DCGAN의 train_step 메소드 구현과정에서, 다음과 같은 코드에 의문이 생겨 질문드립니다.
```
real_labels = tf.ones_like(real_predictions)
real_noisy_labels = real_labels + NOISE_PARAM * tf.random.uniform(
tf.shape(real_predictions)
)
fake_labels = tf.zeros_like(fake_predictions)
fake_noisy_labels = fake_labels - NOISE_PARAM * tf.random.uniform(
tf.shape(fake_predictions)
)
```
real_labels 와 fake_labels를 각각 `tf.ones_like`와 `tf.zeros_like`로 만들어준 후에 label smoothing을 진행하는 것으로 이해했는데요. `tf.random.uniform`이 0에서 1.0사이의 무작위 수를 생성하므로 real_noisy_labels에 -연산이, fake_noisy_labels에 + 연산이 들어가야 맞는 것 아닌가요?