def eval(self, inputs, labels=None):
_, probs = self.forward(inputs)
logits_idx = tf.cast(tf.argmax(cl.softmax(probs, axis=1), axis=1),
tf.int32)
if labels is None:
return probs, logits_idx, None
return probs, logits_idx, self.accuracy(probs, labels)
def accuracy(self, probs, labels):
logits_idx = tf.cast(tf.argmax(cl.softmax(probs, axis=1), axis=1),
tf.int32)
correct_prediction = tf.equal(tf.cast(labels, tf.int32), logits_idx)
correct = tf.reduce_sum(tf.cast(correct_prediction, tf.float32))
accuracy = tf.reduce_mean(correct /
tf.cast(tf.shape(probs)[0], tf.float32))
cl.summary.scalar('accuracy', accuracy, verbose=cfg.summary_verbose)
return accuracy
def dynamicRouting_v1(votes,
num_routing=3,
use_bias=True,
leaky=True):
""" Dynamic routing algorithm.
See [Sabour et al., 2017](https://arxiv.org/abs/1710.09829).
Args:
votes: A 5-D or 7-D tensor with shape [batch_size, ..., in_channels/num_inputs, num_outputs] + out_caps_dims.
num_routing: Integer, number of routing iterations.
use_bias: Boolean, whether the layer uses a bias.
leaky: Boolean, whether the algorithm uses leaky routing.
Returns:
poses: A 4-D or 6-D tensor.
probs: A 2-D or 4-D tensor.
"""
vote_shape = cl.shape(votes)
logit_shape = vote_shape[:-2] + [1, 1]
logits = tf.fill(logit_shape, 0.0)
squash_on = -2 if vote_shape[-1] == 1 else [-2, -1]
if use_bias:
bias_shape = [1 for i in range(len(vote_shape) - 3)] + vote_shape[-3:]
biases = tf.get_variable("biases",
bias_shape,
initializer=tf.constant_initializer(0.1),
dtype=tf.float32)
vote_stopped = tf.stop_gradient(votes, name="stop_gradient")
for i in range(num_routing):
with tf.variable_scope("iter_" + str(i)):
if leaky:
route = _leaky_routing(logits)
else:
route = cl.softmax(logits, axis=-3)
if i == num_routing - 1:
if use_bias:
preactivate = cl.reduce_sum(tf.multiply(route, votes), axis=-4, keepdims=True) + biases
else:
preactivate = cl.reduce_sum(tf.multiply(route, votes), axis=-4, keepdims=True)
poses = cl.ops.squash(preactivate, axis=squash_on)
else:
if use_bias:
preactivate = cl.reduce_sum(tf.multiply(route, vote_stopped), axis=1, keepdims=True) + biases
else:
preactivate = cl.reduce_sum(tf.multiply(route, vote_stopped), axis=1, keepdims=True)
poses = cl.ops.squash(preactivate, axis=squash_on)
logits += cl.reduce_sum(vote_stopped * poses, axis=-4, keepdims=True)
poses = tf.squeeze(poses, axis=-4)
probs = tf.norm(poses, axis=(-2, -1))
return(poses, probs)
def _leaky_routing(logits):
leak_shape = cl.shape(logits)
leak = tf.zeros(leak_shape[:-3] + [1, 1, 1])
leaky_logits = tf.concat([leak, logits], axis=-3)
leaky_routing = cl.softmax(leaky_logits, axis=-3)
return tf.split(leaky_routing, [1, leak_shape[-3]], axis=-3)[1]
def create_network(self, inputs, labels):
""" Setup capsule network.
Args:
inputs: Tensor or array with shape [batch_size, height, width, channels] or [batch_size, height * width * channels].
labels: Tensor or array with shape [batch_size].
Returns:
poses: [batch_size, num_label, 16, 1].
probs: Tensor with shape [batch_size, num_label], the probability of entity presence.
