Introducing SpineNet backbone to TF2.

PiperOrigin-RevId: 318945656

official/README.md

@@ -19,9 +19,10 @@ In the near future, we will add:
 
 * State-of-the-art language understanding models:
   More members in Transformer family
-* Start-of-the-art image classification models:
+* State-of-the-art image classification models:
   EfficientNet, MnasNet, and variants
-* A set of excellent objection detection models.
+* State-of-the-art objection detection and instance segmentation models:
+  RetinaNet, Mask R-CNN, SpineNet, and variants
 
 ## Table of Contents
 
@@ -52,6 +53,7 @@ In the near future, we will add:
 | [RetinaNet](vision/detection) | [Focal Loss for Dense Object Detection](https://arxiv.org/abs/1708.02002) |
 | [Mask R-CNN](vision/detection) | [Mask R-CNN](https://arxiv.org/abs/1703.06870) |
 | [ShapeMask](vision/detection) | [ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors](https://arxiv.org/abs/1904.03239) |
+| [SpineNet](vision/detection) | [SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization](https://arxiv.org/abs/1912.05027) |
 
 ### Natural Language Processing
 

official/vision/detection/README.md

@@ -48,6 +48,22 @@ so the checkpoints are not compatible.
 We will unify the implementation soon.
 
 
+### Train a SpineNet-49 based RetinaNet.
+
+```bash
+TPU_NAME="<your GCP TPU name>"
+MODEL_DIR="<path to the directory to store model files>"
+TRAIN_FILE_PATTERN="<path to the TFRecord training data>"
+EVAL_FILE_PATTERN="<path to the TFRecord validation data>"
+VAL_JSON_FILE="<path to the validation annotation JSON file>"
+python3 ~/models/official/vision/detection/main.py \
+  --strategy_type=tpu \
+  --tpu="${TPU_NAME?}" \
+  --model_dir="${MODEL_DIR?}" \
+  --mode=train \
+  --params_override="{ type: retinanet, architecture: {backbone: spinenet, multilevel_features: identity}, spinenet: {model_id: 49}, train_file_pattern: ${TRAIN_FILE_PATTERN?} }, eval: { val_json_file: ${VAL_JSON_FILE?}, eval_file_pattern: ${EVAL_FILE_PATTERN?} } }"
+```
+
 
 ### Train a custom RetinaNet using the config file.
 
@@ -163,6 +179,24 @@ so the checkpoints are not compatible.
 We will unify the implementation soon.
 
 
+### Train a SpineNet-49 based Mask R-CNN.
+
+```bash
+TPU_NAME="<your GCP TPU name>"
+MODEL_DIR="<path to the directory to store model files>"
+TRAIN_FILE_PATTERN="<path to the TFRecord training data>"
+EVAL_FILE_PATTERN="<path to the TFRecord validation data>"
+VAL_JSON_FILE="<path to the validation annotation JSON file>"
+python3 ~/models/official/vision/detection/main.py \
+  --strategy_type=tpu \
+  --tpu="${TPU_NAME?}" \
+  --model_dir="${MODEL_DIR?}" \
+  --mode=train \
+  --model=mask_rcnn \
+  --params_override="{architecture: {backbone: spinenet, multilevel_features: identity}, spinenet: {model_id: 49}, train_file_pattern: ${TRAIN_FILE_PATTERN?} }, eval: { val_json_file: ${VAL_JSON_FILE?}, eval_file_pattern: ${EVAL_FILE_PATTERN?} } }"
+```
+
+
 ### Train a custom Mask R-CNN using the config file.
 
