loss_alpha = 1 # amplitude
loss_beta = 1 # phase
# threshold is useless when normalization is None
loss_type = {'threshold': 'origin', 'normalization': None}
loss_threshold = 1
# training
mark = 'rcf-gan'
if adversarial_training_param['adv_t_sigma_num'] > 0:
model_label = target_source + '_t_net_' + generator_training_param['net_type'] + '_' + mark
else:
model_label = target_source + '_t_normal_' + generator_training_param['net_type'] + '_' + mark
model = Model(model_label,
target_source, target_dim, target_size, target_batch_size,
adversarial_training_param, generator_training_param,
loss_type, loss_alpha, loss_beta, loss_threshold, ae_loss_reg,
epoch)
if load >= 0:
model.train(load)
# parameters for testing
num_white_noise_test = 100
# testing
model.test(num_white_noise_test)
model.test_rec()
model.save_for_scores()
# model.save_for_scores_per_epoc()
model.interpolated_imgs(8)
'activations_a': [('lrelu', 0.2), ('tanh', None)],
'weight_decay': 0,
'lr_step_size_decay': 0,
'lr_decay_gamma': 0.5,
# activations for the t_net, only valid when t_net = True
'activations_t': [('lrelu', 0.2), ('tanh', None)]
}
if adversarial_training_param['adv_t_sigma_num'] > 0:
model_label = target_source + '_t_net_' + generator_training_param[
'net_type'] + '_' + args.mark
else:
model_label = target_source + '_t_normal_' + generator_training_param[
'net_type'] + '_' + args.mark
# training
model = Model(model_label, target_source, target_dim, target_size,
target_batch_size, adversarial_training_param,
generator_training_param, loss_type, loss_alpha, loss_beta,
loss_threshold, ae_loss_reg, args.epochs)
if args.resume_training >= 0:
model.train(args.resume_training, args.save_period)
# testing
model.test()
model.test_rec()
model.save_for_scores()
# model.save_for_scores_per_epoc()
model.interpolated_imgs(8)