Hands-On Generative Adversarial Networks with Keras

Training

To train the model, we use the following steps:

1. Build the model output, including a variable for the shuffled input:

 real_inputs = Input(shape=img_shape) txt_inputs = Input(shape=emb_shape) txt_shuf_inputs = Input(shape=emb_shape) z_inputs = Input(shape=(z_dim, ))

2. Instantiate symbolic variables for the discriminator's output:

 fake_samples = G([z_inputs, txt_inputs]) D_real = D([real_inputs, txt_inputs]) D_wrong = D([real_inputs, txt_shuf_inputs]) D_fake = D([fake_samples, txt_inputs])

3. Define the D graph and optimizer, including the shuffled text input:

 G.trainable = False D.trainable = True D_model = Model(inputs=[real_inputs, txt_inputs, txt_shuf_inputs, z_inputs],                 outputs=[D_real, D_wrong, D_fake])

4. Get the data for the ...