def get_vae_svhn_wae_architecture(input_shape, embeding_loss_weight, generator_reg, generator_reg_type,
include_batch_norm, spec_norm_dec_only, recon_loss_func, verbose=True):
return rae_mnist.get_vae_mnist(input_shape, bottleneck_size=32,
embeding_loss_weight=embeding_loss_weight,
generator_reg=generator_reg, generator_reg_type=generator_reg_type,
include_batch_norm=include_batch_norm, num_filter=128,
spec_norm_dec_only=spec_norm_dec_only, recon_loss_func=recon_loss_func,
verbose=verbose)
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 get_vae_cifar(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=bottleneck_size,
embeding_loss_weight=embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=generator_reg_types,
include_batch_norm=include_batch_norm,
num_filter=num_filter,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
return build_vae_cifar(encoder, decoder, embeding_loss_weight,
recon_loss_func, constant_sigma)
def get_wae_mnist(input_shape, bottleneck_size, embeding_loss_weight,
generator_regs, generator_reg_types, include_batch_norm,
num_filter, spec_norm_dec_only, fully_convolutional,
batch_size, recon_loss_func):
if fully_convolutional:
raise NotImplementedError(
'Fully convolutional architecture for MNIST has not been implemented'
)
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)
return get_wae_given_enc_dec.get_wae(encoder, decoder,
embeding_loss_weight, batch_size,
recon_loss_func)
def get_vae(input_shape, config_dict, major_idx, minor_idx):
model_name = config_dict[major_idx][minor_idx]['model_name']
bottleneck_factor = config_dict[major_idx][minor_idx]['bottleneck_factor']
embeding_loss_weight = config_dict[major_idx][minor_idx][
'embedding_weight']
generator_regs = config_dict[major_idx][minor_idx]['gen_reg_weight']
reg_types = config_dict[major_idx][minor_idx]['gen_reg_type']
include_batch_norm = config_dict[major_idx][minor_idx][
'include_batch_norm']
spec_norm_dec_only = config_dict[major_idx][minor_idx][
'spec_norm_on_dec_only']
recon_loss_func = function_from_name.get_recon_loss_func(
config_dict[major_idx][minor_idx]['recon_loss_type'])
num_filters = config_dict[major_idx][minor_idx]['num_filters']
if model_name.upper() == "MNIST_WAE_PAPER":
return rae_mnist.get_vae_mnist_wae_architecture(
input_shape,
embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
if model_name.upper() == "MNIST_TINY":
return rae_mnist.get_vae_mnist_tiny_architecture(
input_shape,
bottleneck_factor,
embeding_loss_weight=embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "MNIST":
return rae_mnist.get_vae_mnist(
input_shape,
bottleneck_size=bottleneck_factor,
embeding_loss_weight=embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
num_filter=num_filters,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "CELEBA_WAE_PAPER":
return rae_celeba.get_vae_celeba_wae_architecture(
input_shape,
embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "CELEBA_WAE_PAPER_MAN_EMB_SZIE":
return rae_celeba.get_vae_cleba(
input_shape,
num_filter=num_filters,
embeding_loss_weight=embeding_loss_weight,
bottleneck_size=bottleneck_factor,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func,
fully_convolutional=False)
elif model_name.upper() == "CELEBA_FULLY_CONVOLUTIONAL":
return rae_celeba.get_celeba_fully_convolutional(
input_shape,
bottleneck_factor,
embeding_loss_weight,
num_filters=num_filters,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "CIFAR_WAE_PAPER":
return rae_cifar.get_vae_cifar_wae_architecture(
input_shape,
embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "CIFAR_TINY":
return rae_cifar.get_vae_cifar_tiny_architecture(
input_shape,
bottleneck_factor,
embeding_loss_weight=embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "CIFAR":
# Mnist model there is intentional. Only factor that is different is the bottleneck size
return rae_mnist.get_vae_mnist(
input_shape,
bottleneck_size=bottleneck_factor,
embeding_loss_weight=embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
num_filter=num_filters,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
elif model_name.upper() == "SVHN_WAE_PAPER":
return rae_svhn.get_vae_svhn_wae_architecture(
input_shape,
embeding_loss_weight,
generator_regs=generator_regs,
generator_reg_types=reg_types,
include_batch_norm=include_batch_norm,
spec_norm_dec_only=spec_norm_dec_only,
recon_loss_func=recon_loss_func)
else:
raise ValueError("Specified model: " + model_name +
" not implemented yet.")