A tuple of input tensors of (features, labels).
       model: A tf.keras.Model instance.
       metrics: A nested structure of metrics objects.
 

official/vision/modeling/decoders/nasfpn.py

@@ -135,25 +135,6 @@ class NASFPN(tf.keras.Model):
     self._conv_op = (tf.keras.layers.SeparableConv2D
                      if self._config_dict['use_separable_conv']
                      else tf.keras.layers.Conv2D)
-    if self._config_dict['use_separable_conv']:
-      self._conv_kwargs = {
-          'depthwise_initializer': tf.keras.initializers.VarianceScaling(
-              scale=2, mode='fan_out', distribution='untruncated_normal'),
-          'pointwise_initializer': tf.keras.initializers.VarianceScaling(
-              scale=2, mode='fan_out', distribution='untruncated_normal'),
-          'bias_initializer': tf.zeros_initializer(),
-          'depthwise_regularizer': self._config_dict['kernel_regularizer'],
-          'pointwise_regularizer': self._config_dict['kernel_regularizer'],
-          'bias_regularizer': self._config_dict['bias_regularizer'],
-      }
-    else:
-      self._conv_kwargs = {
-          'kernel_initializer': tf.keras.initializers.VarianceScaling(
-              scale=2, mode='fan_out', distribution='untruncated_normal'),
-          'bias_initializer': tf.zeros_initializer(),
-          'kernel_regularizer': self._config_dict['kernel_regularizer'],
-          'bias_regularizer': self._config_dict['bias_regularizer'],
-      }
     self._norm_op = (tf.keras.layers.experimental.SyncBatchNormalization
                      if self._config_dict['use_sync_bn']
                      else tf.keras.layers.BatchNormalization)
@@ -240,6 +221,28 @@ class NASFPN(tf.keras.Model):
     else:
       return x
 
+  @property
+  def _conv_kwargs(self):
+    if self._config_dict['use_separable_conv']:
+      return {
+          'depthwise_initializer': tf.keras.initializers.VarianceScaling(
+              scale=2, mode='fan_out', distribution='untruncated_normal'),
+          'pointwise_initializer': tf.keras.initializers.VarianceScaling(
+              scale=2, mode='fan_out', distribution='untruncated_normal'),
+          'bias_initializer': tf.zeros_initializer(),
+          'depthwise_regularizer': self._config_dict['kernel_regularizer'],
+          'pointwise_regularizer': self._config_dict['kernel_regularizer'],
+          'bias_regularizer': self._config_dict['bias_regularizer'],
+      }
+    else:
+      return {
+          'kernel_initializer': tf.keras.initializers.VarianceScaling(
+              scale=2, mode='fan_out', distribution='untruncated_normal'),
+          'bias_initializer': tf.zeros_initializer(),
+          'kernel_regularizer': self._config_dict['kernel_regularizer'],
+          'bias_regularizer': self._config_dict['bias_regularizer'],
+      }
+
   def _global_attention(self, feat0, feat1):
     m = tf.math.reduce_max(feat0, axis=[1, 2], keepdims=True)
     m = tf.math.sigmoid(m)

