def __init__(self, config={}, **kwargs):
TorchNNRepresentation.__init__(
self, config=config,
**kwargs) # calls all constructors up to BaseDNN (MRO)
self.output_keys_list = self.network.encoder.output_keys_list + [
"recon_x"
]
def set_network(self, network_name, network_parameters):
TorchNNRepresentation.set_network(self, network_name,
network_parameters)
# add a decoder to the network for the BiGAN
decoder_class = decoders.get_decoder(network_name)
self.network.decoder = decoder_class(config=network_parameters)
# add a discriminator to the network for the BiGAN
discriminator_class = discriminators.get_discriminator(network_name)
self.network.discriminator = discriminator_class(
config=network_parameters)
def set_logger(self, logger_config):
TorchNNRepresentation.set_logger(self, logger_config)
# Save the graph in the logger
if self.logger is not None:
dummy_size = (
1,
self.config.network.parameters.n_channels,
) + self.config.network.parameters.input_size
dummy_input = torch.Tensor(size=dummy_size).uniform_(0, 1).type(
self.config.dtype).to(self.config.device)
self.eval()
with torch.no_grad():
self.logger.add_graph(self, dummy_input, verbose=False)
def default_config():
default_config = TorchNNRepresentation.default_config()
# network parameters
default_config.network = Dict()
default_config.network.name = "Burgess"
default_config.network.parameters = Dict()
default_config.network.parameters.n_channels = 1
default_config.network.parameters.input_size = (64, 64)
default_config.network.parameters.n_latents = 10
default_config.network.parameters.n_conv_layers = 4
default_config.network.parameters.feature_layer = 2
default_config.network.parameters.encoder_conditional_type = "gaussian"
# initialization parameters
default_config.network.weights_init = Dict()
default_config.network.weights_init.name = "pytorch"
default_config.network.weights_init.parameters = Dict()
# loss parameters
default_config.loss = Dict()
default_config.loss.name = "SimCLR"
default_config.loss.parameters = Dict()
default_config.loss.parameters.temperature = 0.5
default_config.loss.parameters.distance = 'cosine'
# optimizer parameters
default_config.optimizer = Dict()
default_config.optimizer.name = "Adam"
default_config.optimizer.parameters = Dict()
default_config.optimizer.parameters.lr = 1e-3
default_config.optimizer.parameters.weight_decay = 1e-5
return default_config
def default_config():
default_config = TorchNNRepresentation.default_config()
# network parameters
default_config.network = Dict()
default_config.network.name = "Burgess"
default_config.network.parameters = Dict()
default_config.network.parameters.n_channels = 1
default_config.network.parameters.input_size = (64, 64)
default_config.network.parameters.n_latents = 10
default_config.network.parameters.n_conv_layers = 4
default_config.network.parameters.feature_layer = 2
default_config.network.parameters.encoder_conditional_type = "gaussian"
# initialization parameters
default_config.network.initialization = Dict()
default_config.network.initialization.name = "pytorch"
default_config.network.initialization.parameters = Dict()
# loss parameters
default_config.loss = Dict()
default_config.loss.name = "Triplet"
default_config.loss.parameters = Dict()
default_config.loss.parameters.distance = "squared_euclidean"
default_config.loss.parameters.margin = 1.0
default_config.loss.parameters.use_attention = True
# optimizer parameters
default_config.optimizer = Dict()
default_config.optimizer.name = "Adam"
default_config.optimizer.parameters = Dict()
default_config.optimizer.parameters.lr = 1e-3
default_config.optimizer.parameters.weight_decay = 1e-5
return default_config
def default_config():
default_config = TorchNNRepresentation.default_config()
# network parameters
default_config.network = Dict()
default_config.network.name = "Dumoulin"
default_config.network.parameters = Dict()
default_config.network.parameters.n_channels = 1
default_config.network.parameters.input_size = (64, 64)
default_config.network.parameters.n_latents = 10
default_config.network.parameters.n_conv_layers = 4
default_config.network.parameters.feature_layer = 2
default_config.network.parameters.encoder_conditional_type = "gaussian"
# initialization parameters
default_config.network.initialization = Dict()
default_config.network.initialization.name = "pytorch"
default_config.network.initialization.parameters = Dict()
# loss parameters
default_config.loss = Dict()
default_config.loss.name = "BiGAN"
default_config.loss.parameters = Dict()
# optimizer parameters
default_config.optimizer = Dict()
default_config.optimizer.name = "Adam"
default_config.optimizer.parameters = Dict()
default_config.optimizer.parameters.lr = 1e-3
default_config.optimizer.parameters.weight_decay = 1e-5
return default_config
def default_config():
default_config = TorchNNRepresentation.default_config()
# network parameters
default_config.network = Dict()
default_config.network.name = "Burgess"
default_config.network.parameters = Dict()
default_config.network.parameters.n_channels = 1
default_config.network.parameters.input_size = (64, 64)
default_config.network.parameters.n_latents = 10
default_config.network.parameters.n_conv_layers = 4
default_config.network.parameters.feature_layer = 2
default_config.network.parameters.encoder_conditional_type = "gaussian"
# weights_init parameters
default_config.network.weights_init = Dict()
default_config.network.weights_init.name = "pytorch"
default_config.network.weights_init.parameters = Dict()
# loss parameters
default_config.loss = Dict()
default_config.loss.name = "VAE"
default_config.loss.parameters = Dict()
default_config.loss.parameters.reconstruction_dist = "bernoulli"
# optimizer parameters
default_config.optimizer = Dict()
default_config.optimizer.name = "Adam"
default_config.optimizer.parameters = Dict()
default_config.optimizer.parameters.lr = 1e-3
default_config.optimizer.parameters.weight_decay = 1e-5
return default_config
def __init__(self, config={}, **kwargs):
base_class.__init__(self, config=config, **kwargs)
if (config.load_pretrained_model) and os.path.exists(config.pretrained_model_filepath):
model = TorchNNRepresentation.load(config.pretrained_model_filepath, map_location=self.config.device)
if hasattr(model, "config"):
self.config = model.config
self.config.update(self.config)
self.config.update(kwargs)
if self.network.encoder.config.use_attention:
self.network.encoder.lf.load_state_dict(model.network.encoder.lf.state_dict())
self.network.encoder.gf.load_state_dict(model.network.encoder.gf.state_dict())
self.network.encoder.ef.load_state_dict(model.network.encoder.ef.state_dict())
self.network.decoder.load_state_dict(model.network.decoder.state_dict())
else:
self.network.load_state_dict(model.network.state_dict())
def set_network(self, network_name, network_parameters):
TorchNNRepresentation.set_network(self, network_name,
network_parameters)
# add a decoder to the network for the SimCLR
self.network.projection_head = ProjectionHead(
config=network_parameters)
def set_network(self, network_name, network_parameters):
TorchNNRepresentation.set_network(self, network_name,
network_parameters)
def set_network(self, network_name, network_parameters):
TorchNNRepresentation.set_network(self, network_name, network_parameters)
# add attention head
if self.network.encoder.config.use_attention:
self.network.fc_cast = nn.Linear(self.config.network.parameters.n_latents * 4,
self.config.network.parameters.n_latents)
def __init__(self, config={}, **kwargs):
TorchNNRepresentation.__init__(self, config=config, **kwargs)
def set_network(self, network_name, network_parameters):
TorchNNRepresentation.set_network(self, network_name,
network_parameters)
# add a decoder to the network for the VAE
decoder_class = decoders.get_decoder(network_name)
self.network.decoder = decoder_class(config=network_parameters)