 First, create a YAML config file, e.g. *my_maskrcnn.yaml*.

official/vision/detection/configs/base_config.py

@@ -17,10 +17,12 @@
 
 BACKBONES = [
     'resnet',
+    'spinenet',
 ]
 
 MULTILEVEL_FEATURES = [
     'fpn',
+    'identity',
 ]
 
 # pylint: disable=line-too-long
@@ -118,6 +120,9 @@ BASE_CFG = {
     'resnet': {
         'resnet_depth': 50,
     },
+    'spinenet': {
+        'model_id': '49',
+    },
     'fpn': {
         'fpn_feat_dims': 256,
         'use_separable_conv': False,

official/vision/detection/modeling/architecture/factory.py

@@ -23,6 +23,7 @@ from official.vision.detection.modeling.architecture import heads
 from official.vision.detection.modeling.architecture import identity
 from official.vision.detection.modeling.architecture import nn_ops
 from official.vision.detection.modeling.architecture import resnet
+from official.vision.detection.modeling.architecture import spinenet
 
 
 def norm_activation_generator(params):
@@ -42,6 +43,9 @@ def backbone_generator(params):
         activation=params.norm_activation.activation,
         norm_activation=norm_activation_generator(
             params.norm_activation))
+  elif params.architecture.backbone == 'spinenet':
+    spinenet_params = params.spinenet
+    backbone_fn = spinenet.SpineNetBuilder(model_id=spinenet_params.model_id)
   else:
     raise ValueError('Backbone model `{}` is not supported.'
                      .format(params.architecture.backbone))

