def __init__(self, nn='v1', name='ssvae', gen_weight=1, class_weight=100):
super().__init__()
self.name = name
self.z_dim = 64
self.y_dim = 10
self.gen_weight = gen_weight
self.class_weight = class_weight
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, self.y_dim)
self.dec = nn.Decoder(self.z_dim, self.y_dim)
self.cls = nn.Classifier(self.y_dim)
# Set prior as fixed parameter attached to Module
self.z_prior_m = torch.nn.Parameter(torch.zeros(1), requires_grad=False)
self.z_prior_v = torch.nn.Parameter(torch.ones(1), requires_grad=False)
self.z_prior = (self.z_prior_m, self.z_prior_v)
def __init__(self,
nn='v1',
name='ssvae',
rec_weight=1,
kl_xy_x_weight=10,
kl_xy_y_weight=10,
gen_weight=1,
class_weight=100,
CNN=False):
super().__init__()
self.name = name
self.CNN = CNN
self.x_dim = 784
self.z_dim = 64
self.y_dim = 10
self.rec_weight = rec_weight
self.kl_xy_x_weight = kl_xy_x_weight
self.kl_xy_y_weight = kl_xy_y_weight
self.gen_weight = gen_weight
self.class_weight = class_weight
nn = getattr(nns, nn)
if CNN:
self.enc_xy = nn.Encoder_XY(z_dim=self.z_dim, y_dim=self.y_dim)
self.enc_x = nn.Encoder_X(z_dim=self.z_dim)
self.enc_y = nn.Encoder_Y(z_dim=self.z_dim, y_dim=self.y_dim)
else:
self.enc_xy = nn.Encoder(z_dim=self.z_dim,
y_dim=self.y_dim,
x_dim=self.x_dim)
self.enc_x = nn.Encoder(z_dim=self.z_dim,
y_dim=0,
x_dim=self.x_dim)
self.enc_y = nn.Encoder(z_dim=self.z_dim,
y_dim=self.y_dim,
x_dim=0)
self.dec = nn.Decoder(z_dim=self.z_dim, y_dim=0, x_dim=self.x_dim)
self.cls = nn.Classifier(y_dim=self.y_dim, input_dim=self.z_dim * 2)