def __init__(self, learning_rate, hidden_dims, label=''):
super().__init__()
self.label = label
self.learning_rate = learning_rate
self.hidden_dims = hidden_dims
self.weight_std = 10**-3
self.sigma_1_prior = Variable(torch.Tensor([0.001]), requires_grad=False)
self.sigma_2_prior = Variable(torch.Tensor([10**-7]), requires_grad=False)
self.prior_weight = 0.5
self.num_samples = 1
# layers
self.W1_mu = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W1_rho = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W2_mu = nn.Parameter(xavier_init((self.hidden_dims[0], 128*7*7)).type(
torch.FloatTensor), requires_grad=True)
self.W2_rho = nn.Parameter(xavier_init((self.hidden_dims[0], 128*7*7)).type(
torch.FloatTensor),
requires_grad=True)
# self.W3_mu = nn.Parameter(xavier_init((self.hidden_dims[1], 32**2*3)).type(torch.FloatTensor),
# requires_grad=True)
# self.W3_rho = nn.Parameter(xavier_init((self.hidden_dims[1], 32**2*3)).type(torch.FloatTensor),
# requires_grad=True)
# # deconv layers
self.deconv1 = nn.ConvTranspose2d(128, 64, 4, stride=2, padding=1)
self.deconv2 = nn.ConvTranspose2d(64, 1, 4, stride=2, padding=1)
# batch norm
self.batch_norm1 = nn.BatchNorm1d(self.hidden_dims[0])
self.batch_norm2 = nn.BatchNorm1d(128*7*7)
self.batch_norm3 = nn.BatchNorm2d(64)
def __init__(self, learning_rate, hidden_dims, label=''):
super().__init__()
self.label = label
self.learning_rate = learning_rate
self.hidden_dims = hidden_dims
self.weight_std = 10**-4
self.sigma_1_prior = Variable(torch.Tensor([0.001]), requires_grad=False)
self.sigma_2_prior = Variable(torch.Tensor([10**-7]), requires_grad=False)
self.prior_weight = 0.5
self.num_samples = 3
# layers
self.W1_mu = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W1_rho = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W2_mu = nn.Parameter(xavier_init((self.hidden_dims[0], 128*7*7)).type(
torch.FloatTensor), requires_grad=True)
self.W2_rho = nn.Parameter(xavier_init((self.hidden_dims[0], 128*7*7)).type(
torch.FloatTensor),
requires_grad=True)
# self.W3_mu = nn.Parameter(xavier_init((self.hidden_dims[1], 32**2*3)).type(torch.FloatTensor),
# requires_grad=True)
# self.W3_rho = nn.Parameter(xavier_init((self.hidden_dims[1], 32**2*3)).type(torch.FloatTensor),
# requires_grad=True)
# # deconv layers
self.W1_deconv_mu = nn.Parameter(xavier_init((128, 64, 6, 6)).type(torch.FloatTensor),
requires_grad=True)
self.W1_deconv_rho = nn.Parameter(xavier_init((128, 64, 6, 6)).type(
torch.FloatTensor), requires_grad=True)
self.W2_deconv_mu = nn.Parameter(xavier_init((64, 3, 4, 4)).type(torch.FloatTensor),
requires_grad=True)
self.W2_deconv_rho = nn.Parameter(xavier_init((64, 3, 4, 4)).type(torch.FloatTensor),
requires_grad=True)
def __init__(self, learning_rate, hidden_dims, label=''):
super().__init__()
self.label = label
self.learning_rate = learning_rate
self.hidden_dims = hidden_dims
self.weight_std = 10**-3
self.sigma_1_prior = Variable(torch.Tensor([0.001]), requires_grad=False)
self.sigma_2_prior = Variable(torch.Tensor([10**-7]), requires_grad=False)
self.prior_weight = 0.5
self.num_samples = 3
# convlutional layers
self.W1_conv_mu = nn.Parameter(xavier_init((32, 1, 5, 5)).type(torch.FloatTensor),
requires_grad=True)
self.W1_conv_rho = nn.Parameter(xavier_init((32, 1, 5, 5)).type(
torch.FloatTensor), requires_grad=True)
self.W2_conv_mu = nn.Parameter(xavier_init((64, 32, 5, 5)).type(torch.FloatTensor),
requires_grad=True)
self.W2_conv_rho = nn.Parameter(xavier_init((64, 32, 5, 5)).type(torch.FloatTensor),
requires_grad=True)
# linear layers
self.W1_mu = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(
torch.FloatTensor), requires_grad=True)
self.W1_rho = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(
torch.FloatTensor), requires_grad=True)
# self.W2_mu = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(torch.FloatTensor), requires_grad=True)
# self.W2_rho = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(torch.FloatTensor), requires_grad=True)
self.W3_mu = nn.Parameter(xavier_init((self.hidden_dims[1], 1)).type(torch.FloatTensor),
requires_grad=True)
self.W3_rho = nn.Parameter(xavier_init((self.hidden_dims[1], 1)).type(torch.FloatTensor),
requires_grad=True)
def __init__(self, learning_rate, hidden_dims, label=''):
super().__init__()
self.label = label
self.learning_rate = learning_rate
self.hidden_dims = hidden_dims
self.weight_std = 10**-3
self.sigma_1_prior = Variable(torch.Tensor([0.001]), requires_grad=False)
self.sigma_2_prior = Variable(torch.Tensor([10**-7]), requires_grad=False)
self.prior_weight = 0.5
self.num_samples = 2
# layers
self.W1_mu = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W1_rho = nn.Parameter(xavier_init((noise_dim, self.hidden_dims[0])).type(torch.FloatTensor), requires_grad=True)
self.W2_mu = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(torch.FloatTensor), requires_grad=True)
self.W2_rho = nn.Parameter(xavier_init((self.hidden_dims[0], self.hidden_dims[1])).type(torch.FloatTensor), requires_grad=True)
self.W3_mu = nn.Parameter(xavier_init((self.hidden_dims[1], 28**2)).type(
torch.FloatTensor),
requires_grad=True)
self.W3_rho = nn.Parameter(xavier_init((self.hidden_dims[1], 28**2)).type(
torch.FloatTensor),
requires_grad=True)
# layer bias
self.b1_mu = nn.Parameter(torch.zeros((self.hidden_dims[0],)).type(torch.FloatTensor),
requires_grad=True)
self.b1_rho = nn.Parameter(torch.zeros((self.hidden_dims[0],)).type(torch.FloatTensor),
requires_grad=True)
self.b2_mu = nn.Parameter(torch.zeros((self.hidden_dims[1],)).type(torch.FloatTensor), requires_grad=True)
self.b2_rho = nn.Parameter(torch.zeros((self.hidden_dims[1],)).type(torch.FloatTensor), requires_grad=True)
self.b3_mu = nn.Parameter(torch.zeros((28**2,)).type(torch.FloatTensor), requires_grad=True)
self.b3_rho = nn.Parameter(torch.zeros((28**2,)).type(torch.FloatTensor), requires_grad=True)