official/vision/modeling/heads/dense_prediction_heads.py

@@ -140,6 +140,9 @@ class RetinaNetHead(tf.keras.layers.Layer):
       for i in range(self._config_dict['num_convs']):
         if level == self._config_dict['min_level']:
           cls_conv_name = 'classnet-conv_{}'.format(i)
+          if 'kernel_initializer' in conv_kwargs:
+            conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+                conv_kwargs['kernel_initializer'])
           self._cls_convs.append(conv_op(name=cls_conv_name, **conv_kwargs))
         cls_norm_name = 'classnet-conv-norm_{}_{}'.format(level, i)
         this_level_cls_norms.append(bn_op(name=cls_norm_name, **bn_kwargs))
@@ -170,6 +173,9 @@ class RetinaNetHead(tf.keras.layers.Layer):
       for i in range(self._config_dict['num_convs']):
         if level == self._config_dict['min_level']:
           box_conv_name = 'boxnet-conv_{}'.format(i)
+          if 'kernel_initializer' in conv_kwargs:
+            conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+                conv_kwargs['kernel_initializer'])
           self._box_convs.append(conv_op(name=box_conv_name, **conv_kwargs))
         box_norm_name = 'boxnet-conv-norm_{}_{}'.format(level, i)
         this_level_box_norms.append(bn_op(name=box_norm_name, **bn_kwargs))
@@ -211,6 +217,9 @@ class RetinaNetHead(tf.keras.layers.Layer):
           for i in range(self._config_dict['num_convs']):
             if level == self._config_dict['min_level']:
               att_conv_name = '{}-conv_{}'.format(att_name, i)
+              if 'kernel_initializer' in conv_kwargs:
+                conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+                    conv_kwargs['kernel_initializer'])
               att_convs_i.append(conv_op(name=att_conv_name, **conv_kwargs))
             att_norm_name = '{}-conv-norm_{}_{}'.format(att_name, level, i)
             this_level_att_norms.append(bn_op(name=att_norm_name, **bn_kwargs))
@@ -436,6 +445,9 @@ class RPNHead(tf.keras.layers.Layer):
       for i in range(self._config_dict['num_convs']):
         if level == self._config_dict['min_level']:
           conv_name = 'rpn-conv_{}'.format(i)
+          if 'kernel_initializer' in conv_kwargs:
+            conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+                conv_kwargs['kernel_initializer'])
           self._convs.append(conv_op(name=conv_name, **conv_kwargs))
         norm_name = 'rpn-conv-norm_{}_{}'.format(level, i)
         this_level_norms.append(bn_op(name=norm_name, **bn_kwargs))

official/vision/modeling/heads/instance_heads.py

@@ -133,6 +133,9 @@ class DetectionHead(tf.keras.layers.Layer):
     self._conv_norms = []
     for i in range(self._config_dict['num_convs']):
       conv_name = 'detection-conv_{}'.format(i)
+      if 'kernel_initializer' in conv_kwargs:
+        conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+            conv_kwargs['kernel_initializer'])
       self._convs.append(conv_op(name=conv_name, **conv_kwargs))
       bn_name = 'detection-conv-bn_{}'.format(i)
       self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))
@@ -324,6 +327,11 @@ class MaskHead(tf.keras.layers.Layer):
     self._conv_norms = []
     for i in range(self._config_dict['num_convs']):
       conv_name = 'mask-conv_{}'.format(i)
+      for initializer_name in ['kernel_initializer', 'depthwise_initializer',
+                               'pointwise_initializer']:
+        if initializer_name in conv_kwargs:
+          conv_kwargs[initializer_name] = tf_utils.clone_initializer(
+              conv_kwargs[initializer_name])
       self._convs.append(conv_op(name=conv_name, **conv_kwargs))
       bn_name = 'mask-conv-bn_{}'.format(i)
       self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))

