def __init__(self, indim, outdim, hiddim=6):
Module.__init__(self, indim, outdim)
self._network = Network()
self._in_layer = LinearLayer(indim + outdim)
self._hid_layer = LSTMLayer(hiddim)
self._out_layer = LinearLayer(outdim)
self._bias = BiasUnit()
self._network.addInputModule(self._in_layer)
self._network.addModule(self._hid_layer)
self._network.addModule(self._bias)
self._network.addOutputModule(self._out_layer)
self._hid_to_out_connection = FullConnection(self._hid_layer,
self._out_layer)
self._in_to_hid_connection = FullConnection(self._in_layer,
self._hid_layer)
self._network.addConnection(self._hid_to_out_connection)
self._network.addConnection(self._in_to_hid_connection)
self._network.addConnection(FullConnection(self._bias,
self._hid_layer))
self._network.sortModules()
self.time = self._network.time
self.backprojectionFactor = 0.01
def __init__(self, numRows, numColumns, name=None):
""" initialize with the number of rows and columns. the table
values are all set to zero.
"""
Module.__init__(self, 2, 1, name)
ParameterContainer.__init__(self, numRows * numColumns)
self.numRows = numRows
self.numColumns = numColumns
def activate(self, state, action):
""" The super class commonly ignores the state and simply passes the
action through the module. implement _forwardImplementation()
in subclasses.
"""
self.state = state
return Module.activate(self, action)
def sortModules(self):
"""Prepare the network for activation by sorting the internal
datastructure.
Needs to be called before activation."""
if self.sorted:
return
# Sort the modules.
self._topologicalSort()
# Sort the connections by name.
for m in self.modules:
self.connections[m].sort(key=lambda x: x.name)
self.motherconnections.sort(key=lambda x: x.name)
# Create a single array with all parameters.
tmp = [pc.params for pc in self._containerIterator()]
total_size = sum(scipy.size(i) for i in tmp)
ParameterContainer.__init__(self, total_size)
if total_size > 0:
self.params[:] = scipy.concatenate(tmp)
self._setParameters(self.params)
# Create a single array with all derivatives.
tmp = [pc.derivs for pc in self._containerIterator()]
self.resetDerivatives()
self.derivs[:] = scipy.concatenate(tmp)
self._setDerivatives(self.derivs)
# TODO: make this a property; indim and outdim are invalid before
# .sortModules is called!
# Determine the input and output dimensions of the network.
self.indim = sum(m.indim for m in self.inmodules)
self.outdim = sum(m.outdim for m in self.outmodules)
self.indim = 0
for m in self.inmodules:
self.indim += m.indim
self.outdim = 0
for m in self.outmodules:
self.outdim += m.outdim
# Initialize the network buffers.
self.bufferlist = []
Module.__init__(self, self.indim, self.outdim, name=self.name)
self.sorted = True
def adaptAgentObject(agent_object):
"""Return an object that adapts a pybrain agent to the rlglue interface.
"""
# This is pretty hacky: We first take a bogus agent with a bogus module
# for our function adaptAgent to work, then substitue the bogus agent with
# our actual agent.
agent = adaptAgent(LearningAgent)(Module(1, 1))
agent.agent = agent_object
return agent
def __init__(self, outdim, hiddim=15):
""" Create an EvolinoNetwork with for sequences of dimension outdim and
hiddim dimension of the RNN Layer."""
indim = 0
Module.__init__(self, indim, outdim)
self._network = RecurrentNetwork()
self._in_layer = LinearLayer(indim + outdim)
self._hid_layer = LSTMLayer(hiddim)
self._out_layer = LinearLayer(outdim)
self._bias = BiasUnit()
self._network.addInputModule(self._in_layer)
self._network.addModule(self._hid_layer)
self._network.addModule(self._bias)
self._network.addOutputModule(self._out_layer)
self._in_to_hid_connection = FullConnection(self._in_layer,
self._hid_layer)
self._bias_to_hid_connection = FullConnection(self._bias,
self._hid_layer)
self._hid_to_out_connection = FullConnection(self._hid_layer,
self._out_layer)
self._network.addConnection(self._in_to_hid_connection)
self._network.addConnection(self._bias_to_hid_connection)
self._network.addConnection(self._hid_to_out_connection)
self._recurrent_connection = FullConnection(self._hid_layer,
self._hid_layer)
self._network.addRecurrentConnection(self._recurrent_connection)
self._network.sortModules()
self._network.reset()
self.offset = self._network.offset
self.backprojectionFactor = 0.01
def __init__(self, timedim, shape,
hiddendim, outsize, blockshape=None, name=None):
"""Initialize an MdrnnLayer.
The dimensionality of the sequence - for example 2 for a
picture or 3 for a video - is given by `timedim`, while the sidelengths
along each dimension are given by the tuple `shape`.
The layer will have `hiddendim` hidden units per swiping direction. The
number of swiping directions is given by 2**timedim, which corresponds
to one swipe from each corner to its opposing corner and back.
To indicate how many outputs per timesteps are used, you have to specify
`outsize`.
In order to treat blocks of the input and not single voxels, you can
also specify `blockshape`. For example the layer will then feed (2, 2)
chunks into the network at each timestep which correspond to the (2, 2)
rectangles that the input can be split into.
"""
self.timedim = timedim
self.shape = shape
blockshape = tuple([1] * timedim) if blockshape is None else blockshape
self.blockshape = shape
self.hiddendim = hiddendim
self.outsize = outsize
self.indim = reduce(operator.mul, shape, 1)
self.blocksize = reduce(operator.mul, blockshape, 1)
self.sequenceLength = self.indim / self.blocksize
self.outdim = self.sequenceLength * self.outsize
self.bufferlist = [('cellStates', self.sequenceLength * self.hiddendim)]
Module.__init__(self, self.indim, self.outdim, name=name)
# Amount of parameters that are required for the input to the hidden
self.num_in_params = self.blocksize * self.hiddendim * (3 + self.timedim)
# Amount of parameters that are needed for the recurrent connections.
# There is one of the parameter for every time dimension.
self.num_rec_params = outsize * hiddendim * (3 + self.timedim)
# Amount of parameters that are needed for the output.
self.num_out_params = outsize * hiddendim
# Amount of parameters that are needed from the bias to the hidden and
# the output
self.num_bias_params = (3 + self.timedim) * self.hiddendim + self.outsize
# Total list of parameters.
self.num_params = sum((self.num_in_params,
self.timedim * self.num_rec_params,
self.num_out_params,
self.num_bias_params))
ParameterContainer.__init__(self, self.num_params)
# Some layers for internal use.
self.hiddenlayer = MDLSTMLayer(self.hiddendim, self.timedim)
# Every point in the sequence has timedim predecessors.
self.predlayers = [LinearLayer(self.outsize) for _ in range(timedim)]
# We need a single layer to hold the input. We will swipe a connection
# over the corrects part of it, in order to feed the correct input in.
self.inlayer = LinearLayer(self.indim)
# Make some layers the same to save memory.
self.inlayer.inputbuffer = self.inlayer.outputbuffer = self.inputbuffer
# In order to allocate not too much memory, we just set the size of the
# layer to 1 and correct it afterwards.
self.outlayer = LinearLayer(self.outdim)
self.outlayer.inputbuffer = self.outlayer.outputbuffer = self.outputbuffer
self.bias = BiasUnit()
def __init__(self, dim, name=None):
Module.__init__(self, dim, dim * 2, name)
self.setArgs(dim=dim, name=self.name)
def reset(self):
"""Reset all component modules and the network."""
Module.reset(self)
for m in self.modules:
m.reset()