def __init__(self,
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
data, # [NxP] matrix of observations
kernelType='ARD',
encoderType_qX='FreeForm2', # MLP', 'Kernel'.
encoderType_rX='FreeForm2', # MLP', 'Kernel'.
Xu_optimise=False,
numHiddenUnits_encoder=10,
numHiddenUnits_decoder=10,
numHiddenLayers_decoder=2,
continuous=True
):
SGPDV.__init__(self,
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
data, # [NxP] matrix of observations
kernelType=kernelType,
encoderType_qX=encoderType_qX,
encoderType_rX=encoderType_rX,
Xu_optimise=Xu_optimise,
numberOfEncoderHiddenUnits=numHiddenUnits_encoder
)
self.HU_decoder = numHiddenUnits_decoder
self.numHiddenLayers_decoder = numHiddenLayers_decoder
self.continuous = continuous
# Construct appropriately sized matrices to initialise theano shares
self.W_zh = sharedZeroMatrix(self.HU_decoder, self.Q, 'W_zh')
self.W_hy1 = sharedZeroMatrix(self.P, self.HU_decoder, 'W_hy')
self.b_zh = sharedZeroVector(self.HU_decoder, 'b_zh', broadcastable=(False,True))
self.b_hy1 = sharedZeroVector(self.P, 'b_zh', broadcastable=(False,True))
self.likelihoodVariables = [self.W_zh, self.W_hy1, self.b_zh, self.b_hy1]
if self.numHiddenLayers_decoder == 2:
self.W_hh = sharedZeroMatrix(self.HU_decoder, self.HU_decoder, 'W_hh')
self.b_hh = sharedZeroVector(self.HU_decoder, 'b_hh', broadcastable=(False,True))
self.likelihoodVariables.extend([self.W_hh, self.b_hh])
if self.continuous:
self.W_hy2 = sharedZeroMatrix(self.P, self.HU_decoder, 'W_hy2')
self.b_hy2 = sharedZeroVector(self.P, 'b_hy2', broadcastable=(False,True))
self.likelihoodVariables.extend([self.W_hy2, self.b_hy2])
self.gradientVariables.extend(self.likelihoodVariables)
# Keep track of bounds and gradients for post analysis
self.all_bounds = []
self.all_gradients = []
def __init__(self,
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
data, # [NxP] matrix of observations
kernelType='RBF',
encoderType_qX='FreeForm', # 'FreeForm', 'MLP', 'Kernel'.
encoderType_rX='FreeForm', # 'FreeForm', 'MLP', 'Kernel', 'NoEncoding'.
encoderType_ru='FreeForm', # 'FreeForm', 'MLP', 'NoEncoding'
z_optimise=False,
numHiddenUnits_encoder=0,
numHiddentUnits_decoder=10,
continuous=True
):
#self, numberOfInducingPoints, batchSize, dimX, dimZ, data, numHiddenUnits, kernelType_='RBF', continuous_=True, encode_qX=True,encode_rX=False, encode_ru=False, encoder_type='kernel' ):
SGPDV.__init__(self,
numberOfInducingPoints,
batchSize,
dimX,
dimZ,
data,
kernelType,
encoderType_qX, # 'FreeForm', 'MLP', 'Kernel'.
encoderType_rX, # 'FreeForm', 'MLP', 'Kernel', 'NoEncoding'.
encoderType_ru, # 'FreeForm', 'MLP', 'NoEncoding'
z_optimise,
numHiddenUnits_encoder
)
self.K = dimZ # max number of features
self.D = data.shape[1] # dimensionality of features
self.continuous = continuous
# Suitably sized zero matrices
K_D_mat = np.zeros((self.K,self.D), dtype=floatX)
K_D_D_ten = np.zeros((self.K,self.D,self.D), dtype=floatX)
K_2_mat = np.zeros((self.K,2), dtype=floatX)
N_K_mat = np.zeros((self.N,self.K), dtype=floatX)
scalar = np.zeros(1, dtype=floatX)
#self.A = th.shared( K_D_mat )
self.Phi_IBP = th.shared( K_D_D_ten, name='Phi_IBP')
self.phi_IBP = th.shared( K_D_mat, name='phi_IBP')
self.tau_IBP = th.shared( K_2_mat, name='tau_IBP')
self.mu_IBP = th.shared( N_K_mat, name='mu_IBP')
self.log_alpha_IBP = th.shared(scalar, name='log_alpha_IBP')
self.log_sigma_y = th.shared(scalar, name='log_sigma_y')
self.log_sigma_A = th.shared(scalar, name='log_sigma_A')
self.alpha_IBP = T.exp(self.log_alpha_IBP)
self.sigma_y = T.exp(self.log_sigma_y)
self.sigma_A = T.exp(self.log_sigma_A)
self.gradientVariables.extend([self.A,self.Phi_IBP,self.phi_IBP,self.tau_IBP,self.alpha_IBP])
self.z_IBP_samp = T.nnet.sigmoid(self.z)
def __init__(
self,
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
data, # [NxP] matrix of observations
kernelType='ARD',
encoderType_qX='FreeForm2', # MLP', 'Kernel'.
encoderType_rX='FreeForm2', # MLP', 'Kernel'.
Xu_optimise=False,
numHiddenUnits_encoder=10,
numHiddenUnits_decoder=10,
numHiddenLayers_decoder=2,
continuous=True):
SGPDV.__init__(
self,
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
data, # [NxP] matrix of observations
kernelType=kernelType,
encoderType_qX=encoderType_qX,
encoderType_rX=encoderType_rX,
Xu_optimise=Xu_optimise,
numberOfEncoderHiddenUnits=numHiddenUnits_encoder)
self.HU_decoder = numHiddenUnits_decoder
self.numHiddenLayers_decoder = numHiddenLayers_decoder
self.continuous = continuous
# Construct appropriately sized matrices to initialise theano shares
self.W_zh = sharedZeroMatrix(self.HU_decoder, self.Q, 'W_zh')
self.W_hy1 = sharedZeroMatrix(self.P, self.HU_decoder, 'W_hy')
self.b_zh = sharedZeroVector(self.HU_decoder,
'b_zh',
broadcastable=(False, True))
self.b_hy1 = sharedZeroVector(self.P,
'b_zh',
broadcastable=(False, True))
self.likelihoodVariables = [
self.W_zh, self.W_hy1, self.b_zh, self.b_hy1
]
if self.numHiddenLayers_decoder == 2:
self.W_hh = sharedZeroMatrix(self.HU_decoder, self.HU_decoder,
'W_hh')
self.b_hh = sharedZeroVector(self.HU_decoder,
'b_hh',
broadcastable=(False, True))
self.likelihoodVariables.extend([self.W_hh, self.b_hh])
if self.continuous:
self.W_hy2 = sharedZeroMatrix(self.P, self.HU_decoder, 'W_hy2')
self.b_hy2 = sharedZeroVector(self.P,
'b_hy2',
broadcastable=(False, True))
self.likelihoodVariables.extend([self.W_hy2, self.b_hy2])
self.gradientVariables.extend(self.likelihoodVariables)
# Keep track of bounds and gradients for post analysis
self.all_bounds = []
self.all_gradients = []