GitLab

cc @lysandrejik

Custom schedulers are currently initiated by wrapping Pytorch's LambdaLR
class and passing a method of the wrapping class to the __init__
function of LambdaLR. This approach is not appropriate for several
reasons:

1. one does not need to define a class when it only defines a
__init__() method;
2. instantiating the parent class by passing a method of the child class
creates a cyclical reference which leads to memory leaks. See issues #1742 and #1134.

In this commit we replace the wrapper classes with functions that
instantiate `LambdaLR` with a custom learning rate function. We use a
closure to specify the parameter of the latter. We also do a bit of
renaming within the function to explicit the behaviour and removed
docstrings that were subsequently not necessary.

As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.

For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.

This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
    1) lowercasing tokens added with .add_tokens()
    2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers

https://github.com/huggingface/transformers/issues/1545

cf. https://docs.pytest.org/en/latest/example/simple.html

[WIP] copy paste bert token classification for roberta