This repository provides a library for efficient training of masked language models (MLM), built with fairseq. We fork fairseq to give researchers more flexibility when using our training scripts, while also making it easier to adapt our code contributions into other projects.

Why DinkyTrain?

The Dinky runs between Princeton Junction and Princeton and is the shortest scheduled commuter rail line in the United States. We also aim to make pre-training short and accessible to everyone.

Our Contributions

• DeepSpeed transformer kernel integration
• A training recipe for efficient MLM pre-training
• An easy-to-follow guideline of using fairseq for MLM pre-training.

Other fairseq features:

• Multi-GPU training on one machine or across multiple machines (data and model parallel)
• Gradient accumulation enables training with large mini-batches even on a single GPU
• Mixed precision training (trains faster with less GPU memory on NVIDIA tensor cores)
• Extensible: easily register new models, criterions, tasks, optimizers and learning rate schedulers
• Flexible configuration based on Hydra allowing a combination of code, command-line and file based configuration
• Full parameter and optimizer state sharding
• Offloading parameters to CPU

See the fairseq repo and its documentation for more details on how to use and extend fairseq.

DinkyTrain for Efficient MLM Pre-training

Quick Links

• Overview
• Installation
• Data Pre-processing
• Pre-training
• Fine-tuning on GLUE and SQuAD
• Convert to HuggingFace
• Model List
• Bugs or Questions?
• Citation

Overview

You can reproduce the pre-training experiments of our recent paper Should You Mask 15% in Masked Language Modeling?, where we find that higher masking rates can lead to more efficient pre-training.

Installation

• PyTorch version >= 1.5.0
• Python version >= 3.6
• To install fairseq and develop locally:

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install --editable ./

• For faster training (FP16) install NVIDIA's apex library:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
  --global-option="--fast_multihead_attn" ./

• For faster training (DeepSpeed cuda kernel) install DeepSpeed library and compile the DeepSpeed kernel

DS_BUILD_TRANSFORMER=1 DS_BUILD_STOCHASTIC_TRANSFORMER=1 pip install deepspeed

• For large datasets install PyArrow: pip install pyarrow
• If you use Docker make sure to increase the shared memory size either with --ipc=host or --shm-size as command line options to nvidia-docker run .

Trouble-shooting:

• If using lower version of Python, you might encounter import problems with importlib.metadata. Try pip install importlib-metadata.
• To install apex and deepspeed, you will need nvcc (CUDA compiler).
• When installing apex, if you encounter the error Cuda extensions are bing compiled with a version of Cuda that does not match ..., go to setup.py and comment out the line that raised the error (at your own risk).
• Both apex and deepspeed installation require a high gcc version to support c++14. If you encounter relevant errors, update your gcc.

Data Pre-processing

Tokenization: First, download the GPT2 BPE vocabulary:

wget -O gpt2_bpe/encoder.json https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json
wget -O gpt2_bpe/vocab.bpe https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe

Then, tokenize your raw data:

python -m examples.roberta.multiprocessing_bpe_encoder \
    --encoder-json gpt2_bpe/encoder.json \
    --vocab-bpe gpt2_bpe/vocab.bpe \
    --inputs ${SPLIT}.raw \
    --outputs ${SPLIT}.bpe \
    --keep-empty \
    --workers 8

Finally, index and binarize your data:

fairseq-preprocess \
    --only-source \
    --srcdict gpt2_bpe/dict.txt \
    --trainpref ${TRAIN_SPLIT}.bpe \
    --validpref ${VALID_SPLIT}.bpe \
    --testpref ${TEST_SPLIT}.bpe \
    --destdir output-bin \
    --workers 8

Alternatively: Use our pre-processed data: We preprocessed Wikipedia+BookCorpus and shared it on Huggingface dataset. It is ~22GB and contains two epochs of data, each epoch being sliced into 8 shards. You can download it using git:

git lfs install # Git lfs is needed for downloading
git clone https://huggingface.co/datasets/princeton-nlp/wikibook_fairseq_format