official/vision/detection/modeling/architecture/nn_blocks.py

+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Contains common building blocks for neural networks."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+from official.modeling import tf_utils
+
+
+@tf.keras.utils.register_keras_serializable(package='Vision')
+class ResidualBlock(tf.keras.layers.Layer):
+  """A residual block."""
+
+  def __init__(self,
+               filters,
+               strides,
+               use_projection=False,
+               kernel_initializer='VarianceScaling',
+               kernel_regularizer=None,
+               bias_regularizer=None,
+               activation='relu',
+               use_sync_bn=False,
+               norm_momentum=0.99,
+               norm_epsilon=0.001,
+               **kwargs):
+    """A residual block with BN after convolutions.
+
+    Args:
+      filters: `int` number of filters for the first two convolutions. Note that
+        the third and final convolution will use 4 times as many filters.
+      strides: `int` block stride. If greater than 1, this block will ultimately
+        downsample the input.
+      use_projection: `bool` for whether this block should use a projection
+        shortcut (versus the default identity shortcut). This is usually `True`
+        for the first block of a block group, which may change the number of
+        filters and the resolution.
+      kernel_initializer: kernel_initializer for convolutional layers.
+      kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
+        Default to None.
+      bias_regularizer: tf.keras.regularizers.Regularizer object for Conv2d.
+        Default to None.
+      activation: `str` name of the activation function.
+      use_sync_bn: if True, use synchronized batch normalization.
+      norm_momentum: `float` normalization omentum for the moving average.
+      norm_epsilon: `float` small float added to variance to avoid dividing by
+        zero.
+      **kwargs: keyword arguments to be passed.
+    """
+    super(ResidualBlock, self).__init__(**kwargs)
+
+    self._filters = filters
+    self._strides = strides
+    self._use_projection = use_projection
+    self._use_sync_bn = use_sync_bn
+    self._activation = activation
+    self._kernel_initializer = kernel_initializer
+    self._norm_momentum = norm_momentum
+    self._norm_epsilon = norm_epsilon
+    self._kernel_regularizer = kernel_regularizer
+    self._bias_regularizer = bias_regularizer
+
+    if use_sync_bn:
+      self._norm = tf.keras.layers.experimental.SyncBatchNormalization
+    else:
+      self._norm = tf.keras.layers.BatchNormalization
+    if tf.keras.backend.image_data_format() == 'channels_last':
+      self._bn_axis = -1
+    else:
+      self._bn_axis = 1
+    self._activation_fn = tf_utils.get_activation(activation)
+
+  def build(self, input_shape):
+    if self._use_projection:
+      self._shortcut = tf.keras.layers.Conv2D(
+          filters=self._filters,
+          kernel_size=1,
+          strides=self._strides,
+          use_bias=False,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer)
+      self._norm0 = self._norm(
+          axis=self._bn_axis,
+          momentum=self._norm_momentum,
+          epsilon=self._norm_epsilon)
+
+    self._conv1 = tf.keras.layers.Conv2D(
+        filters=self._filters,
+        kernel_size=3,
+        strides=self._strides,
+        padding='same',
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)
+    self._norm1 = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)
+
+    self._conv2 = tf.keras.layers.Conv2D(
+        filters=self._filters,
+        kernel_size=3,
+        strides=1,
+        padding='same',
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)
+    self._norm2 = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)
+
+    super(ResidualBlock, self).build(input_shape)
+
+  def get_config(self):
+    config = {
+        'filters': self._filters,
+        'strides': self._strides,
+        'use_projection': self._use_projection,
+        'kernel_initializer': self._kernel_initializer,
+        'kernel_regularizer': self._kernel_regularizer,
+        'bias_regularizer': self._bias_regularizer,
+        'activation': self._activation,
+        'use_sync_bn': self._use_sync_bn,
+        'norm_momentum': self._norm_momentum,
+        'norm_epsilon': self._norm_epsilon
+    }
+
+    base_config = super(ResidualBlock, self).get_config()
+    return dict(list(base_config.items()) + list(config.items()))
+
+  def call(self, inputs):
+    shortcut = inputs
+    if self._use_projection:
+      shortcut = self._shortcut(shortcut)
+      shortcut = self._norm0(shortcut)
+
+    x = self._conv1(inputs)
+    x = self._norm1(x)
+    x = self._activation_fn(x)
+
+    x = self._conv2(x)
+    x = self._norm2(x)
+
+    return self._activation_fn(x + shortcut)
+
+
+@tf.keras.utils.register_keras_serializable(package='Vision')
