CMPUT 651 Project at University of Alberta in Fall 2019

1. Ensure torch and sklearn are installed.
2. Set up the project.

$ pip install -e .
$ pip install torchtext
$ pip install spacy
$ python3 -m spacy download en

3. Download pre-trained GloVe word vectors. Place the Common Crawl (840B tokens, 2.2M vocab, cased, 300d vectors, 2.03 GB download) word vectors in /data/glove/ directory.
4. Download the trial dataset and training dataset, unzip the .zip files and move the image directories in /data.
5. Download the InferSent model trained with GloVe in /data.

$ mkdir encoder
$ curl -Lo encoder/infersent1.pkl https://dl.fbaipublicfiles.com/infersent/infersent1.pkl

6. Clone the Facebook AI Research Sequence-to-Sequence Toolkit (for PyTorch implementation of RoBERTa) in the project directory.

$ git clone git@github.com:pytorch/fairseq.git

7. Download the pretrained RoBERTa model roberta.large here. Decompress the file, and place the folder in the project directory.