pytorch reimplementation of quickthoughts paper: https://arxiv.org/pdf/1803.02893.pdf
This is a very rough implementation atm, so feel free to reach out if you're having trouble running the code.
There's also a lot of code for some experiments I ran here in this repo.
First you'll need to pull down the bookcorpus submodule and follow the instructions there to get all.txt
.
Next move all.txt
to the root folder and then run make data
, which will process the data into cleaned.txt
.
Next grab the pretrained word vectors by running make pull-vec
.
You'll be able to train with make train
.
The model will pull from a config file, which can be edited to tune parameters.