Conditional-Generation

Prerequisites

Python 3.5 +

Setup

To setup, you need to clone the repo with submodules.

git clone --recurse-submodules git@github.com:arianhosseini/Conditional-Generation.git

Create and actvate a virtual environment. Then install the dependencies.

python -m venv venv
source venv/bin/activate

pip install -r requirements.txt

Generate text data

run_generation.py script generates data from Book Corpus using the given sampling and refinement method. It outputs a txt file with the seed sentences and the generated text.

python run_generation.py \ 
    --model_type xlnet 
    --model_name_or_path xlnet-base-cased
    --mode ltr
    --refine gibbs

Evaluate the generated text

First download the embedding data and the GloVe pretrained models

# Embedding data
mkdir GloVe
curl -Lo GloVe/glove.840B.300d.zip http://nlp.stanford.edu/data/glove.840B.300d.zip
unzip GloVe/glove.840B.300d.zip -d GloVe/
# Pretrained model
mkdir encoder
curl -Lo encoder/infersent1.pkl https://dl.fbaipublicfiles.com/infersent/infersent1.pkl

Then run the evalution script. This script takes in a text file with seed sentences and generated text (output of run_generation.py). It computes the Frechet Distance on InferSent embeddings of real and generated text. The script prints the FD score.

python run_evaluation.py  --input-file my_input_file.txt --verbose

Batch processing

If you want to run the generation and evalutation in batch, you can use the bash scripts provided.

bash run_generation.sh
bash run_evaluation.sh

arianhosseini / Conditional-Generation