An implementation of Skip-Thought Vectors in PyTorch. Here's a blog explaining the subtleties about Skip-Thoughts
- Download BookCorpus or any other data-set and concatenate all sentences into one file and put it in
./data/
directory - Modify the following line in
Train.ipynb
notebook accordingly:d = DataLoader("./data/dummy_corpus.txt")
- There is no early stopping.
- The
Train
notebook runs at the rate of 1 epoch / 2 days on an Nvidia 1080 Ti. - Your model is saved when
./saved_models
when the average training loss in the last 20 iterations dips below the previous best.
Only implemented on classification tasks
- Download the movie review dataset and put
rt-polarity.neg
andrt-polarity.pos
in the./tasks/mr_data
directory. - You may also test on other classification tasks by downloading the datasets and providing their path and tasks type in
Evaluate.ipynb