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
The Adapter-Bot: All-In-One Controllable Conversational Model

The Adapter-Bot: All-In-One Controllable Conversational Model This is the implementation of the paper: The Adapter-Bot: All-In-One Controllable Conver

CAiRE 37 Nov 04, 2022
An example to implement a new backbone with OpenMMLab framework.

Backbone example on OpenMMLab framework English | 简体中文 Introduction This is an template repo about how to use OpenMMLab framework to develop a new bac

Ma Zerun 22 Dec 29, 2022
Attempt at implementation of a simple GAN using Keras

Simple GAN This is my attempt to make a wrapper class for a GAN in keras which can be used to abstract the whole architecture process. Simple GAN Over

Deven96 7 May 23, 2019
A hyperparameter optimization framework

Optuna: A hyperparameter optimization framework Website | Docs | Install Guide | Tutorial Optuna is an automatic hyperparameter optimization software

7.4k Jan 04, 2023
Infrastructure as Code (IaC) for a self-hosted version of Gnosis Safe on AWS

Welcome to Yearn Gnosis Safe! Setting up your local environment Infrastructure Deploying Gnosis Safe Prerequisites 1. Create infrastructure for secret

Numan 16 Jul 18, 2022
HODEmu, is both an executable and a python library that is based on Ragagnin 2021 in prep.

HODEmu HODEmu, is both an executable and a python library that is based on Ragagnin 2021 in prep. and emulates satellite abundance as a function of co

Antonio Ragagnin 1 Oct 13, 2021
U-Net Implementation: Convolutional Networks for Biomedical Image Segmentation" using the Carvana Image Masking Dataset in PyTorch

U-Net Implementation By Christopher Ley This is my interpretation and implementation of the famous paper "U-Net: Convolutional Networks for Biomedical

Christopher Ley 1 Jan 06, 2022
FlingBot: The Unreasonable Effectiveness of Dynamic Manipulations for Cloth Unfolding

This repository contains code for training and evaluating FlingBot in both simulation and real-world settings on a dual-UR5 robot arm setup for Ubuntu 18.04

Columbia Artificial Intelligence and Robotics Lab 70 Dec 06, 2022
A fast, scalable, high performance Gradient Boosting on Decision Trees library, used for ranking, classification, regression and other machine learning tasks for Python, R, Java, C++. Supports computation on CPU and GPU.

Website | Documentation | Tutorials | Installation | Release Notes CatBoost is a machine learning method based on gradient boosting over decision tree

CatBoost 6.9k Jan 04, 2023
Numba-accelerated Pythonic implementation of MPDATA with examples in Python, Julia and Matlab

PyMPDATA PyMPDATA is a high-performance Numba-accelerated Pythonic implementation of the MPDATA algorithm of Smolarkiewicz et al. used in geophysical

Atmospheric Cloud Simulation Group @ Jagiellonian University 15 Nov 23, 2022
Video-Music Transformer

VMT Video-Music Transformer (VMT) is an attention-based multi-modal model, which generates piano music for a given video. Paper https://arxiv.org/abs/

Chin-Tung Lin 5 Jul 13, 2022
Code for the Paper "Diffusion Models for Handwriting Generation"

Code for the Paper "Diffusion Models for Handwriting Generation"

62 Dec 21, 2022
Official Implementation for "ReStyle: A Residual-Based StyleGAN Encoder via Iterative Refinement" https://arxiv.org/abs/2104.02699

ReStyle: A Residual-Based StyleGAN Encoder via Iterative Refinement Recently, the power of unconditional image synthesis has significantly advanced th

967 Jan 04, 2023
Official PyTorch implementation of "Evolving Search Space for Neural Architecture Search"

Evolving Search Space for Neural Architecture Search Usage Install all required dependencies in requirements.txt and replace all ..path/..to in the co

Yuanzheng Ci 10 Oct 24, 2022
Pure python implementation reverse-mode automatic differentiation

MiniGrad A minimal implementation of reverse-mode automatic differentiation (a.k.a. autograd / backpropagation) in pure Python. Inspired by Andrej Kar

Kenny Song 76 Sep 12, 2022
JAX code for the paper "Control-Oriented Model-Based Reinforcement Learning with Implicit Differentiation"

Optimal Model Design for Reinforcement Learning This repository contains JAX code for the paper Control-Oriented Model-Based Reinforcement Learning wi

Evgenii Nikishin 43 Sep 28, 2022
Embeds a story into a music playlist by sorting the playlist so that the order of the music follows a narrative arc.

playlist-story-builder This project attempts to embed a story into a music playlist by sorting the playlist so that the order of the music follows a n

Dylan R. Ashley 0 Oct 28, 2021
Implementation of the "Point 4D Transformer Networks for Spatio-Temporal Modeling in Point Cloud Videos" paper.

Point 4D Transformer Networks for Spatio-Temporal Modeling in Point Cloud Videos Introduction Point cloud videos exhibit irregularities and lack of or

Hehe Fan 101 Dec 29, 2022
Official Pytorch implementation of "DivCo: Diverse Conditional Image Synthesis via Contrastive Generative Adversarial Network" (CVPR'21)

DivCo: Diverse Conditional Image Synthesis via Contrastive Generative Adversarial Network Pytorch implementation for our DivCo. We propose a simple ye

64 Nov 22, 2022
Official implementation of FCL-taco2: Fast, Controllable and Lightweight version of Tacotron2 @ ICASSP 2021

FCL-Taco2: Towards Fast, Controllable and Lightweight Text-to-Speech synthesis (ICASSP 2021) Paper | Demo Block diagram of FCL-taco2, where the decode

Disong Wang 39 Sep 28, 2022