使用tf.GradientTape训练GAN时生成器梯度返回[None]如何解决

问题描述 #

训练GAN时判别器梯度计算正常，生成器梯度返回None，优化器应用梯度时报错：ValueError: No gradients provided for any variable: ['carrier_freq:0'].，当前生成器仅包含单个可训练权重，损失可正常输出数值。

报错关联代码 #

generator = make_generator()
discriminator = make_discriminator()

g_loss_fn = keras.losses.BinaryCrossentropy(from_logits=True)
g_optimizer = keras.optimizers.Adam(learning_rate=0.04)
d_loss_fn = keras.losses.BinaryCrossentropy(from_logits=True)
d_optimizer = keras.optimizers.Adam(learning_rate=0.03)

def train_step(train_set):
    # modulate or don't modulate sample
    for batch in train_set:
        # get a random DEMAND noise sample to mix with speech
        noise_indices = tf.random.uniform([batch_size], minval=0, maxval=len(demand_dataset), dtype=tf.int32)
        
        # labels of 0 representing legit samples
        legit_labels = tf.zeros(batch_size, dtype=tf.uint8)
        # labels of 1 representing adversarial samples
        adversarial_labels = tf.ones(batch_size, dtype=tf.uint8)
        # concat legit and adversarial labels
        concat_labels = tf.concat((legit_labels, adversarial_labels), axis=0)
        
        # calculate gradients
        with tf.GradientTape(persistent=True) as tape:
            legit_predictions = discriminator(legit_path(batch, noise_indices))
            adversarial_predictions = discriminator(adversarial_path(batch, noise_indices))
            # concat legit and adversarial predictions to match double batch of concat_labels
            d_predictions = tf.concat((legit_predictions, adversarial_predictions), axis=0)
            d_loss = d_loss_fn(concat_labels, d_predictions)
            g_loss = g_loss_fn(legit_labels, adversarial_predictions)
            print('Discriminator loss: ' + str(d_loss))
            print('Generator loss: ' + str(g_loss))
        d_grads = tape.gradient(d_loss, discriminator.trainable_weights)
        g_grads = tape.gradient(g_loss, generator.trainable_weights)
        print(g_grads)
        d_optimizer.apply_gradients(zip(d_grads, discriminator.trainable_weights))
        g_optimizer.apply_gradients(zip(g_grads, generator.trainable_weights))
        
        discriminator_loss(d_loss)
        generator_loss(g_loss)
    return d_loss, g_loss

关联逻辑说明 #

判别器目标是区分合法样本与对抗样本，输入包含两倍batch大小数据：一路是预处理后的合法数据，另一路是生成器修改得到的对抗样本。生成器仅含单个权重，由封装乘加操作的lambda层构成，判别器为全连接网络。
损失均采用BinaryCrossentropy计算：判别器损失基于样本是否被修改的真实标签计算，生成器损失统计对抗样本被判别为合法样本的概率。

问题根因排查方向 #

• 生成对抗样本的adversarial_path函数未在tf.GradientTape上下文范围内调用生成器，或者根本没有调用生成器，导致梯度带没有记录生成器权重的计算链路，无法反向传播梯度。
• 生成器的lambda层权重未被正确注册为可训练参数，可打印generator.trainable_weights确认carrier_freq是否在返回列表中，排除权重未被识别为可训练参数的问题。
• 生成器内部或adversarial_path流程中使用了不可微操作，例如将张量转为numpy数组、使用非TensorFlow内置算子、添加tf.stop_gradient截断梯度、强制类型转换为无梯度的整数类型、对生成的对抗样本做量化/编码这类不可微预处理，都会打断梯度传播链路。
• 生成器的乘加操作没有关联到可训练的carrier_freq变量，例如lambda层内直接用了常量做计算，没有引用生成器的可训练权重。

内容的提问来源于stack exchange，提问作者jost