+class BottleneckBlock(tf.keras.layers.Layer):
+  """A standard bottleneck block."""
+
+  def __init__(self,
+               filters,
+               strides,
+               use_projection=False,
+               kernel_initializer='VarianceScaling',
+               kernel_regularizer=None,
+               bias_regularizer=None,
+               activation='relu',
+               use_sync_bn=False,
+               norm_momentum=0.99,
+               norm_epsilon=0.001,
+               **kwargs):
+    """A standard bottleneck block with BN after convolutions.
+
+    Args:
+      filters: `int` number of filters for the first two convolutions. Note that
+        the third and final convolution will use 4 times as many filters.
+      strides: `int` block stride. If greater than 1, this block will ultimately
+        downsample the input.
+      use_projection: `bool` for whether this block should use a projection
+        shortcut (versus the default identity shortcut). This is usually `True`
+        for the first block of a block group, which may change the number of
+        filters and the resolution.
+      kernel_initializer: kernel_initializer for convolutional layers.
+      kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
+        Default to None.
+      bias_regularizer: tf.keras.regularizers.Regularizer object for Conv2d.
+        Default to None.
+      activation: `str` name of the activation function.
+      use_sync_bn: if True, use synchronized batch normalization.
+      norm_momentum: `float` normalization omentum for the moving average.
+      norm_epsilon: `float` small float added to variance to avoid dividing by
+        zero.
+      **kwargs: keyword arguments to be passed.
+    """
+    super(BottleneckBlock, self).__init__(**kwargs)
+
+    self._filters = filters
+    self._strides = strides
+    self._use_projection = use_projection
+    self._use_sync_bn = use_sync_bn
+    self._activation = activation
+    self._kernel_initializer = kernel_initializer
+    self._norm_momentum = norm_momentum
+    self._norm_epsilon = norm_epsilon
+    self._kernel_regularizer = kernel_regularizer
+    self._bias_regularizer = bias_regularizer
+    if use_sync_bn:
+      self._norm = tf.keras.layers.experimental.SyncBatchNormalization
+    else:
+      self._norm = tf.keras.layers.BatchNormalization
+    if tf.keras.backend.image_data_format() == 'channels_last':
+      self._bn_axis = -1
+    else:
+      self._bn_axis = 1
+    self._activation_fn = tf_utils.get_activation(activation)
+
+  def build(self, input_shape):
+    if self._use_projection:
+      self._shortcut = tf.keras.layers.Conv2D(
+          filters=self._filters * 4,
+          kernel_size=1,
+          strides=self._strides,
+          use_bias=False,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer)
+      self._norm0 = self._norm(
+          axis=self._bn_axis,
+          momentum=self._norm_momentum,
+          epsilon=self._norm_epsilon)
+
+    self._conv1 = tf.keras.layers.Conv2D(
+        filters=self._filters,
+        kernel_size=1,
+        strides=1,
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)
+    self._norm1 = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)
+
+    self._conv2 = tf.keras.layers.Conv2D(
+        filters=self._filters,
+        kernel_size=3,
+        strides=self._strides,
+        padding='same',
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)
+    self._norm2 = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)
+
+    self._conv3 = tf.keras.layers.Conv2D(
+        filters=self._filters * 4,
+        kernel_size=1,
+        strides=1,
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)
+    self._norm3 = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)
+
+    super(BottleneckBlock, self).build(input_shape)
+
+  def get_config(self):
+    config = {
+        'filters': self._filters,
+        'strides': self._strides,
+        'use_projection': self._use_projection,
+        'kernel_initializer': self._kernel_initializer,
+        'kernel_regularizer': self._kernel_regularizer,
+        'bias_regularizer': self._bias_regularizer,
+        'activation': self._activation,
+        'use_sync_bn': self._use_sync_bn,
+        'norm_momentum': self._norm_momentum,
+        'norm_epsilon': self._norm_epsilon
+    }
+
+    base_config = super(BottleneckBlock, self).get_config()
+    return dict(list(base_config.items()) + list(config.items()))
+
+  def call(self, inputs):
+    shortcut = inputs
+    if self._use_projection:
+      shortcut = self._shortcut(shortcut)
+      shortcut = self._norm0(shortcut)
+
+    x = self._conv1(inputs)
+    x = self._norm1(x)
+    x = self._activation_fn(x)
+
+    x = self._conv2(x)
+    x = self._norm2(x)
+    x = self._activation_fn(x)
+
+    x = self._conv3(x)
+    x = self._norm3(x)
+
+    return self._activation_fn(x + shortcut)

