How to use the cortex.add_step function in cortex

def build_normal_GAN():
    cortex.add_step('discriminator._cost', P='fake.P', X=0.,
                    name='fake_cost')
    cortex.add_step('discriminator._cost', P='real.P', X=1.,
                    name='real_cost')

    cortex.build()
    
    cortex.add_cost(
        lambda x, y: x + y, 'fake_cost.output', 'real_cost.output',
        name='discriminator_cost')
    
    m = 2
    if m == 0:
        cortex.add_cost(gen_cost, F='fake.P',
                        cells=['discriminator'],
                        name='generator_cost')
    elif m == 1:

# Generative model
    cortex.prepare_cell('gaussian', name='noise', dim=dim_in)
    cortex.prepare_cell('DistributionMLP', name='generator', dim_hs=[500, 500, 500],
                        h_act='softplus', batch_normalization=True, dim=dim,
                        weight_normalization=False, bn_mean_only=False,
                        distribution_type=distribution_type)

    # Discriminator
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        distribution_type='binomial',
                        dim=1, dropout=d_dropout,
                        dim_in=dim, dim_hs=[500, 200], h_act='softplus')
    
    # GRAPH --------------------------------------------------------------------

    cortex.add_step('discriminator', 'data.input', name='real')

    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples', constants=['noise.samples'])
    cortex.prepare_samples('generator.P', n_posterior_samples)
    cortex.add_step('discriminator', 'generator.samples', name='fake',
                    constants=['generator.samples'])
    
    cortex.add_step('discriminator._cost', P='fake.P', X=0., name='fake_cost')
    cortex.add_step('discriminator._cost', P='real.P', X=1., name='real_cost')
    cortex.add_step('noise.grid2d', random_idx=True, name='noise_grid')
    cortex.add_step('generator', 'noise_grid.output', name='gen_grid')
    
    cortex.build()

    #cortex.add_cost('l2_decay', 0.002, 'discriminator.mlp.weights')
    cortex.add_cost('l2_decay', 0.002, 'generator.mlp.weights')

distribution_type=distribution_type)

    # Discriminator
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        distribution_type='binomial',
                        dim=1, dropout=d_dropout,
                        dim_in=dim, dim_hs=[500, 200], h_act='softplus')
    
    # GRAPH --------------------------------------------------------------------

    cortex.add_step('discriminator', 'data.input', name='real')

    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples', constants=['noise.samples'])
    cortex.prepare_samples('generator.P', n_posterior_samples)
    cortex.add_step('discriminator', 'generator.samples', name='fake',
                    constants=['generator.samples'])
    
    cortex.add_step('discriminator._cost', P='fake.P', X=0., name='fake_cost')
    cortex.add_step('discriminator._cost', P='real.P', X=1., name='real_cost')
    cortex.add_step('noise.grid2d', random_idx=True, name='noise_grid')
    cortex.add_step('generator', 'noise_grid.output', name='gen_grid')
    
    cortex.build()

    #cortex.add_cost('l2_decay', 0.002, 'discriminator.mlp.weights')
    cortex.add_cost('l2_decay', 0.002, 'generator.mlp.weights')
    
    cortex.add_cost(lambda x, y: x + y, 'fake_cost.output', 'real_cost.output',
                name='discriminator_cost')
    cortex.add_cost(reweighted_MLE, G='generator.P',
                    G_samples='generator.samples',

cell_type='CNN2D',
        input_shape=cortex._manager.datasets['data']['image_shape'],
        filter_shapes=((4, 4), (4, 4), (4, 4)),
        strides=((2, 2), (2, 2), (1, 1)),
        pads=((1, 1), (1, 1), (1, 1)),
        n_filters=[128, 256, 1028], dim_out=1, h_act='softplus',
        batch_normalization=True)
    
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        mlp=discriminator,
                        dim=1,
                        distribution_type='binomial')
    
    cortex.prepare_cell('Baseline', name='baseline')
    
    cortex.add_step('discriminator', 'data.input', name='real')
    
    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples')
    cortex.add_step('discriminator', 'generator.output', name='fake')
    
    build_normal_GAN()
    cortex.profile()
    optimizer_args = {}
    
    train_session = cortex.create_session(batch_size=batch_size)
    cortex.build_session(test=test)
    
    trainer = cortex.setup_trainer(
        train_session,
        optimizer=optimizer,
        epochs=10000,

pads=((1, 1), (1, 1), (1, 1)),
        n_filters=[128, 256, 1028], dim_out=1, h_act='softplus',
        batch_normalization=True)
    
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        mlp=discriminator,
                        dim=1,
                        distribution_type='binomial')
    
    cortex.prepare_cell('Baseline', name='baseline')
    
    cortex.add_step('discriminator', 'data.input', name='real')
    
    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples')
    cortex.add_step('discriminator', 'generator.output', name='fake')
    
    build_normal_GAN()
    cortex.profile()
    optimizer_args = {}
    
    train_session = cortex.create_session(batch_size=batch_size)
    cortex.build_session(test=test)
    
    trainer = cortex.setup_trainer(
        train_session,
        optimizer=optimizer,
        epochs=10000,
        learning_rate=learning_rate,
        learning_rate_decay=.99,
        batch_size=batch_size,
    )

strides=((2, 2), (2, 2), (1, 1)),
        pads=((1, 1), (1, 1), (1, 1)),
        n_filters=[128, 256, 1028], dim_out=1, h_act='softplus',
        batch_normalization=True)
    
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        mlp=discriminator,
                        dim=1,
                        distribution_type='binomial')
    
    cortex.prepare_cell('Baseline', name='baseline')
    
    cortex.add_step('discriminator', 'data.input', name='real')
    
    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples')
    cortex.add_step('discriminator', 'generator.output', name='fake')
    
    build_normal_GAN()
    cortex.profile()
    optimizer_args = {}
    
    train_session = cortex.create_session(batch_size=batch_size)
    cortex.build_session(test=test)
    
    trainer = cortex.setup_trainer(
        train_session,
        optimizer=optimizer,
        epochs=10000,
        learning_rate=learning_rate,
        learning_rate_decay=.99,
        batch_size=batch_size,

h_act='softplus', batch_normalization=True, dim=dim,
                        weight_normalization=False, bn_mean_only=False,
                        distribution_type=distribution_type)

    # Discriminator
    cortex.prepare_cell('DistributionMLP', name='discriminator',
                        distribution_type='binomial',
                        dim=1, dropout=d_dropout,
                        dim_in=dim, dim_hs=[500, 200], h_act='softplus')
    
    # GRAPH --------------------------------------------------------------------

    cortex.add_step('discriminator', 'data.input', name='real')

    cortex.prepare_samples('noise', batch_size)
    cortex.add_step('generator', 'noise.samples', constants=['noise.samples'])
    cortex.prepare_samples('generator.P', n_posterior_samples)
    cortex.add_step('discriminator', 'generator.samples', name='fake',
                    constants=['generator.samples'])
    
    cortex.add_step('discriminator._cost', P='fake.P', X=0., name='fake_cost')
    cortex.add_step('discriminator._cost', P='real.P', X=1., name='real_cost')
    cortex.add_step('noise.grid2d', random_idx=True, name='noise_grid')
    cortex.add_step('generator', 'noise_grid.output', name='gen_grid')
    
    cortex.build()

    #cortex.add_cost('l2_decay', 0.002, 'discriminator.mlp.weights')
    cortex.add_cost('l2_decay', 0.002, 'generator.mlp.weights')
    
    cortex.add_cost(lambda x, y: x + y, 'fake_cost.output', 'real_cost.output',
                name='discriminator_cost')