Code repository accompanying NAACL 2019 tutorial on "Transfer Learning in Natural Language Processing"

The tutorial was given on June 2 at NAACL 2019 in Minneapolis, MN, USA by Sebastian Ruder, Matthew Peters, Swabha Swayamdipta and Thomas Wolf.

Here is the webpage of NAACL tutorials for more information.

The slides for the tutorial can be found here: https://tinyurl.com/NAACLTransfer.

A Google Colab notebook with all the code for the tutorial can be found here: https://tinyurl.com/NAACLTransferColab.

The present repository can also be accessed with the following short url: https://tinyurl.com/NAACLTransferCode

Abstract

The classic supervised machine learning paradigm is based on learning in isolation a single predictive model for a task using a single dataset. This approach requires a large number of training examples and performs best for well-defined and narrow tasks. Transfer learning refers to a set of methods that extend this approach by leveraging data from additional domains or tasks to train a model with better generalization properties.

Over the last two years, the field of Natural Language Processing (NLP) has witnessed the emergence of several transfer learning methods and architectures, which significantly improved upon the state-of-the-art on a wide range of NLP tasks.

These improvements together with the wide availability and ease of integration of these methods are reminiscent of the factors that led to the success of pretrained word embeddings and ImageNet pretraining in computer vision, and indicate that these methods will likely become a common tool in the NLP landscape as well as an important research direction.

We will present an overview of modern transfer learning methods in NLP, how models are pre-trained, what information the representations they learn capture, and review examples and case studies on how these models can be integrated and adapted in downstream NLP tasks.

Overview

This codebase tries to present in the simplest and most compact way a few of the major Transfer Learning techniques, which have emerged over the past years. The code in this repository does not attempt to be state-of-the-art. However, effort has been made to achieve reasonable performance and with some modifications to be competitive with the current state of the art.

Special effort has been made to

• ensure the present code can be used as easily as possible, in particular by hosting pretrained models and datasets;
• keep the present codebase as compact and self-contained as possible to make it easy to manipulate and understand.

Currently the codebase comprises:

• pretraining_model.py: a transformer model with a GPT-2-like architecture as the basic pretrained model;
• pretraining_train.py: a pretraining script to train this model with a language modeling objective on a selection of large datasets (WikiText-103, SimpleBooks-92) using distributed training if available;
• finetuning_model.py: several architectures based on the transformer model for fine-tuning (with a classification head on top, with adapters);
• finetuning_train.py: a fine-tuning script to fine-tune these architectures on a classification task (IMDb).

Installation

To use this codebase, simply clone the Github repository and install the requirements like this:

git clone https://github.com/huggingface/naacl_transfer_learning_tutorial
cd naacl_transfer_learning_tutorial
pip install -r requirements.txt

Pre-training

To pre-train the transformer, run the pretraining_train.py script like this:

python ./pretraining_train.py

or using distributed training like this (for a 8 GPU server):

python -m torch.distributed.launch --nproc_per_node 8 ./pretraining_train.py

The pre-training script will:

• download wikitext-103 for pre-training (default),
• instantiate a 50M parameters transformer model and train it for 50 epochs,
• log the experiements in Tensorboard and in a folder under ./runs,
• save checkpoints in the log folder.

Pretraining to a validation perplexity of ~29 on WikiText-103 will take about 15h on 8 V100 GPUs (can be stopped earlier). If you are interested in SOTA, there are a few reasons the validation perplexity is a bit higher than the equivalent Transformer-XL perplexity (around 24). The main reason is the use of an open vocabulary (sub-words for Bert tokenizer) instead of a closed vocabulary (see this blog post by Sebastian Mielke for some explanation)

Various pre-training options are available, you can list them with:

python ./pretraining_train.py --help

Fine-tuning

To fine-tune the pre-trained transformer, run the finetuning_train.py script like this:

python ./finetuning_train.py --model_checkpoint PATH-TO-YOUR-PRETRAINED-MODEL-FOLDER

PATH-TO-YOUR-PRETRAINED-MODEL-FOLDER can be for instance ./runs/May17_17-47-12_my_big_server

or using distributed training like this (for a 8 GPU server):

python -m torch.distributed.launch --nproc_per_node 8 ./finetuning_train.py  --model_checkpoint PATH-TO-YOUR-PRETRAINED-MODEL-FOLDER

Various fine-tuning options are available, you can list them with:

python ./finetuning_train.py --help