/usr/share/pyshared/mdp/hinet/layer.py is in python-mdp 3.3-1.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 | """
Module for Layers.
Note that additional args and kwargs for train or execute are currently not
supported.
"""
import mdp
from mdp import numx
# TODO: maybe turn self.nodes into a read only property with self._nodes
# TODO: Find a better way to deal with additional args for train/execute?
# Maybe split them by default, but can be disabled via switch?
class Layer(mdp.Node):
"""Layers are nodes which consist of multiple horizontally parallel nodes.
The incoming data is split up according to the dimensions of the internal
nodes. For example if the first node has an input_dim of 50 and the second
node 100 then the layer will have an input_dim of 150. The first node gets
x[:,:50], the second one x[:,50:].
Any additional arguments are forwarded unaltered to each node.
Warning: This might change in the next release (2.5).
Since they are nodes themselves layers can be stacked in a flow (e.g. to
build a layered network). If one would like to use flows instead of nodes
inside of a layer one can use a FlowNode.
"""
def __init__(self, nodes, dtype=None):
"""Setup the layer with the given list of nodes.
The input and output dimensions for the nodes must be already set
(the output dimensions for simplicity reasons). The training phases for
the nodes are allowed to differ.
Keyword arguments:
nodes -- List of the nodes to be used.
"""
self.nodes = nodes
# check nodes properties and get the dtype
dtype = self._check_props(dtype)
# calculate the the dimensions
self.node_input_dims = numx.zeros(len(self.nodes))
input_dim = 0
for index, node in enumerate(nodes):
input_dim += node.input_dim
self.node_input_dims[index] = node.input_dim
output_dim = self._get_output_dim_from_nodes()
super(Layer, self).__init__(input_dim=input_dim,
output_dim=output_dim,
dtype=dtype)
def _get_output_dim_from_nodes(self):
"""Calculate the output_dim from the nodes and return it.
If the output_dim of a node is not set the None is returned.
"""
output_dim = 0
for node in self.nodes:
if node.output_dim is not None:
output_dim += node.output_dim
else:
return None
return output_dim
def _check_props(self, dtype):
"""Check the compatibility of the properties of the internal nodes.
Return the found dtype and check the dimensions.
dtype -- The specified layer dtype.
"""
dtype_list = [] # the dtypes for all the nodes
for i, node in enumerate(self.nodes):
# input_dim for each node must be set
if node.input_dim is None:
err = ("input_dim must be set for every node. " +
"Node #%d (%s) does not comply." % (i, node))
raise mdp.NodeException(err)
if node.dtype is not None:
dtype_list.append(node.dtype)
# check that the dtype is None or the same for every node
nodes_dtype = None
nodes_dtypes = set(dtype_list)
nodes_dtypes.discard(None)
if len(nodes_dtypes) > 1:
err = ("All nodes must have the same dtype (found: %s)." %
nodes_dtypes)
raise mdp.NodeException(err)
elif len(nodes_dtypes) == 1:
nodes_dtype = list(nodes_dtypes)[0]
# check that the nodes dtype matches the specified dtype
if nodes_dtype and dtype:
if not numx.dtype(nodes_dtype) == numx.dtype(dtype):
err = ("Cannot set dtype to %s: " %
numx.dtype(nodes_dtype).name +
"an internal node requires %s" % numx.dtype(dtype).name)
raise mdp.NodeException(err)
elif nodes_dtype and not dtype:
dtype = nodes_dtype
return dtype
def _set_dtype(self, t):
for node in self.nodes:
node.dtype = t
self._dtype = t
def _get_supported_dtypes(self):
# we supported the minimal common dtype set
types = set(mdp.utils.get_dtypes('All'))
for node in self.nodes:
types = types.intersection(node.get_supported_dtypes())
return list(types)
def is_trainable(self):
return any(node.is_trainable() for node in self.nodes)
def is_invertible(self):
return all(node.is_invertible() for node in self.nodes)
def _get_train_seq(self):
"""Return the train sequence.
The length is set by the node with maximum length.
"""
max_train_length = 0
for node in self.nodes:
node_length = len(node._get_train_seq())
if node_length > max_train_length:
max_train_length = node_length
return ([[self._train, self._stop_training]] * max_train_length)
def _train(self, x, *args, **kwargs):
"""Perform single training step by training the internal nodes."""
start_index = 0
stop_index = 0
for node in self.nodes:
start_index = stop_index
stop_index += node.input_dim
if node.is_training():
node.train(x[:, start_index : stop_index], *args, **kwargs)
def _stop_training(self, *args, **kwargs):
"""Stop training of the internal nodes."""
for node in self.nodes:
if node.is_training():
node.stop_training(*args, **kwargs)
if self.output_dim is None:
self.output_dim = self._get_output_dim_from_nodes()
def _pre_execution_checks(self, x):
"""Make sure that output_dim is set and then perform normal checks."""