Pre-training

Use our script for efficient pre-training

GPU={number of GPUs} DATA_DIR={data path} [DEEPSPEED=1] bash run_efficient_mlm_recipe.sh

Flags explained

• GPU: number of GPUs.
• DATA_DIR: directory to the processed pre-training data. If you are using our preprocessed dataset, DATA_DIR should be:

DATA_DIR=$(seq 0 15 | sed -e 's/^/wikibook_fairseq_format\/bin-shard/' | sed -e 's/$/-8/' | paste -sd ':')

• DEEPSPEED (optional): if set to 1, the DeepSpeed CUDA kernel will be used.

Please refer to the script for more hyperparameter choices.

Fine-tuning on GLUE and SQuAD

All our checkpoints can be converted to HuggingFace transformers models (see next nextion) and use the transformers package for fine-tuning. Fairseq also supports fine-tuning on GLUE.

First, download the preprocessed GLUE data (you can also process by yourself following the preprocess section above):

git lfs install # Git lfs is needed for downloading
git clone https://huggingface.co/datasets/princeton-nlp/glue_fairseq_format

Then use the following script for fine-tuning

DATA_DIR={path to the data directory} \
TASK={glue task name (mnli qnli qqp rte sst2 mrpc cola stsb)} \
LR={learning rate} \
BSZ={batch size} \
EPOCHS={number of epochs} \
SEED={random seed} \
CKPT_DIR={checkpoint's directory} \
CKPT_NAME={checkpoint's name} \
[DEEPSPEED=1] bash finetune_glue.sh

For fine-tuning on SQuAD, please convert the models to HuggingFace checkpoints following the next section and use HuggingFace's examples.

Convert to HuggingFace

We also provide conversion codes so that you can easily turn Fairseq checkpoints into HuggingFace checkpoints. Usage:

cd scripts
[PRELAYERNORM=1] [FROM_DS=1] python convert_fs_ckpt_to_hf_ckpt.py --fr {fairseq checkpoint} --to {huggingface checkpoint path} --hf_model_config {roberta-base/roberta-large}

Flags explained:

• PRELAYERNORM=1: Using pre layer-norm (default is post layer-norm).
• FROM_DS=1: The Fairseq checkpoint uses DeepSpeed's cuda kernel.
• --fr: The path to the Fairseq checkpoint.
• --to: The path you want to save the HuggingFace checkpoint to.
• --hf_model_config: roberta-base or roberta-large.

IMPORTANT: all our models use pre layer norm, which is not supported by HuggingFace yet. To use it, import the model class from huggingface/modeling_roberta_prelayernorm.py. For example:

from huggingface.modeling_roberta_prelayernorm import RobertaForSequenceClassification

For more configuration, please refer to convert_fs_ckpt_to_hf_ckpt.py.

Model List

Here are the HuggingFace checkpoints of our models in the paper Should You Mask 15% in Masked Language Modeling. Results are development set performance.

We also offer the original (deepspeed) fairseq checkpoints here.

Bugs or Questions?

If you hav an questions, or encounter any problems when using the code, or want to report a bug, you can open an issue. Please try to specify the problem with details so we can help you better and quicker!

Citation

@article{wettig2022should,
   title={Should You Mask 15% in Masked Language Modeling?},
   author={Wettig, Alexander and Gao, Tianyu and Zhong, Zexuan and Chen, Danqi},
   boo={arXiv preprint arXiv:2202.08005},
   year={2022}
}

Acknowledgment

• Our package is based on fairseq:

Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53.

• Our efficient training recipe is based on the following paper:

Peter Izsak, Moshe Berchansky, and Omer Levy. 2021. How to train BERT with an academic budget. In Empirical Methods in Natural Language Processing (EMNLP), pages 10644–10652.