DeLighT: Very Deep and Light-Weight Transformers

Related tags

Deep Learningdelight
Overview

DeLighT: Very Deep and Light-weight Transformers

This repository contains the source code of our work on building efficient sequence models: DeFINE (ICLR'20) and DeLighT (preprint).

Table of contents

  1. Overview
  2. Requirements and installation
  3. Training, evaluation, and results
  4. Multiplication-addition operations
  5. Citation
  6. Acknowledgement
  7. Issues

Overview

In this repository, we share the source code of our paper DeLight, that delivers similar or better performance than transformer-based models with significantly fewer parameters. DeLighT more efficiently allocates parameters both (1) within each Transformer block using DExTra, a deep and light-weight transformation and (2) across blocks using block-wise scaling, that allows for shallower and narrower DeLighT blocks near the input and wider and deeper DeLighT blocks near the output. Overall, DeLighT networks are 2.5 to 4 times deeper than standard transformer models and yet have fewer parameters and operations. For details, see our papers: DeFINE and and DeLighT.

DeLighT unit

Requirements and Installation

  • PyTorch version >= 1.4.0
  • Python version >= 3.6
  • For training new models, you'll also need an NVIDIA GPU and NCCL
  • To use DeLighT, you need to install fairseq and develop locally:
git clone https://github.com/sacmehta/delight
cd delight
pip install --editable ./
  • For faster training 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" ./

Training, Evaluation, and Results

For training, evaluation, and results, see below links. To ease reproduction of our results, we also provide links to training logs.

Neural machine translation

Language Modeling

Multiplication-Addition Operations

We have added module profiling for both Transformer and DeLight networks. This can be enabled using --print-stats argument. A model summary will be printed (by default for 20 tokens), similar to below screenshot. To use larger sequence lengths for source and target for profiling statistics, you can use --src-len-ps and --tgt-len-ps flags.

Model statistics

Citation

If you find our work useful, please consider citing following works:

@misc{mehta2020delight,
    title={DeLighT: Very Deep and Light-weight Transformer},
    author={Sachin Mehta and Marjan Ghazvininejad and Srinivasan Iyer and Luke Zettlemoyer and Hannaneh Hajishirzi},
    year={2020},
    eprint={2008.00623},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}
@inproceedings{mehta2019define,
  title={DeFINE: Deep Factorized Input Token Embeddings for Neural Sequence Modeling},
  author={Mehta, Sachin and Koncel-Kedziorski, Rik and Rastegari, Mohammad and Hajishirzi, Hannaneh},
  booktitle={International Conference on Learning Representations},
  year={2019}
}

Acknowledgements

We would like to thank Fairseq team for building easy-to-use sequence library.

Issues

Thanks for your interest in our work. For any issues, please raise a request.

Owner
Sachin Mehta
Ph.D. Student at University of Washington
Sachin Mehta
✨风纪委员会自动投票脚本,利用Github Action帮你进行裁决操作(为了让其他风纪委员有案件可判,本程序从中午12点才开始运行,有需要请自己修改运行时间)

风纪委员会自动投票 本脚本通过使用Github Action来实现B站风纪委员的自动投票功能,喜欢请给我点个STAR吧! 如果你不是风纪委员,在符合风纪委员申请条件的情况下,本脚本会自动帮你申请 投票时间是早上八点,如果有需要请自行修改.github/workflows/Judge.yml中的时间,

Pesy Wu 25 Feb 17, 2021
Pytorch port of Google Research's LEAF Audio paper

leaf-audio-pytorch Pytorch port of Google Research's LEAF Audio paper published at ICLR 2021. This port is not completely finished, but the Leaf() fro

Dennis Fedorishin 80 Oct 31, 2022
A big endian Gentoo port developed on a Pine64.org RockPro64

Gentoo-aarch64_be A big endian Gentoo port developed on a Pine64.org RockPro64 The endian wars are over... little endian won. As a result, it is incre

Rory Bolt 6 Dec 07, 2022
Reimplementation of NeurIPS'19: "Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting" by Shu et al.

