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(
Merge(num_inputs=2, out_size=500,
use_weightnorm=config.use_weightnorm))
config.weight_decay = 0
config.use_feature_matching = False
config.use_minibatch_discrimination = False
discriminator = Sequential(weight_initializer=config.weight_initializer,
weight_init_std=config.weight_init_std)
discriminator.add(gaussian_noise(std=0.3))
discriminator.add(
Convolution2D(3,
32,
ksize=4,
stride=2,
pad=1,
use_weightnorm=config.use_weightnorm))
discriminator.add(BatchNormalization(32))
discriminator.add(Activation(config.nonlinearity))
discriminator.add(
Convolution2D(32,
64,
ksize=4,
stride=2,
pad=1,
use_weightnorm=config.use_weightnorm))
discriminator.add(BatchNormalization(64))
discriminator.add(Activation(config.nonlinearity))
discriminator.add(
Convolution2D(64,
128,
ksize=4,
stride=2,
pad=1,
config.weight_initializer = "Normal"
config.use_weightnorm = False
config.nonlinearity = "leaky_relu"
config.optimizer = "rmsprop"
config.learning_rate = 0.0001
config.momentum = 0.5
config.gradient_clipping = 1
config.weight_decay = 0
config.use_feature_matching = False
config.use_minibatch_discrimination = False
discriminator = Sequential(weight_initializer=config.weight_initializer,
weight_init_std=config.weight_init_std)
discriminator.add(Linear(None, 500, use_weightnorm=config.use_weightnorm))
# discriminator.add(gaussian_noise(std=0.5))
discriminator.add(Activation(config.nonlinearity))
# discriminator.add(BatchNormalization(500))
if config.use_minibatch_discrimination:
discriminator.add(
MinibatchDiscrimination(None, num_kernels=50, ndim_kernel=5))
discriminator.add(Linear(None, 500, use_weightnorm=config.use_weightnorm))
params = {
"config": config.to_dict(),
"model": discriminator.to_dict(),
}
with open(discriminator_sequence_filename, "w") as f:
json.dump(params, f, indent=4, sort_keys=True, separators=(',', ': '))
discriminator_params = params
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))
# discriminator_z.add(BatchNormalization(1000))
else:
config = Params()
config.num_classes = 10
config.weight_std = 0.1
config.weight_initializer = "Normal"
config.nonlinearity = "relu"
config.optimizer = "adam"
config.learning_rate = 0.0001
config.momentum = 0.9
config.gradient_clipping = 1
config.weight_decay = 0
model = Sequential()
model.add(Convolution2D(1, 32, ksize=4, stride=2, pad=1))
model.add(BatchNormalization(32))
model.add(Activation(config.nonlinearity))
model.add(Convolution2D(32, 64, ksize=4, stride=2, pad=1))
model.add(BatchNormalization(64))
model.add(Activation(config.nonlinearity))
model.add(Convolution2D(64, 128, ksize=3, stride=2, pad=1))
model.add(BatchNormalization(128))
model.add(Activation(config.nonlinearity))
model.add(Linear(None, config.num_classes))
params = {
"config": config.to_dict(),
"model": model.to_dict(),
}
with open(sequence_filename, "w") as f:
json.dump(params, f, indent=4, sort_keys=True, separators=(',', ': '))
config.ndim_input = image_width * image_height config.num_experts = 128 config.weight_init_std = 0.05 config.weight_initializer = "Normal" config.use_weightnorm = True config.nonlinearity = "elu" config.optimizer = "Adam" config.learning_rate = 0.0002 config.momentum = 0.5 config.gradient_clipping = 10 config.weight_decay = 0 # feature extractor feature_extractor = Sequential(weight_initializer=config.weight_initializer, weight_init_std=config.weight_init_std) feature_extractor.add(Linear(config.ndim_input, 1000, use_weightnorm=config.use_weightnorm)) feature_extractor.add(Activation(config.nonlinearity)) feature_extractor.add(gaussian_noise(std=0.3)) feature_extractor.add(Linear(None, 500, use_weightnorm=config.use_weightnorm)) feature_extractor.add(Activation(config.nonlinearity)) feature_extractor.add(gaussian_noise(std=0.3)) feature_extractor.add(Linear(None, 250, use_weightnorm=config.use_weightnorm)) feature_extractor.add(Activation(config.nonlinearity)) feature_extractor.add(gaussian_noise(std=0.3)) feature_extractor.add(Linear(None, config.num_experts, use_weightnorm=config.use_weightnorm)) feature_extractor.add(tanh()) # experts experts = Sequential(weight_initializer=config.weight_initializer, weight_init_std=config.weight_init_std) experts.add(Linear(config.num_experts, config.num_experts, use_weightnorm=config.use_weightnorm)) # b
config.weight_std = 0.01
config.weight_initializer = "Normal"
config.nonlinearity_d = "elu"
config.nonlinearity_g = "elu"
config.optimizer = "adam"
config.learning_rate = 0.0001
config.momentum = 0.5
config.gradient_clipping = 1
config.weight_decay = 0
# Discriminator
encoder = Sequential()
encoder.add(gaussian_noise(std=0.3))
encoder.add(Convolution2D(3, 32, ksize=4, stride=2, pad=1))
encoder.add(BatchNormalization(32))
encoder.add(Activation(config.nonlinearity_d))
encoder.add(Convolution2D(32, 64, ksize=4, stride=2, pad=1))
encoder.add(BatchNormalization(64))
encoder.add(Activation(config.nonlinearity_d))
encoder.add(Convolution2D(64, 128, ksize=4, stride=2, pad=1))
encoder.add(BatchNormalization(128))
encoder.add(Activation(config.nonlinearity_d))
encoder.add(Convolution2D(128, 256, ksize=4, stride=2, pad=1))
encoder.add(BatchNormalization(256))
encoder.add(Activation(config.nonlinearity_d))
encoder.add(Linear(None, ndim_h))
projection_size = 6
# Decoder
decoder = Sequential()