official/vision/detection/modeling/architecture/spinenet.py

+# Lint as: python3
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Implementation of SpineNet model.
+
+X. Du, T-Y. Lin, P. Jin, G. Ghiasi, M. Tan, Y. Cui, Q. V. Le, X. Song
+SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization
+https://arxiv.org/abs/1912.05027
+"""
+import math
+
+from absl import logging
+import tensorflow as tf
+
+from tensorflow.python.keras import backend
+from official.modeling import tf_utils
+from official.vision.detection.modeling.architecture import nn_blocks
+
+layers = tf.keras.layers
+
+FILTER_SIZE_MAP = {
+    1: 32,
+    2: 64,
+    3: 128,
+    4: 256,
+    5: 256,
+    6: 256,
+    7: 256,
+}
+
+# The fixed SpineNet architecture discovered by NAS.
+# Each element represents a specification of a building block:
+#   (block_level, block_fn, (input_offset0, input_offset1), is_output).
+SPINENET_BLOCK_SPECS = [
+    (2, 'bottleneck', (0, 1), False),
+    (4, 'residual', (0, 1), False),
+    (3, 'bottleneck', (2, 3), False),
+    (4, 'bottleneck', (2, 4), False),
+    (6, 'residual', (3, 5), False),
+    (4, 'bottleneck', (3, 5), False),
+    (5, 'residual', (6, 7), False),
+    (7, 'residual', (6, 8), False),
+    (5, 'bottleneck', (8, 9), False),
+    (5, 'bottleneck', (8, 10), False),
+    (4, 'bottleneck', (5, 10), True),
+    (3, 'bottleneck', (4, 10), True),
+    (5, 'bottleneck', (7, 12), True),
+    (7, 'bottleneck', (5, 14), True),
+    (6, 'bottleneck', (12, 14), True),
+]
+
+SCALING_MAP = {
+    '49S': {
+        'endpoints_num_filters': 128,
+        'filter_size_scale': 0.65,
+        'resample_alpha': 0.5,
+        'block_repeats': 1,
+    },
+    '49': {
+        'endpoints_num_filters': 256,
+        'filter_size_scale': 1.0,
+        'resample_alpha': 0.5,
+        'block_repeats': 1,
+    },
+    '96': {
+        'endpoints_num_filters': 256,
+        'filter_size_scale': 1.0,
+        'resample_alpha': 0.5,
+        'block_repeats': 2,
+    },
+    '143': {
+        'endpoints_num_filters': 256,
+        'filter_size_scale': 1.0,
+        'resample_alpha': 1.0,
+        'block_repeats': 3,
+    },
+    '190': {
+        'endpoints_num_filters': 512,
+        'filter_size_scale': 1.3,
+        'resample_alpha': 1.0,
+        'block_repeats': 4,
+    },
+}
+
+
+class BlockSpec(object):
+  """A container class that specifies the block configuration for SpineNet."""
+
+  def __init__(self, level, block_fn, input_offsets, is_output):
+    self.level = level
+    self.block_fn = block_fn
+    self.input_offsets = input_offsets
+    self.is_output = is_output
+
+
+def build_block_specs(block_specs=None):
+  """Builds the list of BlockSpec objects for SpineNet."""
+  if not block_specs:
+    block_specs = SPINENET_BLOCK_SPECS
+  logging.info('Building SpineNet block specs: %s', block_specs)
+  return [BlockSpec(*b) for b in block_specs]
+
+
+@tf.keras.utils.register_keras_serializable(package='Vision')
+class SpineNet(tf.keras.Model):
+  """Class to build SpineNet models."""
+
+  def __init__(self,
+               input_specs=tf.keras.layers.InputSpec(shape=[None, 640, 640, 3]),
+               min_level=3,
+               max_level=7,
+               block_specs=build_block_specs(),
+               endpoints_num_filters=256,
+               resample_alpha=0.5,
+               block_repeats=1,
+               filter_size_scale=1.0,
+               kernel_initializer='VarianceScaling',
+               kernel_regularizer=None,
+               bias_regularizer=None,
+               activation='relu',
+               use_sync_bn=False,
+               norm_momentum=0.99,
+               norm_epsilon=0.001,
+               **kwargs):
+    """SpineNet model."""
+    self._min_level = min_level
+    self._max_level = max_level
+    self._block_specs = block_specs
+    self._endpoints_num_filters = endpoints_num_filters
+    self._resample_alpha = resample_alpha
+    self._block_repeats = block_repeats
+    self._filter_size_scale = filter_size_scale
+    self._kernel_initializer = kernel_initializer
+    self._kernel_regularizer = kernel_regularizer
