Sadra Barikbin

Hi, yes it does: ```python from tokenizer import Tokenizer from tokenizers.normalizers import Replace from tokenizers import Regex from tokenizers.models import BPE my_tokenizer = Tokenizer(BPE()) my_tokenizer.normalizer = Replace(Regex('[0-9]+'), '[NUM]') ```

[NUM] is not identified in the first one. Its id is that of the [UNK] token.

Hi @n1t0 @patrickvonplaten I just came across the third issue @n1t0 mentioned: > We can't extract these added tokens during the pre-processing step of the training, which would be desirable...

Yes, you're right. I had written it for a project, so I attempted for a PR.

@vfdev-5 This class is also useful in the case where user wants to apply such scheduling on a parameter other than LR or a specific parameter group.

I can't name one, but what about the second use case? if someone wanted to apply this scheduler on a specific parameter(LR) group. Currently our `LRScheduler` does not accept a...

It could work, but only by separating their optimizers. That gets difficult. PyTorch's `_LRSchduler` does not accept a `param_group_index`-like parameter.

I searched a little bit and found that for computing AUC there are three approaches: 1. Computing it exactly. It takes O(NlogN) in time and O(N) in space. It is...

You're right that trying to compute the metric at the end of run is excruciating both in terms of time and in terms of memory, given large data samples. Wilconon-Mann-Whitney...

 According to the formulas above, using two float variables initialized with zero at the beginning of the run we do two first formulas at each batch and finally do...