def __init__(self,
nn='v1',
name='vae2',
z_dim=10,
x_dim=24,
c_dim=0,
warmup=False,
var_pen=1,
y_dim=0):
super().__init__()
# print('ladedah')
# print('x_dim', x_dim)
# print('y_dim', y_dim)
self.name = name
self.z_dim = z_dim
self.x_dim = x_dim
self.y_dim = y_dim
self.c_dim = c_dim
self.warmup = warmup
self.var_pen = var_pen
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, self.x_dim, self.y_dim, self.c_dim)
self.dec = nn.Decoder(self.z_dim, self.x_dim, self.y_dim, self.c_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='mask',
name='vae',
z_dim=16,
z1_dim=4,
z2_dim=4,
inference=False):
super().__init__()
self.name = name
self.z_dim = z_dim
self.z1_dim = z1_dim
self.z2_dim = z2_dim
self.channel = 4
self.scale = np.array([[20, 15], [2, 2], [59.5, 26.5], [10.5, 4.5]])
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, self.channel)
self.dec = nn.Decoder_DAG(self.z_dim, self.z1_dim, self.z2_dim)
self.dag = nn.DagLayer(self.z1_dim, self.z1_dim, i=inference)
self.attn = nn.Attention(self.z2_dim)
self.mask_z = nn.MaskLayer(self.z_dim, concept=self.z1_dim)
self.mask_u = nn.MaskLayer(self.z1_dim, concept=self.z1_dim, z1_dim=1)
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='vae',
z_dim=2,
x_dim=24,
warmup=False,
var_pen=1):
super().__init__()
self.name = name
self.z_dim = z_dim
self.x_dim = x_dim
self.warmup = warmup
self.var_pen = var_pen
# Small note: unfortunate name clash with torch.nn
# nn here refers to the specific architecture file found in
# codebase/models/nns/*.py
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, self.x_dim)
self.dec = nn.Decoder(self.z_dim, self.x_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='vae',
z_dim=2,
z_prior_m=None,
z_prior_v=None):
super().__init__()
self.name = name
self.z_dim = z_dim
# Small note: unfortunate name clash with torch.nn
# nn here refers to the specific architecture file found in
# codebase/models/nns/*.py
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim)
self.dec = nn.Decoder(self.z_dim)
# Set prior as fixed parameter attached to Module
if z_prior_m is None:
self.z_prior_m = torch.nn.Parameter(torch.zeros(z_dim),
requires_grad=False)
else:
self.z_prior_m = z_prior_m
if z_prior_v is None:
self.z_prior_v = torch.nn.Parameter(torch.ones(z_dim),
requires_grad=False)
else:
self.z_prior_v = z_prior_v
self.z_prior = (self.z_prior_m, self.z_prior_v)
def __init__(self, nn='v1', encode_dim=None, name='vae', z_dim=2):
super().__init__()
self.name = name
self.z_dim = z_dim
if nn == 'popv':
nn = getattr(nns, nn)
self.enc = nn.Encoder(encode_dim, self.z_dim)
self.dec = nn.Decoder(encode_dim, self.z_dim)
else:
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim)
self.dec = nn.Decoder(self.z_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)
def __init__(self, nn='v2', name='fsvae'):
super().__init__()
self.name = name
self.z_dim = 10
self.y_dim = 10
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, self.y_dim)
self.dec = nn.Decoder(self.z_dim, 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='v8', name='vae3d', z_dim=16,device ='cpu', lambda_kl =0.01):
super().__init__()
self.name = name
self.z_dim = z_dim
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim, device =device)
self.dec = nn.Decoder(self.z_dim, device =device)
self.lambda_kl = lambda_kl
# 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', z_dim=2, k=500, name='gmvae'):
super().__init__()
self.name = name
self.k = k
self.z_dim = z_dim
nn = getattr(nns, nn)
self.enc = nn.Encoder(self.z_dim)
self.dec = nn.Decoder(self.z_dim)
# Mixture of Gaussians prior
self.z_pre = torch.nn.Parameter(torch.randn(1, 2 * self.k, self.z_dim)
/ np.sqrt(self.k * self.z_dim))
# Uniform weighting
self.pi = torch.nn.Parameter(torch.ones(k) / k, requires_grad=False)
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)
