def __init__(self):
self.width = 28
self.height = 28
self.testds, self.trainds = \
makeMnistDataSets('/Users/bayerj/Desktop/MNIST/')
# Initialize MDRNN
self.net = _FeedForwardNetwork()
inlayer = LinearLayer(self.width * self.height)
hiddenlayer = MdrnnLayer(timedim=2,
shape=(self.width, self.height),
blockshape=(1, 1),
hiddendim=4,
outsize=10,
name='mdrnn')
outlayer = SigmoidLayer(self.width * self.height * 10)
con1 = IdentityConnection(inlayer, hiddenlayer)
con2 = IdentityConnection(hiddenlayer, outlayer)
self.net.addInputModule(inlayer)
self.net.addModule(hiddenlayer)
self.net.addOutputModule(outlayer)
self.net.addConnection(con1)
self.net.addConnection(con2)
self.net.sortModules()
def rec_two_layer_network(self, net):
inlayer = LinearLayer(2, 'in')
outlayer = LinearLayer(2, 'out')
con = IdentityConnection(inlayer, outlayer)
rcon = IdentityConnection(inlayer, outlayer)
net.addInputModule(inlayer)
net.addOutputModule(outlayer)
net.addConnection(con)
net.addRecurrentConnection(rcon)
def buildNonGravityNet(recurrent=False):
if recurrent:
net = RecurrentNetwork()
else:
net = FeedForwardNetwork()
l1 = LinearLayer(2)
l2 = LinearLayer(3)
s1 = SigmoidLayer(2)
l3 = LinearLayer(1)
net.addInputModule(l1)
net.addModule(l2)
net.addModule(s1)
net.addOutputModule(l3)
net.addConnection(IdentityConnection(l1, l2, outSliceFrom=1))
net.addConnection(IdentityConnection(l1, l2, outSliceTo=2))
net.addConnection(IdentityConnection(l2, l3, inSliceFrom=2))
net.addConnection(IdentityConnection(l2, l3, inSliceTo=1))
net.addConnection(IdentityConnection(l1, s1))
net.addConnection(IdentityConnection(l2, s1, inSliceFrom=1))
net.addConnection(IdentityConnection(s1, l3, inSliceFrom=1))
if recurrent:
net.addRecurrentConnection(IdentityConnection(s1, l1))
net.addRecurrentConnection(
IdentityConnection(l2, l2, inSliceFrom=1, outSliceTo=2))
net.sortModules()
return net
def buildMinimalLSTMNetwork():
N = RecurrentNetwork('simpleLstmNet')
i = LinearLayer(4, name='i')
h = LSTMLayer(1, peepholes=True, name='lstm')
o = LinearLayer(1, name='o')
N.addInputModule(i)
N.addModule(h)
N.addOutputModule(o)
N.addConnection(IdentityConnection(i, h))
N.addConnection(IdentityConnection(h, o))
N.sortModules()
return N
def buildMinimalMDLSTMNetwork():
N = RecurrentNetwork('simpleMdLstmNet')
i = LinearLayer(4, name = 'i')
h = MDLSTMLayer(1, peepholes = True, name = 'mdlstm')
o = LinearLayer(1, name = 'o')
N.addInputModule(i)
N.addModule(h)
N.addOutputModule(o)
N.addConnection(IdentityConnection(i, h, outSliceTo = 4))
N.addRecurrentConnection(IdentityConnection(h, h, outSliceFrom = 4, inSliceFrom = 1))
N.addConnection(IdentityConnection(h, o, inSliceTo = 1))
N.sortModules()
return N
def sliced_connection_network(self, net):
inlayer = LinearLayer(2, 'in')
outlayer = LinearLayer(2, 'out')
con = IdentityConnection(inlayer, outlayer,
inSliceFrom=0, inSliceTo=1,
outSliceFrom=1, outSliceTo=2,
)
con = IdentityConnection(inlayer, outlayer,
inSliceFrom=1, inSliceTo=2,
outSliceFrom=0, outSliceTo=1,
)
net.addInputModule(inlayer)
net.addOutputModule(outlayer)
net.addConnection(con)
def __init__(self, module, learner = None):
LearningAgent.__init__(self, module, learner)
# exploration module (linear flat network)
self.explorationmodule = buildNetwork(self.indim, self.outdim, bias=False)
# state dependent exploration layer
self.explorationlayer = StateDependentLayer(self.outdim, self.explorationmodule, 'explore')
# add exploration layer to top of network through identity connection
out = self.module.outmodules.pop()
self.module.addOutputModule(self.explorationlayer)
self.module.addConnection(IdentityConnection(out, self.module['explore'], self.module))
self.module.sortModules()
# tell learner the new module
self.learner.setModule(self.module)
# add the log likelihood (loglh) to the dataset and link it to the others
self.history.addField('loglh', self.module.paramdim)
self.history.link.append('loglh')
self.loglh = None
# if this flag is set to True, random weights are drawn after each reward,
# effectively acting like the vanilla policy gradient alg.
self.actaspg = False
def lstm_cell(self, net):
inpt = LinearLayer(4, 'inpt')
forgetgate = GateLayer(1, 'forgetgate')
ingate = GateLayer(1, 'ingate')
outgate = GateLayer(1, 'outgate')
state = LinearLayer(1, 'state')
in_to_fg = IdentityConnection(inpt,
forgetgate,
inSliceFrom=0,
inSliceTo=1,
outSliceFrom=0,
outSliceTo=1,
name='in_to_fg')
in_to_og = IdentityConnection(inpt,
outgate,
inSliceFrom=1,
inSliceTo=2,
outSliceFrom=1,
outSliceTo=2,
name='in_to_og')
in_to_ig = IdentityConnection(inpt,
ingate,
inSliceFrom=2,
inSliceTo=4,
outSliceFrom=0,
outSliceTo=2,
name='in_to_ig')
fg_to_st = IdentityConnection(forgetgate, state, name='fg_to_st')
st_to_fg = IdentityConnection(state,
forgetgate,
outSliceFrom=1,
outSliceTo=2,
name='st_to_fg')
st_to_og = IdentityConnection(state,
outgate,
outSliceFrom=1,
outSliceTo=2,
name='st_to_og')
ig_to_st = IdentityConnection(ingate, state, name='ig_to_st')
net.addInputModule(inpt)
net.addModule(forgetgate)
net.addModule(ingate)
net.addModule(state)
net.addOutputModule(outgate)
net.addConnection(in_to_fg)
net.addConnection(in_to_og)
net.addConnection(in_to_ig)
net.addConnection(fg_to_st)
net.addRecurrentConnection(st_to_fg)
net.addConnection(st_to_og)
net.addConnection(ig_to_st)
def __init__(self, module, learner=None):
assert isinstance(module, FeedForwardNetwork)
assert len(module.outmodules) == 1
LearningAgent.__init__(self, module, learner)
# create gaussian layer
self.explorationlayer = GaussianLayer(self.outdim, name='gauss')
self.explorationlayer.setSigma([-2] * self.outdim)
# add gaussian layer to top of network through identity connection
out = self.module.outmodules.pop()
self.module.addOutputModule(self.explorationlayer)
self.module.addConnection(IdentityConnection(out,
self.module['gauss']))
self.module.sortModules()
# tell learner the new module
self.learner.setModule(self.module)
# add the log likelihood (loglh) to the dataset and link it to the others
self.history.addField('loglh', self.module.paramdim)
self.history.link.append('loglh')
self.loglh = None
def __init__(self, *args, **kwargs):
IdentityConnection.__init__(self, *args, **kwargs)
ParameterContainer.__init__(self, self.indim)
