def conv2d(inputs,
activation,
filters,
out_caps_dims,
kernel_size,
strides,
padding="valid",
routing_method="EMRouting",
name=None,
reuse=None):
"""A 2D convolutional capsule layer.
Args:
inputs: A 6-D tensor with shape [batch_size, in_height, in_width, in_channels] + in_caps_dims.
activation: A 4-D tensor with shape [batch_size, in_height, in_width, in_channels].
filters: Integer, the dimensionality of the output space (i.e. the number of filters in the convolution).
out_caps_dims: A tuple/list of 2 integers, specifying the dimensions of output capsule, e.g. out_caps_dims=[4, 4] representing that each output capsule has shape [4, 4].
kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions.
strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions.
padding: One of "valid" or "same" (case-insensitive), now only support "valid".
routing_method: One of "EMRouting" or "DynamicRouting", the method of routing-by-agreement algorithm.
name: A string, the name of the layer.
reuse: Boolean, whether to reuse the weights of a previous layer by the same name.
Returns:
pose: A 6-D tensor with shape [batch_size, out_height, out_width, out_channels] + out_caps_dims.
activation: A 4-D tensor with shape [batch_size, out_height, out_width, out_channels].
"""
name = "conv2d" if name is None else name
with tf.variable_scope(name) as scope:
if reuse:
scope.reuse_variables()
input_shape = cl.shape(inputs)
input_rank = len(input_shape)
activation_rank = len(activation.shape)
if not input_rank == 6:
raise ValueError('Inputs to `conv2d` should have rank 6. Received inputs rank:', str(input_rank))
if not activation_rank == 4:
raise ValueError('Activation to `conv2d` should have rank 4. Received activation rank:', str(activation_rank))
if isinstance(kernel_size, int):
kernel_size = [kernel_size, kernel_size, input_shape[3]]
elif isinstance(kernel_size, (list, tuple)) and len(kernel_size) == 2:
kernel_size = [kernel_size[0], kernel_size[1], input_shape[3]]
else:
raise ValueError('"kernel_size" should be an integer or tuple/list of 2 integers. Received:', str(kernel_size))
if isinstance(strides, int):
strides = [strides, strides, 1]
elif isinstance(strides, (list, tuple)) and len(strides) == 2:
strides = [strides[0], strides[1], 1]
else:
raise ValueError('"strides" should be an integer or tuple/list of 2 integers. Received:', str(kernel_size))
if not isinstance(out_caps_dims, (list, tuple)) or len(out_caps_dims) != 2:
raise ValueError('"out_caps_dims" should be a tuple/list of 2 integers. Received:', str(out_caps_dims))
elif isinstance(out_caps_dims, tuple):
out_caps_dims = list(out_caps_dims)
# 1. space to batch
# patching everything into [batch_size, out_height, out_width, in_channels] + in_caps_dims (batched)
# and [batch_size, out_height, out_width, in_channels] (activation).
batched = cl.space_to_batch_nd(inputs, kernel_size, strides)
activation = cl.space_to_batch_nd(activation, kernel_size, strides)
# 2. transforming
# transforming to [batch_size, out_height, out_width, in_channels, out_channels/filters] + out_caps_dims
vote = transforming(batched,
num_outputs=filters,
out_caps_dims=out_caps_dims)
# 3. routing
pose, activation = routing(vote, activation, method=routing_method)
return pose, activation
def dense(inputs, activation,
num_outputs,
out_caps_dims,
routing_method='EMRouting',
routing_iter=3,
coordinate_addition=False,
reuse=None,
name=None):
"""A fully connected capsule layer.
Args:
inputs: A 4-D tensor with shape [batch_size, num_inputs] + in_caps_dims or [batch_size, in_height, in_width, in_channels] + in_caps_dims
activation: [batch_size, num_inputs] or [batch_size, in_height, in_width, in_channels]
num_outputs: Integer, the number of output capsules in the layer.
out_caps_dims: A list with two elements, pose shape of output capsules.
routing_iter: Number of iterations during routing algorithm.
