Home

Awesome

ELECTRA

Introduction

ELECTRA is a method for self-supervised language representation learning. It can be used to pre-train transformer networks using relatively little compute. ELECTRA models are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a GAN. At small scale, ELECTRA achieves strong results even when trained on a single GPU. At large scale, ELECTRA achieves state-of-the-art results on the SQuAD 2.0 dataset.

For a detailed description and experimental results, please refer to our ICLR 2020 paper ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators.

This repository contains code to pre-train ELECTRA, including small ELECTRA models on a single GPU. It also supports fine-tuning ELECTRA on downstream tasks including classification tasks (e.g,. GLUE), QA tasks (e.g., SQuAD), and sequence tagging tasks (e.g., text chunking).

This repository also contains code for Electric, a version of ELECTRA inspired by energy-based models. Electric provides a more principled view of ELECTRA as a "negative sampling" cloze model. It can also efficiently produce pseudo-likelihood scores for text, which can be used to re-rank the outputs of speech recognition or machine translation systems. For details on Electric, please refer to out EMNLP 2020 paper Pre-Training Transformers as Energy-Based Cloze Models.

Released Models

We are initially releasing three pre-trained models:

ModelLayersHidden SizeParamsGLUE score (test set)Download
ELECTRA-Small1225614M77.4link
ELECTRA-Base12768110M82.7link
ELECTRA-Large241024335M85.2link

The models were trained on uncased English text. They correspond to ELECTRA-Small++, ELECTRA-Base++, ELECTRA-1.75M in our paper. We hope to release other models, such as multilingual models, in the future.

On GLUE, ELECTRA-Large scores slightly better than ALBERT/XLNET, ELECTRA-Base scores better than BERT-Large, and ELECTRA-Small scores slightly worst than TinyBERT (but uses no distillation). See the expected results section below for detailed performance numbers.

Requirements

Pre-training

Use build_pretraining_dataset.py to create a pre-training dataset from a dump of raw text. It has the following arguments:

Use run_pretraining.py to pre-train an ELECTRA model. It has the following arguments:

If training is halted, re-running the run_pretraining.py with the same arguments will continue the training where it left off.

You can continue pre-training from the released ELECTRA checkpoints by

  1. Setting the model-name to point to a downloaded model (e.g., --model-name electra_small if you downloaded weights to $DATA_DIR/electra_small).
  2. Setting num_train_steps by (for example) adding "num_train_steps": 4010000 to the --hparams. This will continue training the small model for 10000 more steps (it has already been trained for 4e6 steps).
  3. Increase the learning rate to account for the linear learning rate decay. For example, to start with a learning rate of 2e-4 you should set the learning_rate hparam to 2e-4 * (4e6 + 10000) / 10000.
  4. For ELECTRA-Small, you also need to specifiy "generator_hidden_size": 1.0 in the hparams because we did not use a small generator for that model.

Quickstart: Pre-train a small ELECTRA model.

These instructions pre-train a small ELECTRA model (12 layers, 256 hidden size). Unfortunately, the data we used in the paper is not publicly available, so we will use the OpenWebTextCorpus released by Aaron Gokaslan and Vanya Cohen instead. The fully-trained model (~4 days on a v100 GPU) should perform roughly in between GPT and BERT-Base in terms of GLUE performance. By default the model is trained on length-128 sequences, so it is not suitable for running on question answering. See the "expected results" section below for more details on model performance.

Setup

  1. Place a vocabulary file in $DATA_DIR/vocab.txt. Our ELECTRA models all used the exact same vocabulary as English uncased BERT, which you can download here.
  2. Download the OpenWebText corpus (12G) and extract it (i.e., run tar xf openwebtext.tar.xz). Place it in $DATA_DIR/openwebtext.
  3. Run python3 build_openwebtext_pretraining_dataset.py --data-dir $DATA_DIR --num-processes 5. It pre-processes/tokenizes the data and outputs examples as tfrecord files under $DATA_DIR/pretrain_tfrecords. The tfrecords require roughly 30G of disk space.

Pre-training the model.

Run python3 run_pretraining.py --data-dir $DATA_DIR --model-name electra_small_owt to train a small ELECTRA model for 1 million steps on the data. This takes slightly over 4 days on a Tesla V100 GPU. However, the model should achieve decent results after 200k steps (10 hours of training on the v100 GPU).

To customize the training, add --hparams '{"hparam1": value1, "hparam2": value2, ...}' to the run command. --hparams can also be a path to a .json file containing the hyperparameters. Some particularly useful options:

See configure_pretraining.py for the full set of supported hyperparameters.

Evaluating the pre-trained model.

To evaluate the model on a downstream task, see the below finetuning instructions. To evaluate the generator/discriminator on the openwebtext data run python3 run_pretraining.py --data-dir $DATA_DIR --model-name electra_small_owt --hparams '{"do_train": false, "do_eval": true}'. This will print out eval metrics such as the accuracy of the generator and discriminator, and also writing the metrics out to data-dir/model-name/results.

