Open zhanjw opened 2 years ago
It is a bit difficult to understand the behavior of setting the target of normal samples in the training of holistic_head to 1 in generate_target,
def generate_target(self, target, eval=False): targets = list() if eval: targets.append(target==0) targets.append(target) targets.append(target) targets.append(target) return targets else: temp_t = target != 0 targets.append(target == 0) targets.append(temp_t[target != 2]) targets.append(temp_t[target != 1]) targets.append(target != 0) return targets
Can you assist me in understanding it? There is no mention of this in the paper.
I have the same problem about it. Do you understand it now? If it is, could you explain the problem? Thank you!
I think the normality head predict the probability of normality, that's the reason why they minus this score to obatin final anomaly score.
Hi, @zhanjw @CC2033625919
As @caiyu6666 said, we want holistic_head to predict the probability of normality, so we changed the target to learn the feature of normality.
Cheers, Choubo
It is a bit difficult to understand the behavior of setting the target of normal samples in the training of holistic_head to 1 in generate_target,
Can you assist me in understanding it? There is no mention of this in the paper.