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)
You might also like...
RTS3D: Real-time Stereo 3D Detection from 4D Feature-Consistency Embedding Space for Autonomous Driving

RTS3D: Real-time Stereo 3D Detection from 4D Feature-Consistency Embedding Space for Autonomous Driving (AAAI2021). RTS3D is efficiency and accuracy s

the code used for the preprint Embedding-based Instance Segmentation of Microscopy Images.
the code used for the preprint Embedding-based Instance Segmentation of Microscopy Images.

EmbedSeg Introduction This repository hosts the version of the code used for the preprint Embedding-based Instance Segmentation of Microscopy Images.

 UMEC: Unified Model and Embedding Compression for Efficient Recommendation Systems
UMEC: Unified Model and Embedding Compression for Efficient Recommendation Systems

[ICLR 2021] "UMEC: Unified Model and Embedding Compression for Efficient Recommendation Systems" by Jiayi Shen, Haotao Wang*, Shupeng Gui*, Jianchao Tan, Zhangyang Wang, and Ji Liu

Y. Zhang, Q. Yao, W. Dai, L. Chen. AutoSF: Searching Scoring Functions for Knowledge Graph Embedding. IEEE International Conference on Data Engineering (ICDE). 2020
Y. Zhang, Q. Yao, W. Dai, L. Chen. AutoSF: Searching Scoring Functions for Knowledge Graph Embedding. IEEE International Conference on Data Engineering (ICDE). 2020

AutoSF The code for our paper "AutoSF: Searching Scoring Functions for Knowledge Graph Embedding" and this paper has been accepted by ICDE2020. News:

Code for
Code for "Learning the Best Pooling Strategy for Visual Semantic Embedding", CVPR 2021

Learning the Best Pooling Strategy for Visual Semantic Embedding Official PyTorch implementation of the paper Learning the Best Pooling Strategy for V

Visual Tracking by TridenAlign and Context Embedding
Visual Tracking by TridenAlign and Context Embedding

Visual Tracking by TridentAlign and Context Embedding (TACT) Test code for "Visual Tracking by TridentAlign and Context Embedding" Janghoon Choi, Juns

Paddle implementation for "Highly Efficient Knowledge Graph Embedding Learning with Closed-Form Orthogonal Procrustes Analysis" (NAACL 2021)

ProcrustEs-KGE Paddle implementation for Highly Efficient Knowledge Graph Embedding Learning with Orthogonal Procrustes Analysis 🙈 A more detailed re

Paddle implementation for "Cross-Lingual Word Embedding Refinement by â„“1 Norm Optimisation" (NAACL 2021)

L1-Refinement Paddle implementation for "Cross-Lingual Word Embedding Refinement by ℓ1 Norm Optimisation" (NAACL 2021) 🙈 A more detailed readme is co

HNECV: Heterogeneous Network Embedding via Cloud model and Variational inference
HNECV: Heterogeneous Network Embedding via Cloud model and Variational inference

HNECV This repository provides a reference implementation of HNECV as described in the paper: HNECV: Heterogeneous Network Embedding via Cloud model a

Owner
PENG Bo
http://zhihu.com/people/bopengbopeng
PENG Bo
Seach Losses of our paper 'Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search', accepted by ICLR 2021.

CSE-Autoloss Designing proper loss functions for vision tasks has been a long-standing research direction to advance the capability of existing models

Peidong Liu(刘沛东) 54 Dec 17, 2022
Pytorch implementation of AngularGrad: A New Optimization Technique for Angular Convergence of Convolutional Neural Networks

AngularGrad Optimizer This repository contains the oficial implementation for AngularGrad: A New Optimization Technique for Angular Convergence of Con

mario 124 Sep 16, 2022
Official implementation of the ICCV 2021 paper "Conditional DETR for Fast Training Convergence".

The DETR approach applies the transformer encoder and decoder architecture to object detection and achieves promising performance. In this paper, we handle the critical issue, slow training convergence, and present a conditional cross-attention mechanism for fast DETR training. Our approach is motivated by that the cross-attention in DETR relies highly on the content embeddings and that the spatial embeddings make minor contributions, increasing the need for high-quality content embeddings and thus increasing the training difficulty.

null 281 Dec 30, 2022
Deep Text Search is an AI-powered multilingual text search and recommendation engine with state-of-the-art transformer-based multilingual text embedding (50+ languages).

Deep Text Search - AI Based Text Search & Recommendation System Deep Text Search is an AI-powered multilingual text search and recommendation engine w

null 19 Sep 29, 2022
VSR-Transformer - This paper proposes a new Transformer for video super-resolution (called VSR-Transformer).

VSR-Transformer By Jiezhang Cao, Yawei Li, Kai Zhang, Luc Van Gool This paper proposes a new Transformer for video super-resolution (called VSR-Transf

Jiezhang Cao 225 Nov 13, 2022
Keep CALM and Improve Visual Feature Attribution

Keep CALM and Improve Visual Feature Attribution Jae Myung Kim1*, Junsuk Choe1*, Zeynep Akata2, Seong Joon Oh1† * Equal contribution † Corresponding a

NAVER AI 90 Dec 7, 2022
Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening

Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening Introduction This is an implementation of the model used for breast

null 757 Dec 30, 2022
Official implementation of "Open-set Label Noise Can Improve Robustness Against Inherent Label Noise" (NeurIPS 2021)

Open-set Label Noise Can Improve Robustness Against Inherent Label Noise NeurIPS 2021: This repository is the official implementation of ODNL. Require

Hongxin Wei 12 Dec 7, 2022
Use graph-based analysis to re-classify stocks and to improve Markowitz portfolio optimization

Dynamic Stock Industrial Classification Use graph-based analysis to re-classify stocks and experiment different re-classification methodologies to imp

Sheng Yang 10 Dec 5, 2022