def __init__(self, ndim_x=28 * 28, ndim_y=10, ndim_z=10, ndim_h=1000):
super(Model, self).__init__()
self.ndim_x = ndim_x
self.ndim_y = ndim_y
self.ndim_z = ndim_z
self.ndim_h = ndim_h
with self.init_scope():
decoder = nn.Module(
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, ndim_x),
nn.Tanh(),
)
decoder.merge_y = nn.Linear(ndim_y, ndim_h, nobias=True)
decoder.merge_z = nn.Linear(ndim_z, ndim_h, nobias=True)
decoder.merge_bias = nn.Bias(shape=(ndim_h, ))
self.decoder = decoder
encoder = nn.Module(
nn.Linear(ndim_x, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, ndim_h),
nn.ReLU(),
)
encoder.head_y = nn.Linear(ndim_h, ndim_y)
encoder.head_z = nn.Linear(ndim_h, ndim_z)
self.encoder = encoder
self.discriminator_z = nn.Module(
nn.GaussianNoise(std=0.3),
nn.Linear(ndim_z, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, 2),
)
self.discriminator_y = nn.Module(
nn.GaussianNoise(std=0.3),
# nn.GaussianNoise(0, 0.3),
nn.Linear(ndim_y, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, ndim_h),
nn.ReLU(),
# nn.BatchNormalization(ndim_h),
nn.Linear(ndim_h, 2),
)
for param in self.params():
if param.name == "W":
param.data[...] = np.random.normal(0, 0.01, param.data.shape)
def __init__(self, ndim_x=28*28, ndim_y=20, ndim_z=2, ndim_h=1000, cluster_head_distance_threshold=2): super(Model, self).__init__() self.ndim_x = ndim_x self.ndim_y = ndim_y self.ndim_z = ndim_z self.ndim_h = ndim_h self.cluster_head_distance_threshold = cluster_head_distance_threshold with self.init_scope(): self.decoder = nn.Module( nn.Linear(ndim_z, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, ndim_x), nn.Tanh(), ) encoder = nn.Module( nn.Linear(ndim_x, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, ndim_h), nn.ReLU(), ) encoder.head_y = nn.Linear(ndim_h, ndim_y) encoder.head_z = nn.Linear(ndim_h, ndim_z) self.encoder = encoder self.discriminator_z = nn.Module( nn.GaussianNoise(std=0.3), nn.Linear(ndim_z, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, 2), ) self.discriminator_y = nn.Module( nn.GaussianNoise(std=0.3), # nn.GaussianNoise(0, 0.3), nn.Linear(ndim_y, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, ndim_h), nn.ReLU(), # nn.BatchNormalization(ndim_h), nn.Linear(ndim_h, 2), ) self.cluster_head = nn.Linear(ndim_y, ndim_z, nobias=True) for param in self.params(): if param.name == "W": param.data[...] = np.random.normal(0, 0.01, param.data.shape) for param in self.cluster_head.params(): if param.name == "W": param.data[...] = np.random.normal(0, 1, param.data.shape)
