Recently HF trainer was extended to support full fp16 eval via --fp16_full_eval. I'd have expected it to be either equal or faster than eval with fp32 model, but surprisingly I have noticed a 25% slowdown when using it.

This may or may not impact deepspeed as well, which also runs eval in fp16, but we can't compare it to a baseline, since it only runs fp16.

I wonder if someone would like t

The Split class accepts SplitDelimiterBehavior which is really useful. The Punctuation however always uses SplitDelimiterBehavior::Isolated (and Whitespace on the other hand behaves like SplitDelimiterBehavior::Removed).

impl PreTokenizer for Punctuation {
    fn pre_tokenize(&self, pretokenized: &mut PreTokenizedString) -> Result<()> {
        pretokenized.split(|_, s| s.spl

From paper, it mentioned

Instead, the training data generator chooses 15% of tokens at random, e.g., in the sentence my
dog is hairy it chooses hairy.

It means that 15% of token will be choose for sure.

From https://github.com/codertimo/BERT-pytorch/blob/master/bert_pytorch/dataset/dataset.py#L68,
for every single token, it has 15% of chance that go though the followup procedure.

Create a suite of tools to easily manipulate SQuAD format data. It would be useful to have tools to do things such as:

• merging annotations
• converting SQuAD format to Pandas data frame and vice versa
• easier functions to remove samples / paragraphs / annotations
• easier functions to count
• easier functions to return all questions or all answers

Issue to track tutorial requests:

• Deep Learning with PyTorch: A 60 Minute Blitz - #69