def __init__(self, latent_dim=2, z_prior='gaussian'):
self.z_prior = z_prior
self.enc_l1 = L.Linear((784, 1000))
self.enc_b1 = L.BatchNormalization(1000)
self.enc_l2 = L.Linear((1000, 1000))
self.enc_b2 = L.BatchNormalization(1000)
self.enc_l3 = L.Linear((1000, latent_dim))
self.enc_b3 = L.BatchNormalization(latent_dim)
self.dec_l1 = L.Linear((latent_dim, 1000))
self.dec_b1 = L.BatchNormalization(1000)
self.dec_l2 = L.Linear((1000, 1000))
self.dec_b2 = L.BatchNormalization(1000)
self.dec_l3 = L.Linear((1000, 784))
self.D_l1 = L.Linear((latent_dim, 500))
self.D_b1 = L.BatchNormalization(500)
self.D_l2 = L.Linear((500, 500))
self.D_b2 = L.BatchNormalization(500)
self.D_l3 = L.Linear((500, 1))
self.model_params = self.enc_l1.params + self.enc_l2.params + self.enc_l3.params \
+ self.dec_l1.params + self.dec_l2.params + self.dec_l3.params \
+ self.enc_b1.params + self.enc_b2.params + self.enc_b3.params \
+ self.dec_b1.params + self.dec_b2.params
self.D_params = self.D_l1.params + self.D_l2.params + self.D_l3.params
self.rng = RandomStreams(seed=numpy.random.randint(1234))
def __init__(self, layer_sizes):
self.linear_layers = []
self.bn_layers = []
self.act_layers = []
self.params = []
for m, n in zip(layer_sizes[:-1], layer_sizes[1:]):
l = L.Linear(size=(m, n))
bn = L.BatchNormalization(size=(n))
self.linear_layers.append(l)
self.bn_layers.append(bn)
self.params += l.params + bn.params
for i in range(len(self.linear_layers) - 1):
self.act_layers.append(L.relu)
self.act_layers.append(L.softmax)
def __init__(self):
self.l1 = L.Linear(size=(100, 100))
self.l2 = L.Linear(size=(100, 2))
self.params = self.l1.params + self.l2.params
def __init__(self, drate=0.5):
self.drate = drate
self.l1 = L.Linear(size=(100, 100))
self.l2 = L.Linear(size=(100, 2))
self.params = self.l1.params + self.l2.params