Hi, What if the prior probability of the noise label of the training set is not known?

[ZunxiaoXu]

No description provided.

[MuhammadSYahyaS]

I also have similar issue. In the paper itself, it is being said that

The prior probability P(Y~|D~) is counted directly from the noisy dataset.

But, it is also being said that

Beyond their results, we attempt to propose a theoretically sound approach addressing a general instance-based noise regime without knowing or estimating noise rates.

In the implementation, if we set noise_prior = None, then

cores/phase1.py

Line 211 in cf0945e

noise_prior_cur = noise_prior*num_training_samples - noise_prior_delta

will raise an error. I have tried:

• Initializing it with noise_prior = np.ones(num_classes)
• Avoiding the two lines of noise_prior_cur update process

and all on CIFAR-10 dataset, but my best result was the best test acc simply dropped from 69.28 (using the dataset's known noise prior) to 46.54. Any lead about how to set the noise_prior on unknown prior probability of the dataset's noise? @haochenglouis