This repo contains the code used to compose the experimental part of the paper.
To run this code with its dependenecies installed, we recommend to use the docker container xehartnort/tfgpu.
Note that for the wikitext-103 experiments there are a few requirements, not provided in this repository due to their size:
- A path to "wikitext/wikitext.train" and "wikitext/wikitext.no_train", which contains the tokenized sentences of 1024 elements from wikitext-103 used to train GPT2.
- A path to "wikitext/new_gpt2_model", which contains the checkpoint of the GPT-2 model trained for 20 epochs on "wikitext/wikitext.train".
Such requirements can be fulfilled by downloading and extracting the contents of the following file wikitext.zip. The zip contains: "wikitext/new_gpt2_model", "wikitext/wikitext.train" and "wikitext/wikitext.no_train"