# Initialize parameters
log.info("Initializing parameters")
model.init_params()
# Create theano shared variables
log.info('Creating shared variables')
model.init_shared_variables()
# List of weights that will not receive updates during BP
dont_update = []
# Override some weights with pre-trained ones if given
if train_args.init:
log.info('Will override parameters from pre-trained weights')
log.info(' %s' % os.path.basename(train_args.init))
new_params = get_param_dict(train_args.init)
model.update_shared_variables(new_params)
if freeze:
log.info('Pretrained weights will not be updated.')
dont_update = list(new_params.keys())
# Print number of parameters
log.info("Number of parameters: %s" % model.get_nb_params())
# Load data
log.info("Loading data")
model.load_data()
# Dump model information
model.info()
model.init_shared_variables()
# Khoa:
discriminator.init_shared_variables()
if train_args.model_language_model_type is not None:
language_model.init_shared_variables()
# Khoa.
# List of weights that will not receive updates during BP
dont_update = []
# Override some weights with pre-trained ones if given
if train_args.init:
log.info(
'Will override parameters from pre-trained weights init Generator')
log.info(' %s' % os.path.basename(train_args.init))
new_params = get_param_dict(train_args.init)
model.update_shared_variables(new_params)
if freeze:
log.info('Pretrained weights will not be updated.')
dont_update = list(new_params.keys())
if train_args.initdis:
log.info(
'Will override parameters from pre-trained weights init Discriminator'
)
log.info(' %s' % os.path.basename(train_args.initdis))
new_params = get_param_dict(train_args.initdis)
discriminator.update_shared_variables(new_params)
if freeze:
log.info('Pretrained weights will not be updated.')
dont_update = list(new_params.keys())