def cost(self, readouts, outputs):
mu, sigma, coeff = self.components(readouts)
#ipdb.set_trace()
return GMM(outputs, mu, sigma, coeff)
bricks = [mlp_x, transition, mlp_gmm]
for brick in bricks:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.)
brick.initialize()
##############
# Test model
##############
x_g = mlp_x.apply(x)
h = transition.apply(x_g)
mu, sigma, coeff = mlp_gmm.apply(h[-2])
cost = GMM(y, mu, sigma, coeff)
cost = cost.mean()
cost.name = 'sequence_log_likelihood'
cg = ComputationGraph(cost)
model = Model(cost)
#################
# Algorithm
#################
n_batches = 139 * 16
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule(
############## # Test model ############## x_g = mlp_x.apply(x) h = transition.apply(x_g) mu, sigma, coeff = mlp_gmm.apply(h[-2]) #from theano import function #x_tr, x_mask_tr, y_tr = next(data_stream.get_epoch_iterator()) #print function([x], x_g)(x_tr).shape #print function([x], h)(x_tr)[-2].shape #print function([x], mu)(x_tr).shape from play.utils import GMM cost = GMM(y, mu, sigma, coeff) cost = cost*x_mask cost = cost.sum()/x_mask.sum() cost.name = 'sequence_log_likelihood' cg = ComputationGraph(cost) model = Model(cost) #ipdb.set_trace() #x_tr, x_mask_tr, y_tr = next(data_stream.get_epoch_iterator()) #print function([x, x_mask, y], cost)(x_tr, x_mask_tr, y_tr) ################# # Algorithm #################
bricks = [mlp_x, transition, mlp_gmm]
for brick in bricks:
brick.weights_init = IsotropicGaussian(0.01)
brick.biases_init = Constant(0.)
brick.initialize()
##############
# Test model
##############
x_g = mlp_x.apply(x)
h = transition.apply(x_g)
mu, sigma, coeff = mlp_gmm.apply(h[-2])
cost = GMM(y, mu, sigma, coeff)
cost = cost.mean()
cost.name = 'sequence_log_likelihood'
cg = ComputationGraph(cost)
model = Model(cost)
#################
# Algorithm
#################
n_batches = 139*16
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(10.0), Adam(lr)]))