def _build_capsule(self, input_tensor, num_classes):
"""Adds the capsule layers.
A slim convolutional capsule layer transforms the input tensor to capsule
format. The nonlinearity for slim convolutional capsule is squash function
but there is no routing and each spatial instantiation of capsule is derived
as traditional convolutional layer.
In order to connect the convolutional capsule layer to the top fully
connected capsule layer the grid position of convolution capsule is
merged with different types of capsules dimmension and capsule2 learns
different transformations for each of them.
Args:
input_tensor: 5D input tensor, shape: [batch, 1, 256, height, width].
num_classes: Number of object categories. Used as the output dimmension.
Returns:
A 3D tensor of the top capsule layer with 10 capsule embeddings.
"""
# PrimaryCaps Layer Start
capsule1 = layers.conv_slim_capsule(
input_tensor,
input_dim=1,
output_dim=self._hparams.num_prime_capsules,
layer_name='conv_capsule1',
num_routing=1,
input_atoms=256,
output_atoms=8,
stride=2,
kernel_size=9,
padding=self._hparams.padding,
leaky=self._hparams.leaky,
)
capsule1_atom_last = tf.transpose(capsule1, [0, 1, 3, 4, 2])
# PrimaryCaps Layer End, return [batch_size, 32, 6, 6, 8] in format
# [batch_size, capsule_channels, height, width, capsule_dims]
# Since the digitCaps layer is a fully connected capsule layer, reshape to
# [batch_size, 1152, 8] will be easier to deal with
capsule1_3d = tf.reshape(capsule1_atom_last,
[tf.shape(input_tensor)[0], -1, 8])
_, _, _, height, width = capsule1.get_shape()
# get 1152, the number of capsules in primaryCaps layer
input_dim = self._hparams.num_prime_capsules * height.value * width.value
# DigitCaps layer, return [batch_size, 10, 16]
return layers.capsule(
input_tensor=capsule1_3d,
input_dim=input_dim,
output_dim=num_classes,
layer_name='capsule2',
input_atoms=8,
output_atoms=16,
num_routing=self._hparams.routing,
leaky=self._hparams.leaky,
)
def testCapsule(self):
"""Tests the correct output and variable declaration of layers.capsule."""
input_tensor = tf.random_uniform((4, 3, 2))
output = layers.capsule(
input_tensor=input_tensor,
input_dim=3,
output_dim=2,
layer_name='capsule',
input_atoms=2,
output_atoms=5,
num_routing=3,
leaky=False)
self.assertListEqual(output.get_shape().as_list(), [4, 2, 5])
trainable_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
self.assertEqual(len(trainable_vars), 2)
self.assertStartsWith(trainable_vars[0].name, 'capsule')
def _build_capsule(self, input_tensor, num_classes):
"""Adds the capsule layers.
A slim convolutional capsule layer transforms the input tensor to capsule
format. The nonlinearity for slim convolutional capsule is squash function
but there is no routing and each spatial instantiation of capsule is derived
as traditional convolutional layer.
In order to connect the convolutional capsule layer to the top fully
connected capsule layer the grid position of convolution capsule is
merged with different types of capsules dimmension and capsule2 learns
different transformations for each of them.
Args:
input_tensor: 5D input tensor, shape: [batch, 1, 256, height, width].
num_classes: Number of object categories. Used as the output dimmension.
Returns:
A 3D tensor of the top capsule layer with 10 capsule embeddings.
"""
capsule1 = layers.conv_slim_capsule(
input_tensor,
input_dim=1,
output_dim=self._hparams.num_prime_capsules,
layer_name='conv_capsule1',
num_routing=1,
input_atoms=self._hparams.conv1_channel,
output_atoms=self._hparams.prime_capsule_dim,
stride=2,
kernel_size=9,
padding=self._hparams.padding,
leaky=self._hparams.leaky,
)
capsule1_atom_last = tf.transpose(capsule1, [0, 1, 3, 4, 2])
capsule1_3d = tf.reshape(
capsule1_atom_last,
[tf.shape(input_tensor)[0], -1, self._hparams.prime_capsule_dim])
_, _, _, height, width = capsule1.get_shape()
input_dim = self._hparams.num_prime_capsules * height.value * width.value
return layers.capsule(
input_tensor=capsule1_3d,
input_dim=input_dim,
output_dim=num_classes,
layer_name='capsule2',
input_atoms=self._hparams.prime_capsule_dim,
output_atoms=self._hparams.digit_capsule_dim,
num_routing=self._hparams.routing,
leaky=self._hparams.leaky,
)
def _build_capsule(self, input_tensor, num_classes):
"""Adds the capsule layers.
A slim convolutional capsule layer transforms the input tensor to capsule
format. The nonlinearity for slim convolutional capsule is squash function
but there is no routing and each spatial instantiation of capsule is derived
as traditional convolutional layer.
In order to connect the convolutional capsule layer to the top fully
connected capsule layer the grid position of convolution capsule is
merged with different types of capsules dimmension and capsule2 learns
different transformations for each of them.
Args:
input_tensor: 5D input tensor, shape: [batch, 1, 256, height, width].
num_classes: Number of object categories. Used as the output dimmension.
Returns:
A 3D tensor of the top capsule layer with 10 capsule embeddings.
"""
capsule1 = layers.conv_slim_capsule(
input_tensor,
input_dim=1,
output_dim=self._hparams.num_prime_capsules,
layer_name='conv_capsule1',
num_routing=1,
input_atoms=256,
output_atoms=8,
stride=2,
kernel_size=9,
padding=self._hparams.padding,
leaky=self._hparams.leaky,)
capsule1_atom_last = tf.transpose(capsule1, [0, 1, 3, 4, 2])
capsule1_3d = tf.reshape(capsule1_atom_last,
[tf.shape(input_tensor)[0], -1, 8])
_, _, _, height, width = capsule1.get_shape()
input_dim = self._hparams.num_prime_capsules * height.value * width.value
return layers.capsule(
input_tensor=capsule1_3d,
input_dim=input_dim,
output_dim=num_classes,
layer_name='capsule2',
input_atoms=8,
output_atoms=16,
num_routing=self._hparams.routing,
leaky=self._hparams.leaky,)
