def __init__(self, input_shape, latent_space_dim, have_2nd_density_est,
log_dir, sec_stg_beta):
self.log_dir = log_dir
self.latent_space_dim = latent_space_dim
self.have_2nd_density_est = have_2nd_density_est
self.sec_stg_beta = sec_stg_beta
with tf.name_scope('Encoder'):
e_in = Input(shape=input_shape)
x = Dense(1024, activation='relu')(e_in)
x = Dense(1024, activation='relu')(x)
z = Dense(latent_space_dim, activation='linear')(x)
encoder = Model(inputs=e_in, outputs=z)
layer_for_z_sigma = Dense(latent_space_dim, activation='tanh')
with tf.name_scope('Decoder'):
d_in = Input(shape=(latent_space_dim, ))
x = Dense(1024, activation='relu')(d_in)
x = Dense(1024, activation='relu')(x)
d_out = Dense(input_shape[0], activation='linear')(x)
decoder = Model(inputs=d_in, outputs=d_out)
self.encoder, self.decoder, self.auto_encoder = get_vae_given_enc_dec.get_vae(
encoder,
decoder,
embeding_loss_weight=self.sec_stg_beta,
layer_for_z_sigma=layer_for_z_sigma,
recon_loss_func=mean_squared_error,
constant_sigma=None)
def get_vae_mnist(input_shape, bottleneck_size, embeding_loss_weight, generator_regs, generator_reg_types,
include_batch_norm, num_filter, spec_norm_dec_only, recon_loss_func, constant_sigma):
encoder, decoder, _ = rae_mnist.get_vae_mnist(input_shape, bottleneck_size, embeding_loss_weight,
generator_regs, generator_reg_types,
include_batch_norm, num_filter, spec_norm_dec_only,
recon_loss_func=recon_loss_func, verbose=False)
layer_for_z_sigma = Dense(bottleneck_size, activation='tanh', name='log_sigma')
return get_vae_given_enc_dec.get_vae(encoder, decoder, embeding_loss_weight, layer_for_z_sigma, recon_loss_func,
constant_sigma)
def build_vae_cifar(encoder, decoder, embeding_loss_weight, recon_loss_func,
constant_sigma):
bottleneck_size = K.get_variable_shape(encoder.outputs[0])[-1]
layer_for_z_sigma = Dense(bottleneck_size,
activation='tanh',
name='log_sigma')
return get_vae_given_enc_dec.get_vae(encoder, decoder,
embeding_loss_weight,
layer_for_z_sigma, recon_loss_func,
constant_sigma)
def get_vae_svhn(input_shape, embeding_loss_weight, generator_regs,
generator_reg_types, include_batch_norm, spec_norm_dec_only,
recon_loss_func):
encoder, decoder, _ = rae_svhn.get_vae_svhn_wae_architecture(
input_shape,
embeding_loss_weight,
generator_regs,
generator_reg_types,
include_batch_norm,
spec_norm_dec_only,
recon_loss_func=recon_loss_func,
verbose=False)
bottleneck_size = K.get_variable_shape(encoder.outputs[0])[-1]
layer_for_z_sigma = Dense(bottleneck_size,
activation='tanh',
name='log_sigma')
return get_vae_given_enc_dec.get_vae(encoder, decoder,
embeding_loss_weight,
layer_for_z_sigma, recon_loss_func)