def fully_connected(inputs,
activation,
num_outputs,
out_caps_shape,
routing_method='EMRouting',
reuse=None):
'''A capsule fully connected layer.
Args:
inputs: A tensor with shape [batch_size, num_inputs] + in_caps_shape.
activation: [batch_size, num_inputs]
num_outputs: Integer, the number of output capsules in the layer.
out_caps_shape: A list with two elements, pose shape of output capsules.
Returns:
pose: [batch_size, num_outputs] + out_caps_shape
activation: [batch_size, num_outputs]
'''
in_pose_shape = inputs.get_shape().as_list()
# it confuses me a lot. please refer to https://blog.csdn.net/loseinvain/article/details/78762739
# in_pose_shape=tf.shape(inputs) # dynamic shape, i can not get its shape without feed_dict
# num_inputs=in_pose_shape.eval()[1]
num_inputs = 24 * 24 * 32 # 18432
# with tf.Session() as sess:
# num_outputs=sess.run(tf.shape(inputs))
# num_inputs=tf.unstack(tf.shape(num_outputs))[1]
batch_size = in_pose_shape[0]
T_size = get_transformation_matrix_shape(in_pose_shape[-2:],
out_caps_shape)
T_shape = [1, num_inputs, num_outputs] + T_size
# T_matrix = tf.get_variable('transformation_matrix', shape=T_shape)
T_matrix = tf.get_variable(name='transformation_matrix', shape=T_shape)
T_matrix = tf.tile(T_matrix, [batch_size, 1, 1, 1, 1])
inputs = tf.tile(tf.expand_dims(inputs, axis=2), [1, 1, num_outputs, 1, 1])
with tf.variable_scope('transformation'):
# vote: [batch_size, num_inputs, num_outputs] + out_caps_shape
vote = tf.matmul(T_matrix, inputs)
with tf.variable_scope('routing'):
if routing_method == 'EMRouting':
shape = [-1] + activation.get_shape().as_list()[1:] + [1, 1]
activation = tf.reshape(activation, shape=shape)
vote = tf.reshape(vote,
shape=[batch_size, num_inputs, num_outputs, -1])
pose, activation = routing(vote, activation, num_outputs,
out_caps_shape, routing_method)
pose = tf.reshape(pose,
shape=[batch_size, num_outputs] + out_caps_shape)
activation = tf.reshape(activation, shape=[batch_size, -1])
elif routing_method == 'DynamicRouting':
pose, _ = routing(vote,
activation,
num_outputs=num_outputs,
out_caps_shape=out_caps_shape,
method=routing_method)
pose = tf.squeeze(pose, axis=1)
activation = tf.squeeze(euclidean_norm(pose))
return (pose, activation)
def fully_connected(inputs,
activation,
num_outputs,
out_caps_shape,
routing_method='DynamicRouting',
reuse=None):
"""
A capsule fully connected layer.
Args:
inputs: A tensor with shape [batch_size, num_inputs] + in_caps_shape.
activation: [batch_size, num_inputs]
num_outputs: Integer, the number of output capsules in the layer.
out_caps_shape: A list with two elements, pose shape of output capsules.
Returns:
pose: [batch_size, num_outputs] + out_caps_shape
activation: [batch_size, num_outputs]
"""
in_pose_shape = inputs.get_shape().as_list()
num_inputs = in_pose_shape[1]
batch_size = in_pose_shape[0]
T_size = get_transformation_matrix_shape(in_pose_shape[-2:],
out_caps_shape)
T_shape = [1, num_inputs, num_outputs] + T_size
T_matrix = tf.get_variable("transformation_matrix", shape=T_shape)
T_matrix = tf.tile(T_matrix, [batch_size, 1, 1, 1, 1])
inputs = tf.tile(tf.expand_dims(inputs, axis=2), [1, 1, num_outputs, 1, 1])
with tf.variable_scope('transformation'):
# vote: [batch_size, num_inputs, num_outputs] + out_caps_shape
vote = tf.matmul(T_matrix, inputs)
with tf.variable_scope('routing'):
if routing_method == 'EMRouting':
activation = tf.reshape(activation,
shape=activation.get_shape().as_list() +
[1, 1])
vote = tf.reshape(vote,
shape=[batch_size, num_inputs, num_outputs, -1])
pose, activation = routing(vote, activation, num_outputs,
out_caps_shape, routing_method)
pose = tf.reshape(pose,
shape=[batch_size, num_outputs] + out_caps_shape)
activation = tf.reshape(activation, shape=[batch_size, -1])
elif routing_method == 'DynamicRouting':
pose, _ = routing(vote,
activation,
num_outputs=num_outputs,
out_caps_shape=out_caps_shape,
method=routing_method)
pose = tf.squeeze(pose, axis=1)
activation = tf.squeeze(euclidean_norm(pose))
return pose, activation
def primaryCaps(input,
filters,
kernel_size,
strides,
out_caps_shape,
method=None,
regularizer=None):
'''PrimaryCaps layer
Args:
input: [batch_size, in_height, in_width, in_channels].
filters: Integer, the dimensionality of the output space.
kernel_size: ...
strides: ...
out_caps_shape: ...
method: the method of calculating probability of entity existence(logistic, norm, None)
Returns:
pose: [batch_size, out_height, out_width, filters] + out_caps_shape
activation: [batch_size, out_height, out_width, filters]
'''
# pose matrix
pose_size = reduce(lambda x, y: x * y, out_caps_shape)
pose = tf.layers.conv2d(input,
filters * pose_size,
kernel_size=kernel_size,
strides=strides,
activation=None,
activity_regularizer=regularizer)
pose_shape = [-1] + pose.get_shape().as_list()[1:3] + [filters
] + out_caps_shape
pose = tf.reshape(pose, shape=pose_shape)
if method == 'logistic':
# logistic activation unit
activation = tf.layers.conv2d(input,
filters,
kernel_size=kernel_size,
strides=strides,
activation=tf.nn.sigmoid,
activity_regularizer=regularizer)
elif method == 'norm':
activation = euclidean_norm(pose)
else:
activation = None
return (pose, activation)