Tensorflow apply_gradients() 有多重损失

Question

我正在训练一个具有中间输出的模型 (VAEGAN)，我有两次损失，

我从输出层计算的 KL 散度损失
我从中间层计算的相似性 (rec) 损失。

我可以简单地总结它们并应用如下所示的渐变吗？

with tf.GradientTape() as tape:
    z_mean, z_log_sigma, z_encoder_output = self.encoder(real_images, training = True)
    kl_loss = self.kl_loss_fn(z_mean, z_log_sigma) * kl_loss_coeff

    fake_images = self.decoder(z_encoder_output)
    fake_inter_activations, logits_fake = self.discriminator(fake_images, training = True)
    real_inter_activations, logits_real = self.discriminator(real_images, training = True)

    rec_loss = self.rec_loss_fn(fake_inter_activations, real_inter_activations) * rec_loss_coeff

    total_encoder_loss = kl_loss + rec_loss

grads = tape.gradient(total_encoder_loss, self.encoder.trainable_weights)
self.e_optimizer.apply_gradients(zip(grads, self.encoder.trainable_weights))

或者我是否需要像下面那样将它们分开，同时保持磁带的持久性？

with tf.GradientTape(persistent = True) as tape:
    z_mean, z_log_sigma, z_encoder_output = self.encoder(real_images, training = True)
    kl_loss = self.kl_loss_fn(z_mean, z_log_sigma) * kl_loss_coeff
    
    fake_images = self.decoder(z_encoder_output)
    fake_inter_activations, logits_fake = self.discriminator(fake_images, training = True)
    real_inter_activations, logits_real = self.discriminator(real_images, training = True)
    
    rec_loss = self.rec_loss_fn(fake_inter_activations, real_inter_activations) * rec_loss_coeff

grads_kl_loss = tape.gradient(kl_loss, self.encoder.trainable_weights)
self.e_optimizer.apply_gradients(zip(grads_kl_loss, self.encoder.trainable_weights))

grads_rec_loss = tape.gradient(rec_loss, self.encoder.trainable_weights)
self.e_optimizer.apply_gradients(zip(grads_rec_loss, self.encoder.trainable_weights))

Answer 1

是的，您通常可以对损失求和并计算单个梯度。由于a sum的梯度是各自梯度的和，所以summed loss走的步和一步一步走两个步是一样的。

这是一个简单的例子：假设您有两个权重，并且您当前位于 (1, 3) 点（“起点”）。 loss 1的梯度为(2, -4)，loss 2的梯度为(1, 2).

如果一个接一个地应用这些步骤，您将首先移动到 (3, -1)，然后移动到 (4, 1)。
如果先对梯度求和，总步长为(3, -2)。从起点沿着这个方向也可以到达 (4, 1)。

Tensorflow apply_gradients() 有多重损失

Tensorflow apply_gradients() with multiple losses

deep-learning

tensorflow

loss-function

generative-adversarial-network