def inference(self, features):
"""Adds the inference graph ops.
Builds the architecture of the neural net to drive logits from features.
The inference graph includes a convolution layer, a primary capsule layer
and a 10-capsule final layer. Optionally, it also adds the reconstruction
network on top of the 10-capsule final layer.
Args:
features: Dictionary of batched feature tensors like images and labels.
Returns:
A model.Inferred named tuple of expected outputs of the model like
'logits' and 'recons' for the reconstructions.
"""
image_dim = features['height']
image_depth = features['depth']
image = features['images']
image_4d = tf.reshape(image, [-1, image_depth, image_dim, image_dim])
# ReLU Convolution (conv1 layer start)
with tf.variable_scope('conv1') as scope:
kernel = variables.weight_variable(shape=[9, 9, image_depth, 256],
stddev=5e-2,
verbose=self._hparams.verbose)
biases = variables.bias_variable([256],
verbose=self._hparams.verbose)
conv1 = tf.nn.conv2d(image_4d,
kernel, [1, 1, 1, 1],
padding=self._hparams.padding,
data_format='NCHW')
pre_activation = tf.nn.bias_add(conv1, biases, data_format='NCHW')
relu1 = tf.nn.relu(pre_activation, name=scope.name)
if self._hparams.verbose:
tf.summary.histogram('activation', relu1)
# conv1 laeyr end, return [128, 256, 20, 20] in NCHW format.
# Then expand dims to [128, 1, 256, 20, 20]
hidden1 = tf.expand_dims(relu1, 1)
# Capsules, including primary capsules layer and digit capsules layer.
# The final output here is [batch_size, 10, 16]
capsule_output = self._build_capsule(hidden1, features['num_classes'])
# calculate the length of v using vector norm or 2-norm ||v||_2
# return [batch_size, 10]
# equals to the sqrt(reduce_sum(square(v)))
logits = tf.norm(capsule_output, axis=-1)
# Reconstruction
if self._hparams.remake:
remake = self._remake(features, capsule_output)
else:
remake = None
return model.Inferred(logits, remake)
def inference(self, features):
"""Adds the inference graph ops.
Builds the architecture of the neural net to drive logits from features.
The inference graph includes a convolution layer, a primary capsule layer
and a 10-capsule final layer. Optionally, it also adds the reconstruction
network on top of the 10-capsule final layer.
Args:
features: Dictionary of batched feature tensors like images and labels.
Returns:
A model.Inferred named tuple of expected outputs of the model like
'logits' and 'recons' for the reconstructions.
"""
image_height = features['height']
image_width = features['width']
image_depth = features['depth']
image = features['images']
image_4d = tf.reshape(image,
[-1, image_depth, image_height, image_width])
# ReLU Convolution
with tf.variable_scope('conv1') as scope:
kernel = variables.weight_variable(shape=[9, 9, image_depth, 256],
stddev=5e-2,
verbose=self._hparams.verbose)
biases = variables.bias_variable([256],
verbose=self._hparams.verbose)
conv1 = tf.nn.conv2d(image_4d,
kernel, [1, 1, 1, 1],
padding=self._hparams.padding,
data_format='NCHW')
pre_activation = tf.nn.bias_add(conv1, biases, data_format='NCHW')
relu1 = tf.nn.relu(pre_activation, name=scope.name)
if self._hparams.verbose:
tf.summary.histogram('activation', relu1)
hidden1 = tf.expand_dims(relu1, 1)
# Capsules
capsule_output = self._build_capsule(hidden1, features['num_classes'])
logits = tf.norm(capsule_output, axis=-1)
# Reconstruction
if self._hparams.remake:
remake = self._remake(features, capsule_output)
else:
remake = None
return model.Inferred(logits, remake)
def inference(self, features):
"""Adds the inference graph ops.
Builds the architecture of the neural net to drive logits from
features.
The inference graph includes a series of convolution and fully
connected layers and outputs a [batch, 10] tensor as the logits.
Args:
features: Dictionary of batched feature tensors like images and
labels.
Returns:
A model.Inferred named tuple of expected outputs of the model like
'logits' and 'remakes' for the reconstructions (to be added).
"""
image = features['images']
image_dim = features['height']
image_depth = features['depth']
image_4d = tf.reshape(image, [-1, image_depth, image_dim, image_dim])
conv = self._add_convs(image_4d, [image_depth, 512, 256])
hidden1 = tf.contrib.layers.flatten(conv)
with tf.variable_scope('fc1') as scope:
dim = hidden1.get_shape()[1].value
weights = variables.weight_variable(shape=[dim, 1024],
stddev=0.1,
verbose=self._hparams.verbose)
biases = variables.bias_variable(shape=[1024],
verbose=self._hparams.verbose)
pre_activation = tf.matmul(hidden1, weights) + biases
hidden2 = tf.nn.relu(pre_activation, name=scope.name)
with tf.variable_scope('softmax_layer') as scope:
weights = variables.weight_variable(
shape=[1024, features['num_classes']],
stddev=0.1,
verbose=self._hparams.verbose)
biases = variables.bias_variable(shape=[features['num_classes']],
verbose=self._hparams.verbose)
logits = tf.matmul(hidden2, weights) + biases
return model.Inferred(logits, None)
def build_replica(self, tower_idx):
"""Adds a replica graph ops.