[Re] Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting Reimplementation of NeurIPS'19: "Meta-Weight-Net: Learning an Explicit Mapping

Robert Cedergren 1 Mar 13, 2020
Finding Donors for CharityML

Finding-Donors-for-CharityML - Investigated factors that affect the likelihood of charity donations being made based on real census data.

Moamen Abdelkawy 1 Dec 30, 2021
Proto-RL: Reinforcement Learning with Prototypical Representations

Proto-RL: Reinforcement Learning with Prototypical Representations This is a PyTorch implementation of Proto-RL from Reinforcement Learning with Proto

Denis Yarats 74 Dec 06, 2022
ChebLieNet, a spectral graph neural network turned equivariant by Riemannian geometry on Lie groups.

ChebLieNet: Invariant spectral graph NNs turned equivariant by Riemannian geometry on Lie groups Hugo Aguettaz, Erik J. Bekkers, Michaël Defferrard We

haguettaz 12 Dec 10, 2022
MoCoPnet - Deformable 3D Convolution for Video Super-Resolution

Deformable 3D Convolution for Video Super-Resolution Pytorch implementation of l

Xinyi Ying 28 Dec 15, 2022
render sprites into your desktop environment as shaped windows using GTK

spritegtk render static or animated sprites into your desktop environment as dynamic shaped windows using GTK requires pycairo and PYGobject: pip inst

hermit 20 Oct 27, 2022
This is an official source code for implementation on Extensive Deep Temporal Point Process

Extensive Deep Temporal Point Process This is an official source code for implementation on Extensive Deep Temporal Point Process, which is composed o

Haitao Lin 8 Aug 15, 2022
coldcuts is an R package to automatically generate and plot segmentation drawings in R

coldcuts coldcuts is an R package that allows you to draw and plot automatically segmentations from 3D voxel arrays. The name is inspired by one of It

2 Sep 03, 2022
NAS-HPO-Bench-II is the first benchmark dataset for joint optimization of CNN and training HPs.

NAS-HPO-Bench-II API Overview NAS-HPO-Bench-II is the first benchmark dataset for joint optimization of CNN and training HPs. It helps a fair and low-

yoichi hirose 8 Nov 21, 2022
Fully-automated scripts for collecting AI-related papers

AI-Paper-collector Fully-automated scripts for collecting AI-related papers List of Conferences to crawel ACL: 21-19 (including findings) EMNLP: 21-19

Gordon Lee 776 Jan 08, 2023
Breast cancer is been classified into benign tumour and malignant tumour.

Breast cancer is been classified into benign tumour and malignant tumour. Logistic regression is applied in this model.

1 Feb 04, 2022
Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics.

Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics. By Andres Milioto @ University of Bonn. (for the new P

Photogrammetry & Robotics Bonn 314 Dec 30, 2022
ConE: Cone Embeddings for Multi-Hop Reasoning over Knowledge Graphs

ConE: Cone Embeddings for Multi-Hop Reasoning over Knowledge Graphs This is the code of paper ConE: Cone Embeddings for Multi-Hop Reasoning over Knowl

MIRA Lab 33 Dec 07, 2022
The Self-Supervised Learner can be used to train a classifier with fewer labeled examples needed using self-supervised learning.

Published by SpaceML • About SpaceML • Quick Colab Example Self-Supervised Learner The Self-Supervised Learner can be used to train a classifier with

SpaceML 92 Nov 30, 2022
A framework that constructs deep neural networks, autoencoders, logistic regressors, and linear networks

A framework that constructs deep neural networks, autoencoders, logistic regressors, and linear networks without the use of any outside machine learning libraries - all from scratch.

Kordel K. France 2 Nov 14, 2022
Simple tools for logging and visualizing, loading and training

TNT TNT is a library providing powerful dataloading, logging and visualization utilities for Python. It is closely integrated with PyTorch and is desi

1.5k Jan 02, 2023
ADOP: Approximate Differentiable One-Pixel Point Rendering

ADOP: Approximate Differentiable One-Pixel Point Rendering Abstract: We present a novel point-based, differentiable neural rendering pipeline for scen

Darius Rückert 1.9k Jan 06, 2023