official/vision/modeling/heads/segmentation_heads.py

@@ -118,6 +118,9 @@ class MaskScoring(tf.keras.Model):
     self._conv_norms = []
     for i in range(self._config_dict['num_convs']):
       conv_name = 'mask-scoring_{}'.format(i)
+      if 'kernel_initializer' in conv_kwargs:
+        conv_kwargs['kernel_initializer'] = tf_utils.clone_initializer(
+            conv_kwargs['kernel_initializer'])
       self._convs.append(conv_op(name=conv_name, **conv_kwargs))
       bn_name = 'mask-scoring-bn_{}'.format(i)
       self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))
@@ -297,15 +300,7 @@ class SegmentationHead(tf.keras.layers.Layer):
   def build(self, input_shape: Union[tf.TensorShape, List[tf.TensorShape]]):
     """Creates the variables of the segmentation head."""
     use_depthwise_convolution = self._config_dict['use_depthwise_convolution']
-    random_initializer = tf.keras.initializers.RandomNormal(stddev=0.01)
     conv_op = tf.keras.layers.Conv2D
-    conv_kwargs = {
-        'kernel_size': 3 if not use_depthwise_convolution else 1,
-        'padding': 'same',
-        'use_bias': False,
-        'kernel_initializer': random_initializer,
-        '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)
@@ -352,7 +347,8 @@ class SegmentationHead(tf.keras.layers.Layer):
                 kernel_size=3,
                 padding='same',
                 use_bias=False,
-                depthwise_initializer=random_initializer,
+                depthwise_initializer=tf.keras.initializers.RandomNormal(
+                    stddev=0.01),
                 depthwise_regularizer=self._config_dict['kernel_regularizer'],
                 depth_multiplier=1))
         norm_name = 'segmentation_head_depthwise_norm_{}'.format(i)
@@ -362,7 +358,12 @@ class SegmentationHead(tf.keras.layers.Layer):
           conv_op(
               name=conv_name,
               filters=self._config_dict['num_filters'],
-              **conv_kwargs))
+              kernel_size=3 if not use_depthwise_convolution else 1,
+              padding='same',
+              use_bias=False,
+              kernel_initializer=tf.keras.initializers.RandomNormal(
+                  stddev=0.01),
+              kernel_regularizer=self._config_dict['kernel_regularizer']))
       norm_name = 'segmentation_head_norm_{}'.format(i)
       self._norms.append(bn_op(name=norm_name, **bn_kwargs))
 

official/vision/modeling/layers/deeplab.py

@@ -16,6 +16,8 @@
 
 import tensorflow as tf
 
+from official.modeling import tf_utils
+
 
 class SpatialPyramidPooling(tf.keras.layers.Layer):
   """Implements the Atrous Spatial Pyramid Pooling.
@@ -103,8 +105,10 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
 
     conv_sequential = tf.keras.Sequential([
         tf.keras.layers.Conv2D(
-            filters=self.output_channels, kernel_size=(1, 1),
-            kernel_initializer=self.kernel_initializer,
+            filters=self.output_channels,
+            kernel_size=(1, 1),
+            kernel_initializer=tf_utils.clone_initializer(
+                self.kernel_initializer),
             kernel_regularizer=self.kernel_regularizer,
             use_bias=False),
         bn_op(
@@ -121,21 +125,32 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
       if self.use_depthwise_convolution:
         leading_layers += [
             tf.keras.layers.DepthwiseConv2D(
-                depth_multiplier=1, kernel_size=kernel_size,
-                padding='same', depthwise_regularizer=self.kernel_regularizer,
-                depthwise_initializer=self.kernel_initializer,
-                dilation_rate=dilation_rate, use_bias=False)
+                depth_multiplier=1,
+                kernel_size=kernel_size,
+                padding='same',
+                depthwise_regularizer=self.kernel_regularizer,
+                depthwise_initializer=tf_utils.clone_initializer(
+                    self.kernel_initializer),
+                dilation_rate=dilation_rate,
+                use_bias=False)
         ]
         kernel_size = (1, 1)
       conv_sequential = tf.keras.Sequential(leading_layers + [
           tf.keras.layers.Conv2D(
-              filters=self.output_channels, kernel_size=kernel_size,
-              padding='same', kernel_regularizer=self.kernel_regularizer,
-              kernel_initializer=self.kernel_initializer,
-              dilation_rate=dilation_rate, use_bias=False),
-          bn_op(axis=bn_axis, momentum=self.batchnorm_momentum,
-                epsilon=self.batchnorm_epsilon),
-          tf.keras.layers.Activation(self.activation)])
+              filters=self.output_channels,
+              kernel_size=kernel_size,
+              padding='same',
+              kernel_regularizer=self.kernel_regularizer,
+              kernel_initializer=tf_utils.clone_initializer(
+                  self.kernel_initializer),
+              dilation_rate=dilation_rate,
+              use_bias=False),
+          bn_op(
+              axis=bn_axis,
+              momentum=self.batchnorm_momentum,
+              epsilon=self.batchnorm_epsilon),
+          tf.keras.layers.Activation(self.activation)
+      ])
       self.aspp_layers.append(conv_sequential)
 
