Predictive Querying for Autoregressive Neural Sequence Models

This project is for the paper: Predictive Querying for Autoregressive Neural Sequence Models, published in NeurIPs 2022.

An experimental analysis of the probability that a sequence will produce an end-of-sentence token at a given timstep $K$ in the future, conditioned on a partial sequence history. Open ended histories possess greater probability mass on larger values of $K$ than histories that typically represent short phrases.

Setup

Datasets

Our experiments make use of 5 datasets. Links for downloading these files can be found below. After downloading, place the files in the /data folder.

• MOOCs
• Mobile Apps
• Amazon Reviews
• Shakespeare
• Wikitext-v2

Models

We include the checkpoints of the trained models we utilized in our experiments. Those can be found in the models/ directory under the corresponding dataset folder as model_*.pt.

Required Packages

Our code requires the following packages:

• PyTorch
• Scipy
• Numpy
• Sklearn
• Pandas
• HuggingFace Transformers
• argparse (CLI)
• yaml
• json

Other Details

Our experiments are conducted on Ubuntu Linux 20.04 with Python 3.8.

Code Structure

To provide a general guide to the codebase, we have outlined the main folders/files and what they contain below.

• seq_queries/ - General directory where all the core-functionality code is found.
  • samples.py - Contains all query estimation methods, including sampling, search, and hybrid methods.
  • tree.py - Contains the functionality for the hybrid's query sequence path tree.
  • experiments.py - Orchestration functions. The primary function of interest is sample_dynamic_target_token which conducts hitting time query experiments.
  • train.py and model.py and optim.py - Model training and management.
  • data.py - Data preprocessing and loading.
  • arguments.py - Argparse command line interface.
• scripts/ - Collection of scripts that leverage the core functionality (above) for specific tasks. The names are generally self-explanatory. Each of these is prepared to be run so long as the model and datasets can be found in the directory structure. Be sure to adjust paths accordingly. Below are a subset of the most important scripts that we provide.
  • get_importance_sampling.py
  • get_uniform_mc_sampling.py
  • get_naive_sampling.py
  • get_beam_search_lower_bound.py
  • get_ground_truth.py
  • beam_search_ablation.py
  • entropy_ablation.py
  • temperature_ablation.py
  • train/* - All training scripts
• data/ and models/ - Where data and models should be placed, with pre-trained model checkpoints already present. The code assumes each dataset will be present inside of a folder with its corresponding name. The names we utilized are apps, amazon, moocs, shakespeare, and wikitext.

Experimentation

All of our experimentation, including model training, query estimation, and all ablations, is completed through a single argparse CLI. There are two ways to conduct experiments in this setting. The first is to write a small script with get_args() included from the arguments.py file and then execute it, sending non-default arguments to the execution through the CLI directly. However, since there are many argparse arguments to leverage and this code is quite flexible, we also provide the ability to populate argparse arguments with a .yaml file. Providing a .yaml file will overwrite all other arguments, and an example can be found in config/sample.yaml.

The paragraph above assumes that you wish to run a custom set of experiments with our software. If this is not the case, please refer to the many pre-made scripts in the scripts/ folder for guidance and edit those as necessary. These can be run simply with python scripts/{sample_script}.py and is our recommended means of running this code.

For Bibtex Citations

If you find out work helpful in your own research, please cite our paper with the following:

@inproceedings{boyd2022query,
  author={Boyd, Alex  and Showalter, Sam  and Mandt, Stephan and Smyth, Padhraic},
  title={Predictive Querying for Autoregressive Neural Sequence Models},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022},
}

Reaching Out

Please create an issue for code-related questions. For more in-depth discussion, free feel to email us at [email protected]