if self.output_dim is None:
# first make sure that the output_dim is set for all nodes
in_start = 0
in_stop = 0
for node in self.nodes:
in_start = in_stop
in_stop += node.input_dim
node._pre_execution_checks(x[:,in_start:in_stop])
self.output_dim = self._get_output_dim_from_nodes()
if self.output_dim is None:
err = "output_dim must be set at this point for all nodes"
raise mdp.NodeException(err)
super(Layer, self)._pre_execution_checks(x)
def _execute(self, x, *args, **kwargs):
"""Process the data through the internal nodes."""
in_start = 0
in_stop = 0
out_start = 0
out_stop = 0
y = None
for node in self.nodes:
out_start = out_stop
out_stop += node.output_dim
in_start = in_stop
in_stop += node.input_dim
if y is None:
node_y = node.execute(x[:,in_start:in_stop], *args, **kwargs)
y = numx.zeros([node_y.shape[0], self.output_dim],
dtype=node_y.dtype)
y[:,out_start:out_stop] = node_y
else:
y[:,out_start:out_stop] = node.execute(x[:,in_start:in_stop],
*args, **kwargs)
return y
def _inverse(self, x, *args, **kwargs):
"""Combine the inverse of all the internal nodes."""
in_start = 0
in_stop = 0
out_start = 0
out_stop = 0
y = None
for node in self.nodes:
# compared with execute, input and output are switched
out_start = out_stop
out_stop += node.input_dim
in_start = in_stop
in_stop += node.output_dim
if y is None:
node_y = node.inverse(x[:,in_start:in_stop], *args, **kwargs)
y = numx.zeros([node_y.shape[0], self.input_dim],
dtype=node_y.dtype)
y[:,out_start:out_stop] = node_y
else:
y[:,out_start:out_stop] = node.inverse(x[:,in_start:in_stop],
*args, **kwargs)
return y
## container methods ##
def __len__(self):
return len(self.nodes)
def __getitem__(self, key):
return self.nodes.__getitem__(key)
def __contains__(self, item):
return self.nodes.__contains__(item)
def __iter__(self):
return self.nodes.__iter__()
class CloneLayer(Layer):
"""Layer with a single node instance that is used multiple times.
The same single node instance is used to build the layer, so
Clonelayer(node, 3) executes in the same way as Layer([node]*3).
But Layer([node]*3) would have a problem when closing a training phase,
so one has to use CloneLayer.
A CloneLayer can be used for weight sharing in the training phase. It might
be also useful for reducing the memory footprint use during the execution
phase (since only a single node instance is needed).
"""
def __init__(self, node, n_nodes=1, dtype=None):
"""Setup the layer with the given list of nodes.
Keyword arguments:
node -- Node to be cloned.
n_nodes -- Number of repetitions/clones of the given node.
"""
super(CloneLayer, self).__init__((node,) * n_nodes, dtype=dtype)
self.node = node # attribute for convenience
def _stop_training(self, *args, **kwargs):
"""Stop training of the internal node."""
if self.node.is_training():
self.node.stop_training(*args, **kwargs)
if self.output_dim is None:
self.output_dim = self._get_output_dim_from_nodes()
class SameInputLayer(Layer):
"""SameInputLayer is a layer were all nodes receive the full input.
So instead of splitting the input according to node dimensions, all nodes
receive the complete input data.
"""
def __init__(self, nodes, dtype=None):
"""Setup the layer with the given list of nodes.
The input dimensions for the nodes must all be equal, the output
dimensions can differ (but must be set as well for simplicity reasons).
Keyword arguments:
nodes -- List of the nodes to be used.
"""
self.nodes = nodes
# check node properties and get the dtype
dtype = self._check_props(dtype)
# check that the input dimensions are all the same
input_dim = self.nodes[0].input_dim
for node in self.nodes:
if not node.input_dim == input_dim:
err = "The nodes have different input dimensions."
raise mdp.NodeException(err)
output_dim = self._get_output_dim_from_nodes()
# intentionally use MRO above Layer, not SameInputLayer
super(Layer, self).__init__(input_dim=input_dim,
output_dim=output_dim,
dtype=dtype)
@staticmethod
def is_invertible():
return False
def _train(self, x, *args, **kwargs):
"""Perform single training step by training the internal nodes."""
for node in self.nodes:
if node.is_training():
node.train(x, *args, **kwargs)
def _pre_execution_checks(self, x):
"""Make sure that output_dim is set and then perform nromal checks."""
if self.output_dim is None:
# first make sure that the output_dim is set for all nodes
for node in self.nodes:
node._pre_execution_checks(x)
self.output_dim = self._get_output_dim_from_nodes()
if self.output_dim is None:
err = "output_dim must be set at this point for all nodes"
raise mdp.NodeException(err)
# intentionally use MRO above Layer, not SameInputLayer
super(Layer, self)._pre_execution_checks(x)
def _execute(self, x, *args, **kwargs):
"""Process the data through the internal nodes."""
out_start = 0
out_stop = 0
y = None
for node in self.nodes:
out_start = out_stop
out_stop += node.output_dim
if y is None:
node_y = node.execute(x, *args, **kwargs)
y = numx.zeros([node_y.shape[0], self.output_dim],
dtype=node_y.dtype)
y[:,out_start:out_stop] = node_y
else:
y[:,out_start:out_stop] = node.execute(x, *args, **kwargs)
return y
|