"""
self.raw_imgs = inputs
self.labels = labels
with tf.variable_scope('Conv1_layer'):
# Conv1, return with shape [batch_size, 20, 20, 256]
inputs = tf.reshape(
self.raw_imgs,
shape=[-1, self.height, self.width, self.channels])
conv1 = tf.layers.conv2d(inputs,
filters=256,
kernel_size=9,
strides=1,
padding='VALID',
activation=tf.nn.relu)
with tf.variable_scope('PrimaryCaps_layer'):
primaryCaps, activation = cl.layers.primaryCaps(
conv1,
filters=64, # MNIST 32
kernel_size=9,
strides=2,
out_caps_dims=[8, 1],
method="norm")
with tf.variable_scope('DigitCaps_layer'):
routing_method = "EMRouting"
num_inputs = np.prod(cl.shape(primaryCaps)[1:4])
primaryCaps = tf.reshape(primaryCaps, shape=[-1, num_inputs, 8, 1])
activation = tf.reshape(activation, shape=[-1, num_inputs])
self.poses, self.probs = cl.layers.dense(
primaryCaps,
activation,
num_outputs=self.num_label,
out_caps_dims=[16, 1],
routing_method=routing_method)
cl.summary.histogram('activation',
self.probs,
verbose=cfg.summary_verbose)
# Decoder structure
# Reconstructe the inputs with 3 FC layers
with tf.variable_scope('Decoder'):
labels = tf.one_hot(self.labels,
depth=self.num_label,
axis=-1,
dtype=tf.float32)
self.labels_one_hoted = tf.reshape(labels,
(-1, self.num_label, 1, 1))
masked_caps = tf.multiply(self.poses, self.labels_one_hoted)
num_inputs = np.prod(masked_caps.get_shape().as_list()[1:])
active_caps = tf.reshape(masked_caps, shape=(-1, num_inputs))
fc1 = tf.layers.dense(active_caps,
units=512,
activation=tf.nn.relu)
fc2 = tf.layers.dense(fc1, units=1024, activation=tf.nn.relu)
num_outputs = self.height * self.width * self.channels
self.recon_imgs = tf.layers.dense(fc2,
units=num_outputs,
activation=tf.sigmoid)
recon_imgs = tf.reshape(
self.recon_imgs,
shape=[-1, self.height, self.width, self.channels])
cl.summary.image('reconstruction_img',
recon_imgs,
verbose=cfg.summary_verbose)
with tf.variable_scope('accuracy'):
logits_idx = tf.to_int32(
tf.argmax(cl.softmax(self.probs, axis=1), axis=1))
correct_prediction = tf.equal(tf.to_int32(self.labels), logits_idx)
correct = tf.reduce_sum(tf.cast(correct_prediction, tf.float32))
self.accuracy = tf.reduce_mean(
correct / tf.cast(tf.shape(self.probs)[0], tf.float32))
cl.summary.scalar('accuracy',
self.accuracy,
verbose=cfg.summary_verbose)
return self.poses, self.probs
def create_network(self, inputs, labels):
""" Setup capsule network.
Args:
inputs: Tensor or array with shape [batch_size, height, width, channels] or [batch_size, height * width * channels].
labels: Tensor or array with shape [batch_size].
Returns:
poses: Tensor with shape [batch_size, num_label, 16, 1].
probs: Tensor with shape [batch_size, num_label], the probability of entity presence.
"""
self.raw_imgs = inputs
self.labels = labels
probs = []
inputs = tf.reshape(inputs,
shape=[-1, self.height, self.width, self.channels])
conv1 = tf.layers.conv2d(inputs,
filters=32,
kernel_size=5,
strides=2,
padding='VALID',
activation=tf.nn.relu,
name="Conv1_layer")
convCaps, activation = cl.layers.primaryCaps(conv1,
filters=32,
kernel_size=1,
strides=1,
out_caps_dims=[4, 4],
method="logistic",
name="PrimaryCaps_layer")
probs.append(tf.reduce_mean(activation))
routing_method = 'EMRouting'
convCaps, activation = cl.layers.conv2d(convCaps,
activation,
filters=32,
out_caps_dims=[4, 4],
kernel_size=(3, 3),
strides=(2, 2),
routing_method=routing_method,
name="ConvCaps1_layer")
probs.append(tf.reduce_mean(activation))
convCaps, activation = cl.layers.conv2d(convCaps,
activation,
filters=32,
out_caps_dims=[4, 4],
kernel_size=(3, 3),
strides=(1, 1),
routing_method=routing_method,
name="ConvCaps2_layer")
probs.append(tf.reduce_mean(activation))
self.poses, self.probs = cl.layers.dense(convCaps,
activation,
num_outputs=self.num_label,
out_caps_dims=[4, 4],
routing_method=routing_method,
coordinate_addition=True,
name="ClassCaps_layer")
probs.append(tf.reduce_mean(self.probs))
tf.summary.scalar("probs", tf.reduce_mean(probs))
# Decoder structure
# Reconstructe the inputs with 3 FC layers
# [batch_size, 1, 16, 1] => [batch_size, 16] => [batch_size, 512]
with tf.variable_scope('Decoder'):
labels = tf.one_hot(self.labels,
depth=self.num_label,
axis=-1,
dtype=tf.float32)
self.labels_one_hoted = tf.reshape(labels,
(-1, self.num_label, 1, 1))
masked_caps = tf.multiply(self.poses, self.labels_one_hoted)
num_inputs = np.prod(masked_caps.get_shape().as_list()[1:])
active_caps = tf.reshape(masked_caps, shape=(-1, num_inputs))
fc1 = tf.layers.dense(active_caps,
units=512,
activation=tf.nn.relu)
fc2 = tf.layers.dense(fc1, units=1024, activation=tf.nn.relu)
num_outputs = self.height * self.width * self.channels
self.recon_imgs = tf.layers.dense(fc2,
units=num_outputs,
activation=tf.sigmoid)
recon_imgs = tf.reshape(
self.recon_imgs,
shape=[-1, self.height, self.width, self.channels])
cl.summary.image('reconstruction_img',
recon_imgs,
verbose=cfg.summary_verbose)
with tf.variable_scope('accuracy'):
cl.summary.histogram('activation',
tf.nn.softmax(self.probs, 1),
verbose=cfg.summary_verbose)
logits_idx = tf.to_int32(
tf.argmax(cl.softmax(self.probs, axis=1), axis=1))
correct_prediction = tf.equal(tf.to_int32(self.labels), logits_idx)
correct = tf.reduce_sum(tf.cast(correct_prediction, tf.float32))
self.accuracy = tf.reduce_mean(
correct / tf.cast(tf.shape(self.probs)[0], tf.float32))
cl.summary.scalar('accuracy',
self.accuracy,
verbose=cfg.summary_verbose)
return self.poses, self.probs