Builds the architecture of the neural net to derive logits from
batched_dataset. The inference graph defined here should involve
trainable variables otherwise the optimizer will raise a ValueError.
Args:
tower_idx: the index number for this tower. Each tower is named
as tower_{tower_idx} and resides on gpu:{tower_idx}.
Returns:
Inferred namedtuple containing (logits, recons).
"""
# Image specs
image_size = self._specs['image_size']
image_depth = self._specs['depth']
num_classes = self._specs['num_classes']
# Define input_tensor for input batched_images
batched_images = tf.placeholder(
tf.float32,
shape=[None, image_depth, image_size, image_size],
name='batched_images') # (?, 3, h, w)
"""visual"""
tf.add_to_collection('tower_%d_batched_images' % tower_idx,
batched_images)
# declare the threshold placeholder for ensemble evaluation
threshold = tf.placeholder(tf.float32, name='threshold')
tf.add_to_collection('tower_%d_batched_threshold' % tower_idx,
threshold)
# ReLU Convolution
with tf.variable_scope('conv1') as scope:
kernel = variables.weight_variable(shape=[9, 9, image_depth, 256],
stddev=5e-2,
verbose=self._hparams.verbose)
biases = variables.bias_variable([256],
verbose=self._hparams.verbose)
conv1 = tf.nn.conv2d(batched_images,
kernel,
strides=[1, 1, 1, 1],
padding=self._hparams.padding,
data_format='NCHW')
pre_activation = tf.nn.bias_add(conv1,
biases,
data_format='NCHW',
name='logits')
"""visual"""
tf.add_to_collection('tower_%d_visual' % tower_idx, pre_activation)
relu1 = tf.nn.relu(pre_activation, name=scope.name)
if self._hparams.verbose:
tf.summary.histogram(scope.name + '/activation', relu1)
hidden1 = tf.expand_dims(
relu1, 1) # (?, 1, 3, h, w) h,w are different from previous ones.
# Capsules
capsule_output = self._build_capsule(hidden1, num_classes, tower_idx)
logits = tf.norm(capsule_output, axis=-1, name='logits')
"""visual"""
tf.add_to_collection('tower_%d_visual' % tower_idx, logits)
# Declare one-hot format placeholder for batched_labels
batched_labels = tf.placeholder(tf.int32,
shape=[None, num_classes],
name='batched_labels')
tf.add_to_collection('tower_%d_batched_labels' % tower_idx,
batched_labels)
# Reconstruction
remake = None
if self._hparams.remake:
remake = self._remake(capsule_output, batched_images,
batched_labels)
tf.add_to_collection('tower_%d_recons' % tower_idx, remake)
else:
remake = None
return model.Inferred(logits, remake)
def build_replica(self, tower_idx):
"""Adds a replica graph ops.
Builds the architecture of the neural net to derive logits from
batched_dataset. The inference graph defined here should involve
trainable variables otherwise the optimizer will raise a ValueError.
Args:
tower_idx: the index number for this tower. Each tower is named
as tower_{tower_idx} and resides on gpu:{tower_idx}.
Returns:
Inferred namedtuple containing (logits, None).
"""
# Image specs
image_size = self._specs['image_size']
image_depth = self._specs['depth']
num_classes = self._specs['num_classes']
# Define input_tensor for input batched_images
batched_images = tf.placeholder(tf.float32,
shape=[None, image_depth, image_size, image_size],
name='batched_images')
"""visual"""
tf.add_to_collection('tower_%d_batched_images' % tower_idx, batched_images)
# Add convolutional layers
conv_out = self._add_convs(batched_images, [image_depth, 512, 256], tower_idx)
hidden1 = tf.contrib.layers.flatten(conv_out) # flatten neurons, shape (?, rest)
# Add fully connected layer 1, activation = relu
with tf.variable_scope('fc1') as scope:
dim = hidden1.get_shape()[1].value
weights = variables.weight_variable(shape=[dim, 1024], stddev=0.1,
verbose=self._hparams.verbose)
biases = variables.bias_variable(shape=[1024],
verbose=self._hparams.verbose)
pre_activation = tf.add(tf.matmul(hidden1, weights), biases, name='logits')
"""visual"""
tf.add_to_collection('tower_%d_visual' % tower_idx, pre_activation)
hidden2 = tf.nn.relu(pre_activation, name=scope.name)
# Add fully connected layer 2, activation = None
with tf.variable_scope('softmax_layer') as scope:
weights = variables.weight_variable(
shape=[1024, num_classes], stddev=0.1,
verbose=self._hparams.verbose)
biases = variables.bias_variable(
shape=[num_classes],
verbose=self._hparams.verbose)
logits = tf.add(tf.matmul(hidden2, weights), biases, name='logits')
"""visual"""
tf.add_to_collection('tower_%d_visual' % tower_idx, logits)
# Declare one-hot format placeholder for batched_labels
batched_labels = tf.placeholder(tf.int32,
shape=[None, num_classes], name='batched_labels') # 'tower_i/batched_labels:0'
"""visual"""
tf.add_to_collection('tower_%d_batched_labels' % tower_idx, batched_labels)
return model.Inferred(logits, None)
def inference(self, features):
logits = variables.bias_variable([1, 3])
return model.Inferred(logits, None)
def inference(self, features):
logits = tf.constant([[0.1, 0.2, 0.7]])
return model.Inferred(logits, None)