+    self._bias_regularizer = bias_regularizer
+    self._use_sync_bn = use_sync_bn
+    self._norm_momentum = norm_momentum
+    self._norm_epsilon = norm_epsilon
+    if activation == 'relu':
+      self._activation = tf.nn.relu
+    elif activation == 'swish':
+      self._activation = tf.nn.swish
+    else:
+      raise ValueError('Activation {} not implemented.'.format(activation))
+    self._init_block_fn = 'bottleneck'
+    self._num_init_blocks = 2
+
+    if use_sync_bn:
+      self._norm = layers.experimental.SyncBatchNormalization
+    else:
+      self._norm = layers.BatchNormalization
+
+    if tf.keras.backend.image_data_format() == 'channels_last':
+      self._bn_axis = -1
+    else:
+      self._bn_axis = 1
+
+    # Build SpineNet.
+    inputs = tf.keras.Input(shape=input_specs.shape[1:])
+
+    net = self._build_stem(inputs=inputs)
+    net = self._build_scale_permuted_network(
+        net=net, input_width=input_specs.shape[1])
+    net = self._build_endpoints(net=net)
+
+    super(SpineNet, self).__init__(inputs=inputs, outputs=net)
+
+  def _block_group(self,
+                   inputs,
+                   filters,
+                   strides,
+                   block_fn_cand,
+                   block_repeats=1,
+                   name='block_group'):
+    """Creates one group of blocks for the SpineNet model."""
+    block_fn_candidates = {
+        'bottleneck': nn_blocks.BottleneckBlock,
+        'residual': nn_blocks.ResidualBlock,
+    }
+    block_fn = block_fn_candidates[block_fn_cand]
+    _, _, _, num_filters = inputs.get_shape().as_list()
+
+    if block_fn_cand == 'bottleneck':
+      use_projection = not (num_filters == (filters * 4) and strides == 1)
+    else:
+      use_projection = not (num_filters == filters and strides == 1)
+
+    x = block_fn(
+        filters=filters,
+        strides=strides,
+        use_projection=use_projection,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer,
+        activation=self._activation,
+        use_sync_bn=self._use_sync_bn,
+        norm_momentum=self._norm_momentum,
+        norm_epsilon=self._norm_epsilon)(
+            inputs)
+    for _ in range(1, block_repeats):
+      x = block_fn(
+          filters=filters,
+          strides=1,
+          use_projection=False,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer,
+          activation=self._activation,
+          use_sync_bn=self._use_sync_bn,
+          norm_momentum=self._norm_momentum,
+          norm_epsilon=self._norm_epsilon)(
+              x)
+    return tf.identity(x, name=name)
+
+  def _build_stem(self, inputs):
+    """Build SpineNet stem."""
+    x = layers.Conv2D(
+        filters=64,
+        kernel_size=7,
+        strides=2,
+        use_bias=False,
+        padding='same',
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)(
+            inputs)
+    x = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)(
+            x)
+    x = tf_utils.get_activation(self._activation)(x)
+    x = layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)
+
+    net = []
+    # Build the initial level 2 blocks.
+    for i in range(self._num_init_blocks):
+      x = self._block_group(
+          inputs=x,
+          filters=int(FILTER_SIZE_MAP[2] * self._filter_size_scale),
+          strides=1,
+          block_fn_cand=self._init_block_fn,
+          block_repeats=self._block_repeats,
+          name='stem_block_{}'.format(i + 1))
+      net.append(x)
+    return net
+
+  def _build_scale_permuted_network(self,
+                                    net,
+                                    input_width,
+                                    weighted_fusion=False):
+    """Build scale-permuted network."""
+    net_sizes = [int(math.ceil(input_width / 2**2))] * len(net)
+    net_block_fns = [self._init_block_fn] * len(net)
+    num_outgoing_connections = [0] * len(net)
+
+    endpoints = {}
+    for i, block_spec in enumerate(self._block_specs):
+      # Find out specs for the target block.
+      target_width = int(math.ceil(input_width / 2**block_spec.level))
+      target_num_filters = int(FILTER_SIZE_MAP[block_spec.level] *
