language-model

https://github.com/huggingface/transformers/blob/546dc24e0883e5e9f5eb06ec8060e3e6ccc5f6d7/src/transformers/models/gpt2/modeling_gpt2.py#L698

Assertions can't be relied upon for control flow because they can be disabled, as per the following:

$ python --help
usage: python [option] ... [-c cmd | -m mod | file | -] [arg] ...
...
-O     : remove assert and __debug__-dependent statem

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.

Many users in our community have been asking to have easier ways to return the output of intermediate nodes. I can see that this could be very useful for debugging and also qualitative evaluation.

I think this feature would be very useful, though the exact design is not yet fully clear.

Issue to track tutorial requests:

• Deep Learning with PyTorch: A 60 Minute Blitz - #69
• Sentence Classification - #79