ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

Overview

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:

Model Layers Hidden Size Params GLUE score (test set) Download
ELECTRA-Small 12 256 14M 77.4 link
ELECTRA-Base 12 768 110M 82.7 link
ELECTRA-Large 24 1024 335M 85.2 link

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:

  • --corpus-dir: A directory containing raw text files to turn into ELECTRA examples. A text file can contain multiple documents with empty lines separating them.
  • --vocab-file: File defining the wordpiece vocabulary.
  • --output-dir: Where to write out ELECTRA examples.
  • --max-seq-length: The number of tokens per example (128 by default).
  • --num-processes: If >1 parallelize across multiple processes (1 by default).
  • --blanks-separate-docs: Whether blank lines indicate document boundaries (True by default).
  • --do-lower-case/--no-lower-case: Whether to lower case the input text (True by default).

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

  • --data-dir: a directory where pre-training data, model weights, etc. are stored. By default, the training loads examples from <data-dir>/pretrain_tfrecords and a vocabulary from <data-dir>/vocab.txt.
  • --model-name: a name for the model being trained. Model weights will be saved in <data-dir>/models/<model-name> by default.
  • --hparams (optional): a JSON dict or path to a JSON file containing model hyperparameters, data paths, etc. See configure_pretraining.py for the supported hyperparameters.

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:

  • "debug": true trains a tiny ELECTRA model for a few steps.
  • "model_size": one of "small", "base", or "large": determines the size of the model
  • "electra_objective": false trains a model with masked language modeling instead of replaced token detection (essentially BERT with dynamic masking and no next-sentence prediction).
  • "num_train_steps": n controls how long the model is pre-trained for.
  • "pretrain_tfrecords": <paths> determines where the pre-training data is located. Note you need to specify the specific files not just the directory (e.g., <data-dir>/pretrain_tf_records/pretrain_data.tfrecord*)
  • "vocab_file": <path> and "vocab_size": n can be used to set a custom wordpiece vocabulary.
  • "learning_rate": lr, "train_batch_size": n, etc. can be used to change training hyperparameters
  • "model_hparam_overrides": {"hidden_size": n, "num_hidden_layers": m}, etc. can be used to changed the hyperparameters for the underlying transformer (the "model_size" flag sets the default values).

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:

  • --data-dir: a directory where data, model weights, etc. are stored. By default, the script loads finetuning data from <data-dir>/finetuning_data/<task-name> and a vocabulary from <data-dir>/vocab.txt.
  • --model-name: a name of the pre-trained model: the pre-trained weights should exist in data-dir/models/model-name.
  • --hparams: a JSON dict containing model hyperparameters, data paths, etc. (e.g., --hparams '{"task_names": ["rte"], "model_size": "base", "learning_rate": 1e-4, ...}'). See configure_pretraining.py for the supported hyperparameters. Instead of a dict, this can also be a path to a .json file containing the hyperparameters. You must specify the "task_names" and "model_size" (see examples below).

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:

  • "debug": true fine-tunes a tiny ELECTRA model for a few steps.
  • "task_names": ["task_name"]: specifies the tasks to train on. A list because the codebase nominally supports multi-task learning, (although be warned this has not been thoroughly tested).
  • "model_size": one of "small", "base", or "large": determines the size of the model; you must set this to the same size as the pre-trained model.
  • "do_train" and "do_eval": train and/or evaluate a model (both are set to true by default). For using "do_eval": true with "do_train": false, you need to specify the init_checkpoint, e.g., python3 run_finetuning.py --data-dir $DATA_DIR --model-name electra_base --hparams '{"model_size": "base", "task_names": ["mnli"], "do_train": false, "do_eval": true, "init_checkpoint": "<data-dir>/models/electra_base/finetuning_models/mnli_model_1"}'
  • "num_trials": n: If >1, does multiple fine-tuning/evaluation runs with different random seeds.
  • "learning_rate": lr, "train_batch_size": n, etc. can be used to change training hyperparameters.
  • "model_hparam_overrides": {"hidden_size": n, "num_hidden_layers": m}, etc. can be used to changed the hyperparameters for the underlying transformer (the "model_size" flag sets the default values).

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

  • Squad 1.1: Download the train and dev datasets and move them under $DATA_DIR/finetuning_data/squadv1/(train|dev).json
  • Squad 2.0: Download the datasets from the SQuAD Website and move them under $DATA_DIR/finetuning_data/squad/(train|dev).json
  • MRQA tasks: Download the data from here. Move the data to $DATA_DIR/finetuning_data/(newsqa|naturalqs|triviaqa|searchqa)/(train|dev).jsonl.

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).

CoLA SST MRPC STS QQP MNLI QNLI RTE SQuAD 1.1 SQuAD 2.0 Chunking
Metrics MCC Acc Acc Spearman Acc Acc Acc Acc EM EM F1
ELECTRA-Large 69.1 96.9 90.8 92.6 92.4 90.9 95.0 88.0 89.7 88.1 97.2
ELECTRA-Base 67.7 95.1 89.5 91.2 91.5 88.8 93.2 82.7 86.8 80.5 97.1
ELECTRA-Small 57.0 91.2 88.0 87.5 89.0 81.3 88.4 66.7 75.8 70.1 96.5
ELECTRA-Small-OWT 56.8 88.3 87.4 86.8 88.3 78.9 87.9 68.5 -- -- --

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 ([email protected]).