+                               self._filter_size_scale)
+      target_block_fn = block_spec.block_fn
+
+      # Resample then merge input0 and input1.
+      parents = []
+      input0 = block_spec.input_offsets[0]
+      input1 = block_spec.input_offsets[1]
+
+      x0 = self._resample_with_alpha(
+          inputs=net[input0],
+          input_width=net_sizes[input0],
+          input_block_fn=net_block_fns[input0],
+          target_width=target_width,
+          target_num_filters=target_num_filters,
+          target_block_fn=target_block_fn,
+          alpha=self._resample_alpha)
+      parents.append(x0)
+      num_outgoing_connections[input0] += 1
+
+      x1 = self._resample_with_alpha(
+          inputs=net[input1],
+          input_width=net_sizes[input1],
+          input_block_fn=net_block_fns[input1],
+          target_width=target_width,
+          target_num_filters=target_num_filters,
+          target_block_fn=target_block_fn,
+          alpha=self._resample_alpha)
+      parents.append(x1)
+      num_outgoing_connections[input1] += 1
+
+      # Merge 0 outdegree blocks to the output block.
+      if block_spec.is_output:
+        for j, (j_feat,
+                j_connections) in enumerate(zip(net, num_outgoing_connections)):
+          if j_connections == 0 and (j_feat.shape[2] == target_width and
+                                     j_feat.shape[3] == x0.shape[3]):
+            parents.append(j_feat)
+            num_outgoing_connections[j] += 1
+
+      # pylint: disable=g-direct-tensorflow-import
+      if weighted_fusion:
+        dtype = parents[0].dtype
+        parent_weights = [
+            tf.nn.relu(tf.cast(tf.Variable(1.0, name='block{}_fusion{}'.format(
+                i, j)), dtype=dtype)) for j in range(len(parents))]
+        weights_sum = tf.add_n(parent_weights)
+        parents = [
+            parents[i] * parent_weights[i] / (weights_sum + 0.0001)
+            for i in range(len(parents))
+        ]
+
+      # Fuse all parent nodes then build a new block.
+      x = tf_utils.get_activation(self._activation)(tf.add_n(parents))
+      x = self._block_group(
+          inputs=x,
+          filters=target_num_filters,
+          strides=1,
+          block_fn_cand=target_block_fn,
+          block_repeats=self._block_repeats,
+          name='scale_permuted_block_{}'.format(i + 1))
+
+      net.append(x)
+      net_sizes.append(target_width)
+      net_block_fns.append(target_block_fn)
+      num_outgoing_connections.append(0)
+
+      # Save output feats.
+      if block_spec.is_output:
+        if block_spec.level in endpoints:
+          raise ValueError('Duplicate feats found for output level {}.'.format(
+              block_spec.level))
+        if (block_spec.level < self._min_level or
+            block_spec.level > self._max_level):
+          raise ValueError('Output level is out of range [{}, {}]'.format(
+              self._min_level, self._max_level))
+        endpoints[block_spec.level] = x
+
+    return endpoints
+
+  def _build_endpoints(self, net):
+    """Match filter size for endpoints before sharing conv layers."""
+    endpoints = {}
+    for level in range(self._min_level, self._max_level + 1):
+      x = layers.Conv2D(
+          filters=self._endpoints_num_filters,
+          kernel_size=1,
+          strides=1,
+          use_bias=False,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer)(
+              net[level])
+      x = self._norm(
+          axis=self._bn_axis,
+          momentum=self._norm_momentum,
+          epsilon=self._norm_epsilon)(
+              x)
+      x = tf_utils.get_activation(self._activation)(x)
+      endpoints[level] = x
+    return endpoints
+
+  def _resample_with_alpha(self,
+                           inputs,
+                           input_width,
+                           input_block_fn,
+                           target_width,
+                           target_num_filters,
+                           target_block_fn,
+                           alpha=0.5):
+    """Match resolution and feature dimension."""
+    _, _, _, input_num_filters = inputs.get_shape().as_list()