     if self.pool_kernel_size is None:
@@ -151,7 +166,8 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
             tf.keras.layers.Conv2D(
                 filters=self.output_channels,
                 kernel_size=(1, 1),
-                kernel_initializer=self.kernel_initializer,
+                kernel_initializer=tf_utils.clone_initializer(
+                    self.kernel_initializer),
                 kernel_regularizer=self.kernel_regularizer,
                 use_bias=False),
             bn_op(
@@ -170,8 +186,10 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
 
     self.projection = tf.keras.Sequential([
         tf.keras.layers.Conv2D(
-            filters=self.output_channels, kernel_size=(1, 1),
-            kernel_initializer=self.kernel_initializer,
+            filters=self.output_channels,
+            kernel_size=(1, 1),
+            kernel_initializer=tf_utils.clone_initializer(
+                self.kernel_initializer),
             kernel_regularizer=self.kernel_regularizer,
             use_bias=False),
         bn_op(
@@ -179,7 +197,8 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
             momentum=self.batchnorm_momentum,
             epsilon=self.batchnorm_epsilon),
         tf.keras.layers.Activation(self.activation),
-        tf.keras.layers.Dropout(rate=self.dropout)])
+        tf.keras.layers.Dropout(rate=self.dropout)
+    ])
 
   def call(self, inputs, training=None):
     if training is None:

official/vision/modeling/layers/nn_blocks.py

@@ -141,7 +141,8 @@ class ResidualBlock(tf.keras.layers.Layer):
           kernel_size=1,
           strides=self._strides,
           use_bias=False,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer)
       self._norm0 = self._norm(
@@ -162,7 +163,7 @@ class ResidualBlock(tf.keras.layers.Layer):
         strides=self._strides,
         padding=conv1_padding,
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm1 = self._norm(
@@ -177,7 +178,7 @@ class ResidualBlock(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm2 = self._norm(
@@ -191,7 +192,8 @@ class ResidualBlock(tf.keras.layers.Layer):
           in_filters=self._filters,
           out_filters=self._filters,
           se_ratio=self._se_ratio,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer)
     else:
@@ -338,7 +340,8 @@ class BottleneckBlock(tf.keras.layers.Layer):
             kernel_size=1,
             strides=1,
             use_bias=False,
-            kernel_initializer=self._kernel_initializer,
+            kernel_initializer=tf_utils.clone_initializer(
+                self._kernel_initializer),
             kernel_regularizer=self._kernel_regularizer,
             bias_regularizer=self._bias_regularizer)
       else:
@@ -347,7 +350,8 @@ class BottleneckBlock(tf.keras.layers.Layer):
             kernel_size=1,
             strides=self._strides,
             use_bias=False,
-            kernel_initializer=self._kernel_initializer,
+            kernel_initializer=tf_utils.clone_initializer(
+                self._kernel_initializer),
             kernel_regularizer=self._kernel_regularizer,
             bias_regularizer=self._bias_regularizer)
 
@@ -362,7 +366,7 @@ class BottleneckBlock(tf.keras.layers.Layer):
         kernel_size=1,
         strides=1,
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm1 = self._norm(
@@ -380,7 +384,7 @@ class BottleneckBlock(tf.keras.layers.Layer):
         dilation_rate=self._dilation_rate,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm2 = self._norm(
@@ -396,7 +400,7 @@ class BottleneckBlock(tf.keras.layers.Layer):
         kernel_size=1,
         strides=1,
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm3 = self._norm(
@@ -412,7 +416,8 @@ class BottleneckBlock(tf.keras.layers.Layer):
           in_filters=self._filters * 4,
           out_filters=self._filters * 4,
           se_ratio=self._se_ratio,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer)
     else:
@@ -616,7 +621,8 @@ class InvertedBottleneckBlock(tf.keras.layers.Layer):
           strides=expand_stride,
           padding='same',
           use_bias=False,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer)
       self._norm0 = self._norm(
@@ -635,7 +641,8 @@ class InvertedBottleneckBlock(tf.keras.layers.Layer):
           depth_multiplier=1,
           dilation_rate=self._dilation_rate,
           use_bias=False,
-          depthwise_initializer=self._kernel_initializer,
+          depthwise_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           depthwise_regularizer=self._depthsize_regularizer,
           bias_regularizer=self._bias_regularizer)
       self._norm1 = self._norm(
@@ -657,7 +664,8 @@ class InvertedBottleneckBlock(tf.keras.layers.Layer):
           se_ratio=self._se_ratio,
           divisible_by=self._divisible_by,
           round_down_protect=self._se_round_down_protect,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer,
           activation=self._se_inner_activation,
@@ -672,7 +680,7 @@ class InvertedBottleneckBlock(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm2 = self._norm(
@@ -829,7 +837,7 @@ class ResidualInner(tf.keras.layers.Layer):
         strides=self.strides,
         use_bias=False,
         padding='same',
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
 