Owner
Google Research
Google Research
M3DSSD: Monocular 3D Single Stage Object Detector

M3DSSD: Monocular 3D Single Stage Object Detector Setup pytorch 0.4.1 Preparation Download the full KITTI detection dataset. Then place a softlink (or

mumianyuxin 64 Dec 27, 2022
Implementation of ICCV2021(Oral) paper - VMNet: Voxel-Mesh Network for Geodesic-aware 3D Semantic Segmentation

VMNet: Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation Created by Zeyu HU Introduction This work is based on our paper VMNet: Voxel-Mes

HU Zeyu 82 Dec 27, 2022
Implementation of Memory-Compressed Attention, from the paper "Generating Wikipedia By Summarizing Long Sequences"

Memory Compressed Attention Implementation of the Self-Attention layer of the proposed Memory-Compressed Attention, in Pytorch. This repository offers

Phil Wang 47 Dec 23, 2022
Pytorch implementation of "M-LSD: Towards Light-weight and Real-time Line Segment Detection"

M-LSD: Towards Light-weight and Real-time Line Segment Detection Pytorch implementation of "M-LSD: Towards Light-weight and Real-time Line Segment Det

123 Jan 04, 2023
Deep Sea Treasure Environment for Multi-Objective Optimization Research

DeepSeaTreasure Environment Installation In order to get started with this environment, you can install it using the following command: python3 -m pip

imec IDLab 6 Nov 14, 2022
🎯 A comprehensive gradient-free optimization framework written in Python

Solid is a Python framework for gradient-free optimization. It contains basic versions of many of the most common optimization algorithms that do not

Devin Soni 565 Dec 26, 2022
Tutorial for the PERFECTING FACTORY 5.0 WITH EDGE-POWERED AI workshop

Workshop Advantech Jetson Nano This tutorial has been designed for the PERFECTING FACTORY 5.0 WITH EDGE-POWERED AI workshop in collaboration with Adva

Edge Impulse 18 Nov 22, 2022
bio_inspired_min_nets_improve_the_performance_and_robustness_of_deep_networks

Code Submission for: Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks Run with docker To build a docker environment, chan

0 Dec 09, 2021
External Attention Network

Beyond Self-attention: External Attention using Two Linear Layers for Visual Tasks paper : https://arxiv.org/abs/2105.02358 Jittor code will come soon

MenghaoGuo 357 Dec 11, 2022
A PyTorch implementation of "CoAtNet: Marrying Convolution and Attention for All Data Sizes".

CoAtNet Overview This is a PyTorch implementation of CoAtNet specified in "CoAtNet: Marrying Convolution and Attention for All Data Sizes", arXiv 2021

Justin Wu 268 Jan 07, 2023
Code for "Adversarial attack by dropping information." (ICCV 2021)

AdvDrop Code for "AdvDrop: Adversarial Attack to DNNs by Dropping Information(ICCV 2021)." Human can easily recognize visual objects with lost informa

Ranjie Duan 52 Nov 10, 2022
A Python Package For System Identification Using NARMAX Models

SysIdentPy is a Python module for System Identification using NARMAX models built on top of numpy and is distributed under the 3-Clause BSD license. N

Wilson Rocha 175 Dec 25, 2022
SelfAugment extends MoCo to include automatic unsupervised augmentation selection.

SelfAugment extends MoCo to include automatic unsupervised augmentation selection. In addition, we've included the ability to pretrain on several new datasets and included a wandb integration.

Colorado Reed 24 Oct 26, 2022
Unofficial Implement PU-Transformer

PU-Transformer-pytorch Pytorch unofficial implementation of PU-Transformer (PU-Transformer: Point Cloud Upsampling Transformer) https://arxiv.org/abs/

Lee Hyung Jun 7 Sep 21, 2022
Differentiable rasterization applied to 3D model simplification tasks

nvdiffmodeling Differentiable rasterization applied to 3D model simplification tasks, as described in the paper: Appearance-Driven Automatic 3D Model

NVIDIA Research Projects 336 Dec 30, 2022
COD-Rank-Localize-and-Segment (CVPR2021)

COD-Rank-Localize-and-Segment (CVPR2021) Simultaneously Localize, Segment and Rank the Camouflaged Objects Full camouflage fixation training dataset i

JingZhang 52 Dec 20, 2022
Best practices for segmentation of the corporate network of any company

Best-practice-for-network-segmentation What is this? This project was created to publish the best practices for segmentation of the corporate network

2k Jan 07, 2023
A PyTorch Reimplementation of TecoGAN: Temporally Coherent GAN for Video Super-Resolution

TecoGAN-PyTorch Introduction This is a PyTorch reimplementation of TecoGAN: Temporally Coherent GAN for Video Super-Resolution (VSR). Please refer to

165 Dec 17, 2022
Little tool in python to watch anime from the terminal (the better way to watch anime)

ani-cli Script working again :), thanks to the fork by Dink4n for the alternative approach to by pass the captcha on gogoanime A cli to browse and wat

Harshith 4.5k Dec 31, 2022
DeepOBS: A Deep Learning Optimizer Benchmark Suite

DeepOBS - A Deep Learning Optimizer Benchmark Suite DeepOBS is a benchmarking suite that drastically simplifies, automates and improves the evaluation

Aaron Bahde 7 May 12, 2020