why the loss in wgan.py is different with the original paper?

[guojting]

It makes me confused that which one is correct? As implemented in wgan.py, we have self.g_loss = tf.reduce_mean(self.d_) self.d_loss = tf.reduce_mean(self.d) - tf.reduce_mean(self.d_) however, according to the original paper of wgan, it seems that we should minimize (-1)*self.g_loss, instead of self.g_loss. Could you tell me why the losses are implemented in the above form? Anyway, it seems that using the implementation in wgan.py or wgan_v2.py, I can still get some results. This makes me more confused.

How about the losses as follows self.g_loss = tf.reduce_mean(tf.scalar_mul(-1,self.d_)) self.d_loss = tf.reduce_mean(self.d_) - tf.reduce_mean(self.d) ?

Thank you!

[guangyuanyu]

The same question

[Joey-Liu]

I also have this question， Is this a mistake?