     self._batch_norm_1 = self._norm(
@@ -843,7 +851,7 @@ class ResidualInner(tf.keras.layers.Layer):
         strides=1,
         use_bias=False,
         padding='same',
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
 
     super(ResidualInner, self).build(input_shape)
@@ -954,7 +962,7 @@ class BottleneckResidualInner(tf.keras.layers.Layer):
         strides=self.strides,
         use_bias=False,
         padding='same',
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
     self._batch_norm_1 = self._norm(
         axis=self._bn_axis,
@@ -966,7 +974,7 @@ class BottleneckResidualInner(tf.keras.layers.Layer):
         strides=1,
         use_bias=False,
         padding='same',
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
     self._batch_norm_2 = self._norm(
         axis=self._bn_axis,
@@ -978,7 +986,7 @@ class BottleneckResidualInner(tf.keras.layers.Layer):
         strides=1,
         use_bias=False,
         padding='same',
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
 
     super(BottleneckResidualInner, self).build(input_shape)
@@ -1286,7 +1294,7 @@ class DepthwiseSeparableConvBlock(tf.keras.layers.Layer):
         padding='same',
         depth_multiplier=1,
         dilation_rate=self._dilation_rate,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._depthsize_regularizer,
         use_bias=False)
     self._norm0 = self._norm(
@@ -1300,7 +1308,7 @@ class DepthwiseSeparableConvBlock(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer)
     self._norm1 = self._norm(
         axis=self._bn_axis,
@@ -1411,7 +1419,7 @@ class TuckerConvBlock(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm0 = self._norm(
@@ -1432,7 +1440,7 @@ class TuckerConvBlock(tf.keras.layers.Layer):
         strides=self._strides,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm1 = self._norm(
@@ -1449,7 +1457,7 @@ class TuckerConvBlock(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm2 = self._norm(

official/vision/modeling/layers/nn_blocks_3d.py

@@ -155,7 +155,7 @@ class BottleneckBlock3D(tf.keras.layers.Layer):
             self._temporal_strides, self._spatial_strides, self._spatial_strides
         ],
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm0 = self._norm(
@@ -169,7 +169,7 @@ class BottleneckBlock3D(tf.keras.layers.Layer):
         strides=[self._temporal_strides, 1, 1],
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm1 = self._norm(
@@ -183,7 +183,7 @@ class BottleneckBlock3D(tf.keras.layers.Layer):
         strides=[1, self._spatial_strides, self._spatial_strides],
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm2 = self._norm(
@@ -197,7 +197,7 @@ class BottleneckBlock3D(tf.keras.layers.Layer):
         strides=[1, 1, 1],
         padding='same',
         use_bias=False,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
     self._norm3 = self._norm(
@@ -211,7 +211,8 @@ class BottleneckBlock3D(tf.keras.layers.Layer):
           out_filters=self._filters * 4,
           se_ratio=self._se_ratio,
           use_3d_input=True,
-          kernel_initializer=self._kernel_initializer,
+          kernel_initializer=tf_utils.clone_initializer(
+              self._kernel_initializer),
           kernel_regularizer=self._kernel_regularizer,
           bias_regularizer=self._bias_regularizer)
     else:

official/vision/modeling/layers/nn_layers.py

@@ -163,7 +163,7 @@ class SqueezeExcitation(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=True,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
 
