SmallInitEmb - LayerNorm(SmallInit(Embedding)) in a Transformer to improve convergence

Overview

SmallInitEmb

LayerNorm(SmallInit(Embedding)) in a Transformer

I find that when training a transformer, the embedding matrix moves slowly, hence it's difficult for the model to jump out of the initial noisy embedding.

(initial embedding)
[[-0.0073  0.0062 -0.0261 ...  0.0086  0.0107 -0.008 ] ... ]
 (after 1 step, the directions of the embedding vectors are not moved much because the numbers change by ~LR = ~4e-4)
[[-0.0069  0.0066 -0.0265 ...  0.009   0.0111 -0.0084] ... ]

So I propose initializing the embedding matrix to tiny values, and put another LayerNorm after it (before all the SA & FFN layers):

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
if self.config.USE_SMALL_EMB and self.layer_id == 0:
    x = self.lnPre(x) # LN(SmallInit(Emb))
x = x + self.att(self.ln1(x))
x = x + self.ffn(self.ln2(x))

And then you get improved convergence (especially for BPE models) because the model can quickly jump out of the tiny initial embedding (small changes after 1 step -> significant changes of directions -> significant changes after LayerNorm).

Loss curve comparison: https://wandb.ai/blinkdl/SmallEmbTest

(the gap between LayerNorm(SmallEmb)) and baseline persists after more training)

Moreover, you can directly train PostLN models without warmup with SmallInit(Emb)

if isinstance(module, (nn.Embedding)):
    nn.init.uniform_(module.weight, a=-1e-4, b=1e-4) # SmallInit(Emb)
...
x = self.ln1(x) # this plays the same role as the lnPre in the above PreLN code
x = x + self.att(x)
x = self.ln2(x)
x = x + self.ffn(x)
(note you shall have another LN after the final ffn)
Owner
PENG Bo
http://zhihu.com/people/bopengbopeng
PENG Bo
ElasticFace: Elastic Margin Loss for Deep Face Recognition

This is the official repository of the paper: ElasticFace: Elastic Margin Loss for Deep Face Recognition Paper on arxiv: arxiv Model Log file Pretrain

Fadi Boutros 113 Dec 14, 2022
[CVPR 2022] CoTTA Code for our CVPR 2022 paper Continual Test-Time Domain Adaptation

CoTTA Code for our CVPR 2022 paper Continual Test-Time Domain Adaptation Prerequisite Please create and activate the following conda envrionment. To r

Qin Wang 87 Jan 08, 2023
Public implementation of "Learning from Suboptimal Demonstration via Self-Supervised Reward Regression" from CoRL'21

Self-Supervised Reward Regression (SSRR) Codebase for CoRL 2021 paper "Learning from Suboptimal Demonstration via Self-Supervised Reward Regression "

19 Dec 12, 2022
Back to the Feature: Learning Robust Camera Localization from Pixels to Pose (CVPR 2021)

Back to the Feature with PixLoc We introduce PixLoc, a neural network for end-to-end learning of camera localization from an image and a 3D model via

Computer Vision and Geometry Lab 610 Jan 05, 2023
Erpnext app for make employee salary on payroll entry based on one or more project with percentage for all project equal 100 %

Project Payroll this app for make payroll for employee based on projects like project on 30 % and project 2 70 % as account dimension it makes genral

Ibrahim Morghim 8 Jan 02, 2023
A PaddlePaddle version of Neural Renderer, refer to its PyTorch version

Neural 3D Mesh Renderer in PadddlePaddle A PaddlePaddle version of Neural Renderer, refer to its PyTorch version Install Run: pip install neural-rende

AgentMaker 13 Jul 12, 2022
The code for Bi-Mix: Bidirectional Mixing for Domain Adaptive Nighttime Semantic Segmentation

