config.optimizer = "Adam" config.learning_rate = 0.0002 config.momentum = 0.5 config.gradient_clipping = 5 config.weight_decay = 0 decoder = Sequential(weight_initializer=config.weight_initializer, weight_init_std=config.weight_init_std) decoder.add(Linear(None, 1000)) decoder.add(Activation(config.nonlinearity)) decoder.add(Linear(None, 1000)) decoder.add(Activation(config.nonlinearity)) decoder.add(Linear(None, config.ndim_x)) decoder.add(sigmoid()) discriminator = Sequential(weight_initializer=config.weight_initializer, weight_init_std=config.weight_init_std) discriminator.add(Merge(num_inputs=2, out_size=1000, nobias=True)) discriminator.add(gaussian_noise(std=0.3)) discriminator.add(Activation(config.nonlinearity)) discriminator.add(Linear(None, 1000)) discriminator.add(Activation(config.nonlinearity)) discriminator.add(Linear(None, 1000)) discriminator.add(Activation(config.nonlinearity)) discriminator.add(Linear(None, 2)) generator = Sequential(weight_initializer=config.weight_initializer, weight_init_std=config.weight_init_std) generator.add(Linear(config.ndim_x, 1000)) generator.add(Activation(config.nonlinearity)) generator.add(Linear(None, 1000)) generator.add(Activation(config.nonlinearity)) if config.distribution_z == "deterministic": generator.add(Linear(None, config.ndim_z))
config.weight_init_std = 0.01
config.weight_initializer = "Normal"
config.nonlinearity = "elu"
config.optimizer = "Adam"
config.learning_rate = 0.0003
config.momentum = 0.9
config.gradient_clipping = 10
config.weight_decay = 0
config.use_weightnorm = False
config.num_mc_samples = 1
# p(x|y,z) - x ~ Bernoulli
p_x_ayz = Sequential(weight_initializer=config.weight_initializer,
weight_init_std=config.weight_init_std)
p_x_ayz.add(
Merge(num_inputs=3, out_size=500,
use_weightnorm=config.use_weightnorm))
p_x_ayz.add(BatchNormalization(500))
p_x_ayz.add(Activation(config.nonlinearity))
p_x_ayz.add(Linear(None, 500, use_weightnorm=config.use_weightnorm))
p_x_ayz.add(BatchNormalization(500))
p_x_ayz.add(Activation(config.nonlinearity))
p_x_ayz.add(Linear(None, 500, use_weightnorm=config.use_weightnorm))
p_x_ayz.add(BatchNormalization(500))
p_x_ayz.add(Activation(config.nonlinearity))
p_x_ayz.add(
Linear(None, config.ndim_x, use_weightnorm=config.use_weightnorm))
# p(a|x,y,z) - a ~ Gaussian
p_a_yz = Sequential(weight_initializer=config.weight_initializer,
weight_init_std=config.weight_init_std)
p_a_yz.add(
config.ndim_x = 28 * 28
config.ndim_y = 10
config.ndim_z = 10
config.distribution_z = "deterministic" # deterministic or gaussian
config.weight_std = 0.01
config.weight_initializer = "Normal"
config.nonlinearity = "relu"
config.optimizer = "Adam"
config.learning_rate = 0.0001
config.momentum = 0.1
config.gradient_clipping = 5
config.weight_decay = 0
# x = decoder(y, z)
decoder = Sequential()
decoder.add(Merge(num_inputs=2, out_size=1000, nobias=True))
decoder.add(Activation(config.nonlinearity))
# decoder.add(BatchNormalization(1000))
decoder.add(Linear(None, 1000))
decoder.add(Activation(config.nonlinearity))
# decoder.add(BatchNormalization(1000))
decoder.add(Linear(None, 1000))
decoder.add(Activation(config.nonlinearity))
# decoder.add(BatchNormalization(1000))
decoder.add(Linear(None, config.ndim_x))
decoder.add(tanh())
discriminator_z = Sequential()
discriminator_z.add(gaussian_noise(std=0.3))
discriminator_z.add(Linear(config.ndim_z, 1000))
discriminator_z.add(Activation(config.nonlinearity))