@@ -173,7 +173,7 @@ class SqueezeExcitation(tf.keras.layers.Layer):
         strides=1,
         padding='same',
         use_bias=True,
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         bias_regularizer=self._bias_regularizer)
 
@@ -1155,7 +1155,7 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
     conv1 = tf.keras.layers.Conv2D(
         filters=self._output_channels,
         kernel_size=(1, 1),
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         use_bias=False)
     norm1 = self._bn_op(
@@ -1175,7 +1175,8 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
                 kernel_size=kernel_size,
                 padding='same',
                 depthwise_regularizer=self._kernel_regularizer,
-                depthwise_initializer=self._kernel_initializer,
+                depthwise_initializer=tf_utils.clone_initializer(
+                    self._kernel_initializer),
                 dilation_rate=dilation_rate,
                 use_bias=False)
         ]
@@ -1186,7 +1187,8 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
               kernel_size=kernel_size,
               padding='same',
               kernel_regularizer=self._kernel_regularizer,
-              kernel_initializer=self._kernel_initializer,
+              kernel_initializer=tf_utils.clone_initializer(
+                  self._kernel_initializer),
               dilation_rate=dilation_rate,
               use_bias=False)
       ]
@@ -1208,7 +1210,7 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
     conv2 = tf.keras.layers.Conv2D(
         filters=self._output_channels,
         kernel_size=(1, 1),
-        kernel_initializer=self._kernel_initializer,
+        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
         kernel_regularizer=self._kernel_regularizer,
         use_bias=False)
     norm2 = self._bn_op(
@@ -1225,7 +1227,8 @@ class SpatialPyramidPooling(tf.keras.layers.Layer):
         tf.keras.layers.Conv2D(
             filters=self._output_channels,
             kernel_size=(1, 1),
-            kernel_initializer=self._kernel_initializer,
+            kernel_initializer=tf_utils.clone_initializer(
+                self._kernel_initializer),
             kernel_regularizer=self._kernel_regularizer,
             use_bias=False),
         self._bn_op(

official/vision/ops/box_ops.py

@@ -616,7 +616,7 @@ def bbox_overlap(boxes, gt_boxes):
         tf.transpose(gt_invalid_mask, [0, 2, 1]))
     iou = tf.where(padding_mask, -tf.ones_like(iou), iou)
 
-    # Fills -1 for for invalid (-1) boxes.
+    # Fills -1 for invalid (-1) boxes.
     boxes_invalid_mask = tf.less(
         tf.reduce_max(boxes, axis=-1, keepdims=True), 0.0)
     iou = tf.where(boxes_invalid_mask, -tf.ones_like(iou), iou)

official/vision/tasks/image_classification.py

@@ -218,7 +218,7 @@ class ImageClassificationTask(base_task.Task):
     """Does forward and backward.
 
     Args:
-      inputs: A tuple of of input tensors of (features, labels).
+      inputs: A tuple of input tensors of (features, labels).
       model: A tf.keras.Model instance.
       optimizer: The optimizer for this training step.
       metrics: A nested structure of metrics objects.
@@ -278,7 +278,7 @@ class ImageClassificationTask(base_task.Task):
     """Runs validatation step.
 
     Args:
-      inputs: A tuple of of input tensors of (features, labels).
+      inputs: A tuple of input tensors of (features, labels).
       model: A tf.keras.Model instance.
       metrics: A nested structure of metrics objects.
 

official/vision/utils/object_detection/balanced_positive_negative_sampler.py

@@ -64,7 +64,7 @@ class BalancedPositiveNegativeSampler(minibatch_sampler.MinibatchSampler):
       sorted_indices_tensor: A sorted int32 tensor of shape [N] which contains
         the signed indices of the examples where the sign is based on the label
         value. The examples that cannot be sampled are set to 0. It samples
-        atmost sample_size*positive_fraction positive examples and remaining
+        at most sample_size*positive_fraction positive examples and remaining
         from negative examples.
       sample_size: Size of subsamples.