Learning Transformer Programs

This repository contains the code for our paper, Learning Transformer Programs. The code can be used to train a modified Transformer to solve a task, and then convert it into a human-readable Python program. The repository also includes a number of example programs, which we learned for the tasks described in the paper. Please see our paper for more details.

Quick links

• Setup
• Learning Programs
  • Training
  • Converting to code
• Example Programs
• Questions?
• Citation

Setup

Install PyTorch and then install the remaining requirements: pip install -r requirements.txt. This code was tested using Python 3.8 and PyTorch version 1.13.1.

In our experiments on NLP tasks, we initialize word embeddings using 300-dimensional pre-trained GloVe embeddings, which can be downloaded here (Common Crawl, cased):

mkdir data
wget https://huggingface.co/stanfordnlp/glove/resolve/main/glove.840B.300d.zip -P data/
unzip data/glove.840B.300d.zip

Learning Programs

Training

The code to learn a Transformer Program can be found in src/run.py. For example, the following command will train a Transformer Program for the sort task, using two layers, four categorical attention heads per-layer, and one-hot input embeddings:

python src/run.py \
     --dataset "sort" \
     --vocab_size 8 \
     --dataset_size 10000 \
     --min_length 1 \
     --max_length 8 \
     --n_epochs 250 \
     --batch_size 512 \
     --lr "5e-2" \
     --n_layers 2 \
     --n_heads_cat 4 \
     --n_heads_num 0 \
     --n_cat_mlps 1 \
     --n_num_mlps 0 \
     --one_hot_embed \
     --count_only \
     --seed 0 \
     --save \
     --save_code \
     --output_dir "output/sort";

This command will train a Transformer Program for the CoNLL 2003 named-entity recognition task, learning input embeddings composed of four 32-dimensional categorical variables:

python src/run.py \
     --dataset "conll_ner" \
     --vocab_size 10000 \
     --min_length 1 \
     --max_length 32 \
     --n_epochs 50 \
     --batch_size 32 \
     --lr "5e-2" \
     --n_vars_cat 4 \
     --d_var 32 \
     --n_layers 2 \
     --n_heads_cat 4 \
     --n_heads_num 0 \
     --n_cat_mlps 1 \
     --n_num_mlps 0 \
     --mlp_vars_in 2 \
     --count_only \
     --seed 0 \
     --replace_numbers 1 \
     --glove_embeddings "data/glove.840B.300d.txt" \
     --do_glove 1 \
     --save \
     --save_code \
     --output_dir "output/conll";

Please see src/run.py for all of the possible arguments. The training data will either be generated (for the RASP tasks) or downloaded from Hugging Face Datasets; see src/utils/data_utils.py for the supported datasets. The scripts directory contains scripts for training Transformer Programs and standard Transformers with the experiment settings used in the paper.

Converting to code

Run the training script with the --save_code flag to convert the model to a Python program at the end of training. To convert a model that has already been trained, use src/decompile.py. For example,

python src/decompile.py --path output/sort/ --output_dir programs/sort/

output/sort/ should be the output directory of a training run.

Example Programs

The programs directory contains example programs for small-scale versions of all of the RASP tasks, as well as named-entity recognition. Each program defines a function called run that takes a sequence of tokens as input and returns a list of predicted labels. For example:

>>> from programs.rasp.sort import sort
>>> sort.run(["<s>", "3", "1", "4", "2", "4", "0", "</s>"])
['<s>', '0', '1', '2', '3', '4', '4', '</s>']

programs/rasp contains the best-performing programs for each task, using both categorical and numerical attention heads. programs/rasp_categorical_only contains the best-performing programs using only categorical variables. programs/conll_ner contains a program for named-entity recognition.

Questions?

If you have any questions about the code or paper, please email Dan ([email protected]) or open an issue.

Citation

@article{friedman2023learning,
    title={Learning {T}ransformer {P}rograms},
    author={Friedman, Dan and Wettig, Alexander and Chen, Danqi},
    journal={arXiv preprint},
    year={2023}
}