Returns:
pose: A 4-D tensor with shape [batch_size, num_outputs] + out_caps_dims
activation: [batch_size, num_outputs]
"""
name = "dense" if name is None else name
with tf.name_scope(name) as scope:
if reuse:
scope.reuse()
if coordinate_addition and len(inputs.shape) == 6 and len(activation.shape) == 4:
vote = transforming(inputs, num_outputs=num_outputs, out_caps_dims=out_caps_dims)
with tf.name_scope("coodinate_addition"):
batch_size, in_height, in_width, in_channels, _, out_caps_height, out_caps_width = cl.shape(vote)
num_inputs = in_height * in_width * in_channels
zeros = np.zeros((in_height, out_caps_width - 1))
coord_offset_h = ((np.arange(in_height) + 0.5) / in_height).reshape([in_height, 1])
coord_offset_h = np.concatenate([zeros, coord_offset_h], axis=-1)
zeros = np.zeros((out_caps_height - 1, out_caps_width))
coord_offset_h = np.stack([np.concatenate([coord_offset_h[i:(i + 1), :], zeros], axis=0) for i in range(in_height)], axis=0)
coord_offset_h = coord_offset_h.reshape((1, in_height, 1, 1, 1, out_caps_height, out_caps_width))
zeros = np.zeros((1, in_width))
coord_offset_w = ((np.arange(in_width) + 0.5) / in_width).reshape([1, in_width])
coord_offset_w = np.concatenate([zeros, coord_offset_w, zeros, zeros], axis=0)
zeros = np.zeros((out_caps_height, out_caps_width - 1))
coord_offset_w = np.stack([np.concatenate([zeros, coord_offset_w[:, i:(i + 1)]], axis=1) for i in range(in_width)], axis=0)
coord_offset_w = coord_offset_w.reshape((1, 1, in_width, 1, 1, out_caps_height, out_caps_width))
vote = vote + tf.constant(coord_offset_h + coord_offset_w, dtype=tf.float32)
vote = tf.reshape(vote, shape=[batch_size, num_inputs, num_outputs] + out_caps_dims)
activation = tf.reshape(activation, shape=[batch_size, num_inputs])
elif len(inputs.shape) == 4 and len(activation.shape) == 2:
vote = transforming(inputs, num_outputs=num_outputs, out_caps_dims=out_caps_dims)
else:
raise TypeError("Wrong rank for inputs or activation")
pose, activation = routing(vote, activation, routing_method, num_iter=routing_iter)
# pose, activation = cl.core.gluing(vote, activation)
assert len(pose.shape) == 4
assert len(activation.shape) == 2
return(pose, activation)
def conv3d(inputs,
activation,
filters,
out_caps_dims,
kernel_size,
strides,
padding="valid",
routing_method="EMRouting",
name=None,
reuse=None):
"""A 3D convolutional capsule layer.
Args:
inputs: A 7-D tensor with shape [batch_size, in_depth, in_height, in_width, in_channels] + in_caps_dims.
activation: A 5-D tensor with shape [batch_size, in_depth, in_height, in_width, in_channels].
filters: Integer, the dimensionality of the output space (i.e. the number of filters in the convolution).
out_caps_dims: A tuple/list of 2 integers, specifying the dimensions of output capsule, e.g. out_caps_dims=[4, 4] representing that each output capsule has shape [4, 4].
kernel_size: An integer or tuple/list of 3 integers, specifying the height and width of the 3D convolution window. Can be a single integer to specify the same value for all spatial dimensions.
strides: An integer or tuple/list of 3 integers, specifying the strides of the convolution along the height and width. Can be a single integer to specify the same value for all spatial dimensions.
padding: One of "valid" or "same" (case-insensitive), now only support "valid".
routing_method: One of "EMRouting" or "DynamicRouting", the method of routing-by-agreement algorithm.
name: String, a name for the operation (optional).
reuse: Boolean, whether to reuse the weights of a previous layer by the same name.
Returns:
pose: A 7-D tensor with shape [batch_size, out_depth, out_height, out_width, out_channels] + out_caps_dims.
activation: A 5-D tensor with shape [batch_size, out_depth, out_height, out_width, out_channels].
"""
name = "conv1d" if name is None else name
with tf.name_scope(name):
input_shape = cl.shape(inputs)
input_rank = len(input_shape)
activation_rank = len(activation.shape)
if input_rank != 7:
raise ValueError('Inputs to `conv3d` should have rank 7. Received input rank:', str(input_rank))
if activation_rank != 5:
raise ValueError('Activation to `conv3d` should have rank 5. Received input shape:', str(activation_rank))
if isinstance(kernel_size, int):
kernel_size = [kernel_size, kernel_size, kernel_size]
elif isinstance(kernel_size, (list, tuple)) and len(kernel_size) == 3:
kernel_size = kernel_size
else:
raise ValueError('"kernel_size" should be an integer or tuple/list of 3 integers. Received:', str(kernel_size))
if isinstance(strides, int):
strides = [strides, strides, strides]
elif isinstance(strides, (list, tuple)) and len(strides) == 3:
strides = strides
else:
raise ValueError('"strides" should be an integer or tuple/list of 3 integers. Received:', str(strides))
if not isinstance(out_caps_dims, (list, tuple)) or len(out_caps_dims) != 2:
raise ValueError('"out_caps_dims" should be a tuple/list of 2 integers. Received:', str(out_caps_dims))
elif isinstance(out_caps_dims, tuple):
out_caps_dims = list(out_caps_dims)
# 1. space to batch
batched = cl.space_to_batch_nd(inputs, kernel_size, strides)
activation = cl.space_to_batch_nd(activation, kernel_size, strides)
# 2. transforming
vote = transforming(batched,
num_outputs=filters,
out_caps_dims=out_caps_dims)
# 3. routing
pose, activation = routing(vote, activation, method=routing_method)
return pose, activation
