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
An open source AutoML toolkit for automate machine learning lifecycle, including feature engineering, neural architecture search, model compression and hyper-parameter tuning.

NNI Doc | 简体中文 NNI (Neural Network Intelligence) is a lightweight but powerful toolkit to help users automate Feature Engineering, Neural Architecture

Microsoft 12.4k Dec 31, 2022
MADT: Offline Pre-trained Multi-Agent Decision Transformer

MADT: Offline Pre-trained Multi-Agent Decision Transformer A link to our paper can be found on Arxiv. Overview Official codebase for Offline Pre-train

Linghui Meng 51 Dec 21, 2022
GNN-based Recommendation Benchma

GRecX A Fair Benchmark for GNN-based Recommendation Preliminary Comparison DiffNet-Yelp dataset (featureless) Algo 73 Oct 17, 2022

Rethinking the Importance of Implementation Tricks in Multi-Agent Reinforcement Learning

RIIT Our open-source code for RIIT: Rethinking the Importance of Implementation Tricks in Multi-AgentReinforcement Learning. We implement and standard

405 Jan 06, 2023
YouRefIt: Embodied Reference Understanding with Language and Gesture

YouRefIt: Embodied Reference Understanding with Language and Gesture YouRefIt: Embodied Reference Understanding with Language and Gesture by Yixin Che

16 Jul 11, 2022
EMNLP 2021: Single-dataset Experts for Multi-dataset Question-Answering

MADE (Multi-Adapter Dataset Experts) This repository contains the implementation of MADE (Multi-adapter dataset experts), which is described in the pa

Princeton Natural Language Processing 68 Jul 18, 2022
A library for performing coverage guided fuzzing of neural networks

TensorFuzz: Coverage Guided Fuzzing for Neural Networks This repository contains a library for performing coverage guided fuzzing of neural networks,

Brain Research 195 Dec 28, 2022
Adversarial vulnerability of powerful near out-of-distribution detection

Adversarial vulnerability of powerful near out-of-distribution detection by Stanislav Fort In this repository we're collecting replications for the ke

Stanislav Fort 9 Aug 30, 2022
Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset (CVPR'19)

Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset (CVPR'19) Tianyu Wang*, Xin Yang*, Ke Xu, Shaozhe Chen, Qiang Zhang, Ry

Steve Wong 177 Dec 01, 2022
General purpose Slater-Koster tight-binding code for electronic structure calculations

tight-binder Introduction General purpose tight-binding code for electronic structure calculations based on the Slater-Koster approximation. The code

9 Dec 15, 2022
This is a simple face recognition mini project that was completed by a team of 3 members in 1 week's time

PeekingDuckling 1. Description This is an implementation of facial identification algorithm to detect and identify the faces of the 3 team members Cla

Eric Kwok 2 Jan 25, 2022
This repository provides data for the VAW dataset as described in the CVPR 2021 paper titled "Learning to Predict Visual Attributes in the Wild"

Visual Attributes in the Wild (VAW) This repository provides data for the VAW dataset as described in the CVPR 2021 Paper: Learning to Predict Visual

Adobe Research 36 Dec 30, 2022
Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021)

Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021) The implementation of Reducing Infromation Bottleneck for W

Jungbeom Lee 81 Dec 16, 2022
GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data

GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data By Shuchang Zhou, Taihong Xiao, Yi Yang, Dieqiao Feng, Qinyao He, W

Taihong Xiao 141 Apr 16, 2021
Text and code for the forthcoming second edition of Think Bayes, by Allen Downey.

Think Bayes 2 by Allen B. Downey The HTML version of this book is here. Think Bayes is an introduction to Bayesian statistics using computational meth

Allen Downey 1.5k Jan 08, 2023
Unsupervised Foreground Extraction via Deep Region Competition

Unsupervised Foreground Extraction via Deep Region Competition [Paper] [Code] The official code repository for NeurIPS 2021 paper "Unsupervised Foregr

28 Nov 06, 2022
Code for a real-time distributed cooperative slam(RDC-SLAM) system for ROS compatible platforms.

RDC-SLAM This repository contains code for a real-time distributed cooperative slam(RDC-SLAM) system for ROS compatible platforms. The system takes in

40 Nov 19, 2022
Hypersim: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding

The Hypersim Dataset For many fundamental scene understanding tasks, it is difficult or impossible to obtain per-pixel ground truth labels from real i

Apple 1.3k Jan 04, 2023
IEGAN — Official PyTorch Implementation Independent Encoder for Deep Hierarchical Unsupervised Image-to-Image Translation

IEGAN — Official PyTorch Implementation Independent Encoder for Deep Hierarchical Unsupervised Image-to-Image Translation Independent Encoder for Deep

30 Nov 05, 2022
Code for "Solving Graph-based Public Good Games with Tree Search and Imitation Learning"

Code for "Solving Graph-based Public Good Games with Tree Search and Imitation Learning" This is the code for the paper Solving Graph-based Public Goo

Victor-Alexandru Darvariu 3 Dec 05, 2022