tf.keras中关于model.trainable=False的设置（in GAN）

提出问题

在看GAN的实现代码的时候，发现了这么一个地方：

class GAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.latent_dim = 100
​
        optimizer = Adam(0.0002, 0.5)
​
        # Build and compile the discriminator
        self.discriminator = self.build_discriminator()
        self.discriminator.compile(loss='binary_crossentropy',
            optimizer=optimizer,
            metrics=['accuracy'])
​
        # Build the generator
        self.generator = self.build_generator()
​
        # The generator takes noise as input and generates imgs
        z = Input(shape=(self.latent_dim,))
        img = self.generator(z)
​
        # For the combined model we will only train the generator
        self.discriminator.trainable = False
​
        # The discriminator takes generated images as input and determines validity
        validity = self.discriminator(img)
​
        # The combined model  (stacked generator and discriminator)
        # Trains the generator to fool the discriminator
        self.combined = Model(z, validity)
        self.combined.compile(loss='binary_crossentropy', optimizer=optimizer)

缩小上面代码的范围，看这一行：

 # For the combined model we will only train the generator
        self.discriminator