def _ulogprob_hid(self, Y, num_is_samples=100): """ Estimates the unnormalized marginal log-probabilities of hidden states. Use this method only if you know what you are doing. """ # approximate this SRBM with an RBM rbm = RBM(self.X.shape[0], self.Y.shape[0]) rbm.W = self.W rbm.b = self.b rbm.c = self.c # allocate memory Q = np.asmatrix(np.zeros([num_is_samples, Y.shape[1]])) for k in range(num_is_samples): # draw importance samples X = rbm.backward(Y) # store importance weights Q[k, :] = self._ulogprob(X, Y) - rbm._clogprob_vis_hid(X, Y) # average importance weights to get estimates return utils.logmeanexp(Q, 0)
def _ulogprob_hid(self, Y, num_is_samples=100):
"""
Estimates the unnormalized marginal log-probabilities of hidden states.
Use this method only if you know what you are doing.
"""
# approximate this SRBM with an RBM
rbm = RBM(self.X.shape[0], self.Y.shape[0])
rbm.W = self.W
rbm.b = self.b
rbm.c = self.c
# allocate memory
Q = np.asmatrix(np.zeros([num_is_samples, Y.shape[1]]))
for k in range(num_is_samples):
# draw importance samples
X = rbm.backward(Y)
# store importance weights
Q[k, :] = self._ulogprob(X, Y) - rbm._clogprob_vis_hid(X, Y)
# average importance weights to get estimates
return utils.logmeanexp(Q, 0)
def load_dbn_param(self,dbnpath,softmaxpath):
weights = cPickle.load(open(dbnpath,'rb'))
vlen,hlen = 0,0
self.nlayers = len(weights)
for i in range(self.nlayers):
weight = weights[i]
vlen,hlen = weight.shape[0],weight.shape[1]
rbm = RBM(vlen,hlen)
rbm.W = weight
self.rbm_layers.append(rbm)
print "RBM layer%d shape:%s" %(i,str(rbm.W.shape))
self.softmax = SoftMax()
self.softmax.load_theta(softmaxpath)
print "softmax parameter: "+str(self.softmax.theta.shape)
input = rbm3.reconstruct_from_output(input)
input = rbm2.reconstruct_from_output(input)
input = rbm1.reconstruct_from_output(input)
for i in xrange(input.shape[0]):
first_input = rbm1.input[i]
last_input = input[i]
delta = [x-y for(x, y) in zip(first_input, last_input)]
delta = numpy.array(delta)
# RBM1 finetune
W = rbm1.W
for j in xrange(W.shape[0]):
W[j] = W[j] + finetuning_lr * delta
rbm1.W = W
delta = rbm1.output_from_input(delta)
# RBM2 finetune
W = rbm2.W
for j in xrange(W.shape[0]):
W[j] = W[j] + finetuning_lr * delta
rbm2.W = W
delta = rbm2.output_from_input(delta)
# RBM3 finetune
W = rbm3.W
for j in xrange(W.shape[0]):
W[j] = W[j] + finetuning_lr * delta
rbm3.W = W
delta = rbm3.output_from_input(delta)
