def _create_parameters(self):
"""
Instantiate the parameters of the network
:return:
"""
self.extra, self.extra_inf = OrderedDict(), OrderedDict()
# input layer
gin = self.layer(self.dimx + 1, self.dimh[0], 'in', priors=self.priors, N=self.N, nonlin=self.nonlinearity,
type_init=self.type_init, n_inducing=self.n_inducing, noise_lvl=self.ind_noise_lvl)
layers = [gin]
# remaining hidden layers
self.hidden_params = []
for i, h in enumerate(self.dimh[1:]):
gh = self.layer(self.dimh[i] + 1, h, 'h' + str(i + 1), priors=self.priors, N=self.N, nonlin=self.nonlinearity,
type_init=self.type_init, n_inducing=self.n_inducing, noise_lvl=self.ind_noise_lvl)
layers.append(gh)
gout = self.layer(self.dimh[-1] + 1, self.dimy, 'out', priors=self.priors, nonlin='linear',
type_init=self.type_init, N=self.N, n_inducing=self.n_inducing, noise_lvl=self.ind_noise_lvl)
layers.append(gout)
if self.task_type == 'regression':
a1, b1 = nnu.multvector(self.dimy, np.log(a0), 'out_a1'), nnu.multvector(self.dimy, np.log(b0), 'out_b1')
a1inf, b1inf = nnu.multvector(self.dimy, np.log(a0), 'out_a1_inf'), nnu.multvector(self.dimy, np.log(b0),
'out_b1_inf')
self.extra['a1'] = a1; self.extra['b1'] = b1
self.extra_inf['a1'] = a1inf; self.extra_inf['b1'] = b1inf
self.layers = layers
self.layers_inf = [copy(layer) for layer in layers]
for layeri in self.layers_inf:
layeri.N = self.N_valid
def _create_parameters(self):
"""
Instantiate the parameters of the network
:return:
"""
self.extra, self.extra_inf = OrderedDict(), OrderedDict()
# input layer
gin = self.layer(self.dimx + 1,
self.dimh[0],
'in',
priors=self.priors,
N=self.N,
nonlin=self.nonlinearity,
type_init=self.type_init,
n_inducing=self.n_inducing,
noise_lvl=self.ind_noise_lvl)
layers = [gin]
# remaining hidden layers
self.hidden_params = []
for i, h in enumerate(self.dimh[1:]):
gh = self.layer(self.dimh[i] + 1,
h,
'h' + str(i + 1),
priors=self.priors,
N=self.N,
nonlin=self.nonlinearity,
type_init=self.type_init,
n_inducing=self.n_inducing,
noise_lvl=self.ind_noise_lvl)
layers.append(gh)
gout = self.layer(self.dimh[-1] + 1,
self.dimy,
'out',
priors=self.priors,
nonlin='linear',
type_init=self.type_init,
N=self.N,
n_inducing=self.n_inducing,
noise_lvl=self.ind_noise_lvl)
layers.append(gout)
if self.task_type == 'regression':
a1, b1 = nnu.multvector(self.dimy, np.log(a0),
'out_a1'), nnu.multvector(
self.dimy, np.log(b0), 'out_b1')
a1inf, b1inf = nnu.multvector(self.dimy, np.log(a0),
'out_a1_inf'), nnu.multvector(
self.dimy, np.log(b0),
'out_b1_inf')
self.extra['a1'] = a1
self.extra['b1'] = b1
self.extra_inf['a1'] = a1inf
self.extra_inf['b1'] = b1inf
self.layers = layers
self.layers_inf = [copy(layer) for layer in layers]
for layeri in self.layers_inf:
layeri.N = self.N_valid
def __init__(self, dim_in, dim_out, name, priors=(0., 0., 0.), N=1, nonlin='relu', type_init='he2', n_inducing=50,
noise_lvl=0.01):
sigma_in = 0.01
params = [nnu.randmat(dim_in, dim_out, 'mu_' + name, type_init=type_init, type_dist='normal'),
nnu.randvector(dim_in, 'sigma_row_mgauss_' + name, sigma=sigma_in),
nnu.randvector(dim_out, 'sigma_col_mgauss_' + name, sigma=sigma_in),
nnu.randmat2(n_inducing, dim_in, 'inducing_x_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.randmat2(n_inducing, dim_out, 'inducing_y_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.multvector(dim_in, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_x_' + name),
nnu.multvector(dim_out, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_y_' + name),
nnu.multvector(n_inducing, np.log(np.sqrt(noise_lvl)), name='dropout_alpha_ffdrop_pd_' + name)]
self.dim_in = dim_in
self.dim_out = dim_out
self.name = name
self.type_init = type_init
self.n_inducing = n_inducing
self.noise_lvl = noise_lvl
super(MatrixGaussDiagLayerFF, self).__init__(params, N=N, nonlin=nonlin, priors=priors)
def __init__(self, dim_in, dim_out, name, priors=(0., 1., 1.), N=1, nonlin='relu', type_init='he2', n_inducing=50,
noise_lvl=0.01):
sigma_in = 0.01
self.a0, self.b0 = 1., .5
params = [nnu.randmat(dim_in, dim_out, 'mu_' + name, type_init=type_init),
nnu.randvector(dim_in, 'sigma_row_mgauss_' + name, sigma=sigma_in),
nnu.randvector(dim_out, 'sigma_col_mgauss_' + name, sigma=sigma_in),
nnu.randmat2(n_inducing, dim_in, 'inducing_x_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.randmat2(n_inducing, dim_out, 'inducing_y_' + name, sigma=sigma_in, type_dist='uniform'),
nnu.multvector(1, np.log(self.a0), 'row_a_' + name), nnu.multvector(1, np.log(self.b0), 'row_b_' + name),
nnu.multvector(1, np.log(self.a0), 'col_a_' + name), nnu.multvector(1, np.log(self.b0), 'col_b_' + name),
nnu.tscalar(np.log(np.sqrt(noise_lvl)), 'scalar_dropout_alpha_x_' + name),
nnu.tscalar(np.log(np.sqrt(noise_lvl)), 'scalar_dropout_alpha_y_' + name)]
self.dim_in = dim_in
self.dim_out = dim_out
self.name = name
self.type_init = type_init
self.n_inducing = n_inducing
self.noise_lvl = noise_lvl
super(MatrixGaussDiagLayerLearnP, self).__init__(params, N=N, nonlin=nonlin, priors=priors)
