def __init__(self,
units_full=128,
units_latent=20,
u_encoder_block=DenseEncoderDecoder(),
u_decoder_block=DenseEncoderDecoder(),
F_encoder_block=DenseEncoderDecoder(),
F_decoder_block=DenseEncoderDecoder(),
operator_initializer=keras.initializers.Identity(),
train_autoencoders_only=False,
**kwargs):
super().__init__(**kwargs) # handles standard args (e.g., name)
# Place configuration as attributes
self.l = units_latent
# u autoencoder
self.u_encoder = u_encoder_block
self.u_Reduce = tf.Variable(self.I_seed(units_full, units_latent),
trainable=True)
self.u_Expand = tf.Variable(self.I_seed(units_latent, units_full),
trainable=True)
self.u_decoder = u_decoder_block
# F autoencoder
self.F_encoder = F_encoder_block
self.F_Reduce = tf.Variable(self.I_seed(units_full, units_latent),
trainable=True)
self.F_Expand = tf.Variable(self.I_seed(units_latent, units_full),
trainable=True)
self.F_decoder = F_decoder_block
# Now create the operator layer that connects v->f
self.operator_initializer = operator_initializer
op_size = (units_latent, units_latent)
self.Operator = tf.Variable(abs(operator_initializer(op_size)),
trainable=True)
# Set the NMSE loss function used for custom losses
self.NMSE = NMSE(name="NormalizedMSE")
# Boolean for whether or not to train Autoencoders only
self.train_autoencoders_only = train_autoencoders_only
'units_full': units_full,
'units_latent': units_latent,
'u_encoder_block': DenseEncoderDecoder(**enc_dec_config),
'u_decoder_block': DenseEncoderDecoder(**enc_dec_config),
'F_encoder_block': DenseEncoderDecoder(**enc_dec_config),
'F_decoder_block': DenseEncoderDecoder(**enc_dec_config),
'operator_initializer': return_sdg_matrix
}
# Aggregate all the training options in one dictionary
training_options = {
'aec_only_epochs': 25,
'init_full_epochs': 300,
'best_model_epochs': 2500,
'num_init_models': 20,
'loss_fn': NMSE(),
'optimizer': keras.optimizers.Adam,
'optimizer_opts': {},
'batch_size': 64
}
####################################################################
### Launch the Experiment
####################################################################
# Get a random number generator seed
random_seed = r.randint(0, 10**(10))
# Set the custom objects used in the model (for loading purposes)
custom_objs = {"NormalizedMeanSquaredError": NMSE}