The original training process as seen in Chapter 3, My First GAN in Under 100 Lines, is as follows:

### Pseudocode
On Epoch:
    Batch Size = ##
    x_train_real = Half of Batch From Real Images
    y_train_real = 1s (Real) x Batch Size
    x_train_gen  = Half of Batch from Generated Images
    y_train_gen  = 0s (Fake) x Batch Size

    train_discriminator(concat(x_train_real,x_train_gen),
                        concat(y_train_real,y_train_gen))

    x_generated_images = generator(batch_size)
    y_labels = 1s for Real
    train_gan(x_generated_images,y_labels)

The updated process is as follows:

### Pseudocode
On Epoch:
    Batch Size = ##
    flipcoin()
        x_train = Half of Batch From Real Images
        y_train = 1s (Real) x Batch Size
    else:
        x_train  = Half of Batch from Generated Images
        y_train  = 0s (Fake) x Batch Size

    train_discriminator(x_train,y_train)

    x_generated_images = generator(batch_size)
    flipcoin()
        y_labels = 1s for Real
    else:
        y_labels = 0s for Fake (Noisy Labels)
    train_gan(x_generated_images,y_labels)

A key change with this training step is that we no longer mix the generated and real images when training the discriminator, as illustrated in the following diagram:

These minor changes to the trainer class work toward improving the performance of GAN, and in some cases, are the reason you can converge. For example, when writing this chapter, we ran into a few instances where the discriminator would diverge, making the generator only produce noise.