BiMix The code for Bi-Mix: Bidirectional Mixing for Domain Adaptive Nighttime Semantic Segmentation arxiv Framework: visualization results: Requiremen

stanley 18 Sep 18, 2022
Guiding evolutionary strategies by (inaccurate) differentiable robot simulators @ NeurIPS, 4th Robot Learning Workshop

Guiding Evolutionary Strategies by Differentiable Robot Simulators In recent years, Evolutionary Strategies were actively explored in robotic tasks fo

Vladislav Kurenkov 4 Dec 14, 2021
WSDM2022 "A Simple but Effective Bidirectional Extraction Framework for Relational Triple Extraction"

BiRTE WSDM2022 "A Simple but Effective Bidirectional Extraction Framework for Relational Triple Extraction" Requirements The main requirements are: py

9 Dec 27, 2022
Code for "Adversarial Attack Generation Empowered by Min-Max Optimization", NeurIPS 2021

Min-Max Adversarial Attacks [Paper] [arXiv] [Video] [Slide] Adversarial Attack Generation Empowered by Min-Max Optimization Jingkang Wang, Tianyun Zha

Jingkang Wang 12 Nov 23, 2022
Multiwavelets-based operator model

Multiwavelet model for Operator maps Gaurav Gupta, Xiongye Xiao, and Paul Bogdan Multiwavelet-based Operator Learning for Differential Equations In Ne

Gaurav 33 Dec 04, 2022
基于Paddle框架的arcface复现

arcface-Paddle 基于Paddle框架的arcface复现 ArcFace-Paddle 本项目基于paddlepaddle框架复现ArcFace,并参加百度第三届论文复现赛,将在2021年5月15日比赛完后提供AIStudio链接~敬请期待 参考项目: InsightFace Padd

QuanHao Guo 16 Dec 15, 2022
Source Code for our paper: Understand me, if you refer to Aspect Knowledge: Knowledge-aware Gated Recurrent Memory Network

KaGRMN-DSG_ABSA This repository contains the PyTorch source Code for our paper: Understand me, if you refer to Aspect Knowledge: Knowledge-aware Gated

XingBowen 4 May 20, 2022
Code for the paper Open Sesame: Getting Inside BERT's Linguistic Knowledge.

Open Sesame This repository contains the code for the paper Open Sesame: Getting Inside BERT's Linguistic Knowledge. Credits We built the project on t

9 Jul 24, 2022
AI4Good project for detecting waste in the environment

Detect waste AI4Good project for detecting waste in environment. www.detectwaste.ml. Our latest results were published in Waste Management journal in

108 Dec 25, 2022
Towards the D-Optimal Online Experiment Design for Recommender Selection (KDD 2021)

Towards the D-Optimal Online Experiment Design for Recommender Selection (KDD 2021) Contact 0 Jan 11, 2022

DECAF: Generating Fair Synthetic Data Using Causally-Aware Generative Networks

DECAF (DEbiasing CAusal Fairness) Code Author: Trent Kyono This repository contains the code used for the "DECAF: Generating Fair Synthetic Data Using

van_der_Schaar \LAB 7 Nov 24, 2022
List some popular DeepFake models e.g. DeepFake, FaceSwap-MarekKowal, IPGAN, FaceShifter, FaceSwap-Nirkin, FSGAN, SimSwap, CihaNet, etc.

deepfake-models List some popular DeepFake models e.g. DeepFake, CihaNet, SimSwap, FaceSwap-MarekKowal, IPGAN, FaceShifter, FaceSwap-Nirkin, FSGAN, Si

Mingcan Xiang 100 Dec 17, 2022
The first public PyTorch implementation of Attentive Recurrent Comparators

arc-pytorch PyTorch implementation of Attentive Recurrent Comparators by Shyam et al. A blog explaining Attentive Recurrent Comparators Visualizing At

Sanyam Agarwal 150 Oct 14, 2022
Open source annotation tool for machine learning practitioners.

doccano doccano is an open source text annotation tool for humans. It provides annotation features for text classification, sequence labeling and sequ

7.1k Jan 01, 2023