Discriminator.trainable = False makes discriminator parameters also non trainable

#1

When we are making the GAN model we set discriminator.trainable = False. Now when I do discriminator.summary() I get 0 trainable parameters. I cannot understand the reason for this. I’d already compiled my discriminator model before making the changes. My GAN algorithm somehow works absolutely fin. I’m able to generate accurate imagess

#2

Hey @raunaqsingh10, yes theoritically the code in the video, is little wrong, I also wrote my code, and i made sure to correct this mistake,

for step in tqdm(range(NO_OF_BATCHES)):
        #randomly select 50% real images
        idx = np.random.randint(0,X_Train.shape[0],HALF_BATCH_SIZE)
        real_imgs = X_Train[idx]
        
        # generate 50% random images
        noise = np.random.normal(0,1,size=(HALF_BATCH_SIZE,NOISE_DIM))
        fake_imgs = generator.predict(noise)
        
        
        # one sided label smoothing
        real_y = np.ones((HALF_BATCH_SIZE,1))*0.9 #Label Smoothing, Works well in practice
        fake_y = np.zeros((HALF_BATCH_SIZE,1))
        
        
        # train on real and fake images
        discriminator.trainable = True
        generator.trainable = False
        d_loss_real = discriminator.train_on_batch(real_imgs,real_y) #updates the weights of discriminator
        d_loss_fake = discriminator.train_on_batch(fake_imgs,fake_y) 
        d_loss = 0.5*d_loss_real + 0.5*d_loss_fake
        
        epoch_d_loss += d_loss
        # discriminator.trainable = False
        
        
        #Train Generator (Complete Model Generator + Frozen Discriminator)
        discriminator.trainable = False
        generator.trainable = True
        noise = np.random.normal(0,1,size=(BATCH_SIZE,NOISE_DIM))
        real_y = np.ones((BATCH_SIZE,1))
        g_loss = model.train_on_batch(noise,real_y)
        epoch_g_loss += g_loss

My model was also able to produce satisfactory results. Though it need a little more time,

Hope this resolved your doubt.
Plz mark the doubt as resolved in my doubts section. :blush: