class BN(object):
'''A BN is a bidirectional NeuralNet. It is equivelant to two opposite direction feed forward nets.'''
def __init__(self, layers, up_weights=None, down_weights=None):
'''Initializes a BN from the layers given'''
self.numlayers = len(layers)
down_layers = [layer.__class__.from_layer(layer) for layer in layers[::-1]]#Copy list so that upnet and downnet layers are different objects
self.upnet = NeuralNet(layers, up_weights)
self.downnet = NeuralNet(down_layers, down_weights)
@classmethod
def from_rbms(cls, rbms):
'''Initializes a BN from a list of RBMS. NOTE: the down weights and upweights are tied.
Modifying one, modifies the other. To untie, call the __untie__ method.'''
layers = []
# First layer of dbn is the visible layer of the bottom rbm
layers.append(rbms[0].get_vislayer())
# Keep all hidden layers
for rbm in rbms:
layers.append(rbm.get_hidlayer())
up_weights = [rbm.get_vishid() for rbm in rbms]
down_weights = [rbm.get_vishid().transpose() for rbm in rbms[::-1]]
return cls(layers, up_weights, down_weights)
def bottom_up(self, data):
'''Expects data to be probabilities'''
self.upnet.layers[0].probs = data
self.upnet.layers[0].activities = sample_binary_stochastic(data)
return self.upnet.forward_pass(data, 1)
def top_down(self, data):
'''Expects data to be binary'''
self.downnet.layers[0].activities = data
return self.downnet.forward_pass(data, 1)
def top_down_prob(self, data):
'''Expects data to be binary'''
self.downnet.layers[0].activities = data
last = len(self.downnet.weights)
data = dot(data, self.downnet.weights[0])
for i in range(1, self.downnet.numlayers):
self.downnet.layers[i].process(data)
data = self.downnet.layers[i].probs
if i < last:
data = dot(data, self.downnet.weights[i])
return [layer.activities for layer in self.downnet.layers]
def __untie_weights__(self):
'''This is an ugly step, and is only necessary when the db is initialized from RBMs.
It unties the recognition weights from the generative ones.'''
numweights = self.numlayers - 1
for i in range(numweights):
self.upnet.weights[i] = self.upnet.weights[i].copy()
def wake_phase(self, data):
'''The first step of wake-sleep and contrastive wake-sleep. Returns wake_deltas, a list of matrices by which the
the weights of the down net should be adjusted. Also returns wake_bias_deltas, wake_visbias_delta, and hidden states
of top layer. Assumes DBN layers are binary stochastic layers.'''
#Get the states and probabilities of every layer after doing a bottom-up pass
hid_states = self.bottom_up(data)
hid_probs = []
for layer in self.upnet.layers:
hid_probs.append(layer.probs)
wake_deltas = []
wake_bias_deltas = []
#Bias deltas for the generative visible units
wake_visbias_delta = (data - self.upnet.layers[0].probs).sum(0)/data.shape[0]
#Iterate over each layer excluding bottom layer
for i in range(self.upnet.numlayers - 1):
upper_state = hid_states[i+1]
upper_activity = hid_probs[i+1]
lower_state = hid_states[i]
lower_activity = hid_probs[i]
delta = dot(upper_state.transpose(), (lower_state - lower_activity))/data.shape[0]
#Get bias deltas as well to update hidden biases in downnet
delta_bias = array([(upper_state - upper_activity).sum(0)/data.shape[0]])
wake_deltas.insert(0,delta)
wake_bias_deltas.insert(0, delta_bias)
return wake_deltas, wake_bias_deltas, wake_visbias_delta, hid_states[-1]
def sleep_phase(self, data):
'''The last step of wake-sleep and contrastive wake-sleep. Returns sleep_deltas, a list of matrices by which the
the weights of the up net should be adjusted. Also returns sleep_bias_deltas. Assumes DBN layers are binary stochastic layers.'''
#Get the states and probabilities of every layer after doing a top-down pass
hid_states = self.top_down(data)
hid_probs = []
for layer in self.downnet.layers:
hid_probs.append(layer.probs)
sleep_deltas = []
sleep_bias_deltas = []
#Iterate over each layer excluding top layer
for i in range(self.downnet.numlayers -1):
lower_state = hid_states[i+1]
upper_state = hid_states[i]
upper_activity = hid_probs[i]
delta = dot(lower_state.transpose(), (upper_state - upper_activity))/data.shape[0]
#Get bias deltas to update hidden biases in upnet
delta_bias = array([(upper_state - upper_activity).sum(0)/data.shape[0]])
sleep_deltas.insert(0,delta)
sleep_bias_deltas.insert(0, delta_bias)
return sleep_deltas, sleep_bias_deltas
def wake_sleep(self, data, learning_rate):
'''Combines wake and sleep phases'''
downnet_deltas, downnet_hidbias_deltas, downnet_visbias_delta, top_state = self.wake_phase(data)
upnet_deltas, upnet_bias_deltas = self.sleep_phase(top_state) #The top state is the input for the top-down pass
recons_error = square(data - self.upnet.layers[0].probs).sum()
print 'BN Reconstruction Error', recons_error
self.downnet.layers[-1].bias += learning_rate*downnet_visbias_delta
for i in range(len(downnet_deltas)):
self.downnet.weights[i] += learning_rate*downnet_deltas[i]
self.downnet.layers[i+1].bias += learning_rate*downnet_hidbias_deltas[i]
for i in range(len(upnet_deltas)):
self.upnet.weights[i] += learning_rate*upnet_deltas[i]
self.upnet.layers[i+1].bias += learning_rate*upnet_bias_deltas[i]
return recons_error