+    if input_block_fn == 'bottleneck':
+      input_num_filters /= 4
+    new_num_filters = int(input_num_filters * alpha)
+
+    x = layers.Conv2D(
+        filters=new_num_filters,
+        kernel_size=1,
+        strides=1,
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)(
+            inputs)
+    x = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)(
+            x)
+    x = tf_utils.get_activation(self._activation)(x)
+
+    # Spatial resampling.
+    if input_width > target_width:
+      x = layers.Conv2D(
+          filters=new_num_filters,
+          kernel_size=3,
+          strides=2,
+          padding='SAME',
+          use_bias=False,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer)(
+              x)
+      x = self._norm(
+          axis=self._bn_axis,
+          momentum=self._norm_momentum,
+          epsilon=self._norm_epsilon)(
+              x)
+      x = tf_utils.get_activation(self._activation)(x)
+      input_width /= 2
+      while input_width > target_width:
+        x = layers.MaxPool2D(pool_size=3, strides=2, padding='SAME')(x)
+        input_width /= 2
+    elif input_width < target_width:
+      scale = target_width // input_width
+      x = layers.UpSampling2D(size=(scale, scale))(x)
+
+    # Last 1x1 conv to match filter size.
+    if target_block_fn == 'bottleneck':
+      target_num_filters *= 4
+    x = layers.Conv2D(
+        filters=target_num_filters,
+        kernel_size=1,
+        strides=1,
+        use_bias=False,
+        kernel_initializer=self._kernel_initializer,
+        kernel_regularizer=self._kernel_regularizer,
+        bias_regularizer=self._bias_regularizer)(
+            x)
+    x = self._norm(
+        axis=self._bn_axis,
+        momentum=self._norm_momentum,
+        epsilon=self._norm_epsilon)(
+            x)
+
+    return x
+
+
+class SpineNetBuilder(object):
+  """SpineNet builder."""
+
+  def __init__(self,
+               model_id,
+               input_specs=tf.keras.layers.InputSpec(shape=[None, 640, 640, 3]),
+               min_level=3,
+               max_level=7,
+               block_specs=build_block_specs(),
+               kernel_initializer='VarianceScaling',
+               kernel_regularizer=None,
+               bias_regularizer=None,
+               activation='relu',
+               use_sync_bn=False,
+               norm_momentum=0.99,
+               norm_epsilon=0.001):
+    if model_id not in SCALING_MAP:
+      raise ValueError(
+          'SpineNet {} is not a valid architecture.'.format(model_id))
+    scaling_params = SCALING_MAP[model_id]
+    self._input_specs = input_specs
+    self._min_level = min_level
+    self._max_level = max_level
+    self._block_specs = block_specs
+    self._endpoints_num_filters = scaling_params['endpoints_num_filters']
+    self._resample_alpha = scaling_params['resample_alpha']
+    self._block_repeats = scaling_params['block_repeats']
+    self._filter_size_scale = scaling_params['filter_size_scale']
+    self._kernel_initializer = kernel_initializer
+    self._kernel_regularizer = kernel_regularizer
+    self._bias_regularizer = bias_regularizer
+    self._activation = activation
+    self._use_sync_bn = use_sync_bn
+    self._norm_momentum = norm_momentum
+    self._norm_epsilon = norm_epsilon
+
+  def __call__(self, inputs, is_training=None):
+    with backend.get_graph().as_default():
+      model = SpineNet(
+          input_specs=self._input_specs,
+          min_level=self._min_level,
+          max_level=self._max_level,
+          block_specs=self._block_specs,
+          endpoints_num_filters=self._endpoints_num_filters,
+          resample_alpha=self._resample_alpha,
+          block_repeats=self._block_repeats,
+          filter_size_scale=self._filter_size_scale,
+          kernel_initializer=self._kernel_initializer,
+          kernel_regularizer=self._kernel_regularizer,
+          bias_regularizer=self._bias_regularizer,
+          activation=self._activation,
+          use_sync_bn=self._use_sync_bn,
+          norm_momentum=self._norm_momentum,
+          norm_epsilon=self._norm_epsilon)
+      return model(inputs)