Fine-tuning

Use run_finetuning.py to fine-tune and evaluate an ELECTRA model on a downstream NLP task. It expects three arguments:

Eval metrics will be saved in data-dir/model-name/results and model weights will be saved in data-dir/model-name/finetuning_models by default. Evaluation is done on the dev set by default. To customize the training, add --hparams '{"hparam1": value1, "hparam2": value2, ...}' to the run command. Some particularly useful options:

Setup

Get a pre-trained ELECTRA model either by training your own (see pre-training instructions above), or downloading the release ELECTRA weights and unziping them under $DATA_DIR/models (e.g., you should have a directory$DATA_DIR/models/electra_large if you are using the large model).

Finetune ELECTRA on a GLUE task

Download the GLUE data by running this script. Set up the data by running mv CoLA cola && mv MNLI mnli && mv MRPC mrpc && mv QNLI qnli && mv QQP qqp && mv RTE rte && mv SST-2 sst && mv STS-B sts && mv diagnostic/diagnostic.tsv mnli && mkdir -p $DATA_DIR/finetuning_data && mv * $DATA_DIR/finetuning_data.

Then run run_finetuning.py. For example, to fine-tune ELECTRA-Base on MNLI

python3 run_finetuning.py --data-dir $DATA_DIR --model-name electra_base --hparams '{"model_size": "base", "task_names": ["mnli"]}'

Or fine-tune a small model pre-trained using the above instructions on CoLA.

python3 run_finetuning.py --data-dir $DATA_DIR --model-name electra_small_owt --hparams '{"model_size": "small", "task_names": ["cola"]}'

Finetune ELECTRA on question answering

The code supports SQuAD 1.1 and 2.0, as well as datasets in the 2019 MRQA shared task

Then run (for example)

python3 run_finetuning.py --data-dir $DATA_DIR --model-name electra_base --hparams '{"model_size": "base", "task_names": ["squad"]}'

This repository uses the official evaluation code released by the SQuAD authors and the MRQA shared task to compute metrics

Finetune ELECTRA on sequence tagging

Download the CoNLL-2000 text chunking dataset from here and put it under $DATA_DIR/finetuning_data/chunk/(train|dev).txt. Then run

python3 run_finetuning.py --data-dir $DATA_DIR --model-name electra_base --hparams '{"model_size": "base", "task_names": ["chunk"]}'

Adding a new task

The easiest way to run on a new task is to implement a new finetune.task.Task, add it to finetune.task_builder.py, and then use run_finetuning.py as normal. For classification/qa/sequence tagging, you can inherit from a finetune.classification.classification_tasks.ClassificationTask, finetune.qa.qa_tasks.QATask, or finetune.tagging.tagging_tasks.TaggingTask. For preprocessing data, we use the same tokenizer as BERT.

Expected Results

Here are expected results for ELECTRA on various tasks (test set for chunking, dev set for the other tasks). Note that variance in fine-tuning can be quite large, so for some tasks you may see big fluctuations in scores when fine-tuning from the same checkpoint multiple times. The below scores show median performance over a large number of random seeds. ELECTRA-Small/Base/Large are our released models. ELECTRA-Small-OWT is the OpenWebText-trained model from above (it performs a bit worse than ELECTRA-Small due to being trained for less time and on a smaller dataset).

CoLASSTMRPCSTSQQPMNLIQNLIRTESQuAD 1.1SQuAD 2.0Chunking
MetricsMCCAccAccSpearmanAccAccAccAccEMEMF1
ELECTRA-Large69.196.990.892.692.490.995.088.089.788.197.2
ELECTRA-Base67.795.189.591.291.588.893.282.786.880.597.1
ELECTRA-Small57.091.288.087.589.081.388.466.775.870.196.5
ELECTRA-Small-OWT56.888.387.486.888.378.987.968.5------

See here for losses / training curves of the models during pre-training.

Electric

To train Electric, use the same pre-training script and command as ELECTRA. Pass "electra_objective": false and "electric_objective": true to the hyperparameters. We plan to release pre-trained Electric models soon!

Citation

If you use this code for your publication, please cite the original paper:

@inproceedings{clark2020electra,
  title = {{ELECTRA}: Pre-training Text Encoders as Discriminators Rather Than Generators},
  author = {Kevin Clark and Minh-Thang Luong and Quoc V. Le and Christopher D. Manning},
  booktitle = {ICLR},
  year = {2020},
  url = {https://openreview.net/pdf?id=r1xMH1BtvB}
}

If you use the code for Electric, please cite the Electric paper:

@inproceedings{clark2020electric,
  title = {Pre-Training Transformers as Energy-Based Cloze Models},
  author = {Kevin Clark and Minh-Thang Luong and Quoc V. Le and Christopher D. Manning},
  booktitle = {EMNLP},
  year = {2020},
  url = {https://www.aclweb.org/anthology/2020.emnlp-main.20.pdf}
}

Contact Info

For help or issues using ELECTRA, please submit a GitHub issue.

For personal communication related to ELECTRA, please contact Kevin Clark (kevclark@cs.stanford.edu).