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
Official code of "R2RNet: Low-light Image Enhancement via Real-low to Real-normal Network."

R2RNet Official code of "R2RNet: Low-light Image Enhancement via Real-low to Real-normal Network." Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu

77 Dec 24, 2022
An essential implementation of BYOL in PyTorch + PyTorch Lightning

Essential BYOL A simple and complete implementation of Bootstrap your own latent: A new approach to self-supervised Learning in PyTorch + PyTorch Ligh

Enrico Fini 48 Sep 27, 2022
Official PyTorch repo for JoJoGAN: One Shot Face Stylization

JoJoGAN: One Shot Face Stylization This is the PyTorch implementation of JoJoGAN: One Shot Face Stylization. Abstract: While there have been recent ad

1.3k Dec 29, 2022
O-CNN: Octree-based Convolutional Neural Networks for 3D Shape Analysis

O-CNN This repository contains the implementation of our papers related with O-CNN. The code is released under the MIT license. O-CNN: Octree-based Co

Microsoft 607 Dec 28, 2022
PyTorch Language Model for 1-Billion Word (LM1B / GBW) Dataset

PyTorch Large-Scale Language Model A Large-Scale PyTorch Language Model trained on the 1-Billion Word (LM1B) / (GBW) dataset Latest Results 39.98 Perp

Ryan Spring 114 Nov 04, 2022
Ensemble Learning Priors Driven Deep Unfolding for Scalable Snapshot Compressive Imaging [PyTorch]

Ensemble Learning Priors Driven Deep Unfolding for Scalable Snapshot Compressive Imaging [PyTorch] Abstract Snapshot compressive imaging (SCI) can rec

integirty 6 Nov 01, 2022
Codes for CVPR2021 paper "PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization"

PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization (CVPR 2021) This is the official implementation of PW

Intelligent Robotics and Machine Vision Lab 42 Dec 18, 2022
FaRL for Facial Representation Learning

FaRL for Facial Representation Learning This repo hosts official implementation of our paper General Facial Representation Learning in a Visual-Lingui

Microsoft 19 Jan 05, 2022
Application of the L2HMC algorithm to simulations in lattice QCD.

l2hmc-qcd 📊 Slides Recent talk on Training Topological Samplers for Lattice Gauge Theory from the Machine Learning for High Energy Physics, on and of

Sam Foreman 37 Dec 14, 2022
ALIbaba's Collection of Encoder-decoders from MinD (Machine IntelligeNce of Damo) Lab

AliceMind AliceMind: ALIbaba's Collection of Encoder-decoders from MinD (Machine IntelligeNce of Damo) Lab This repository provides pre-trained encode

Alibaba 1.4k Jan 01, 2023
Improving Non-autoregressive Generation with Mixup Training

MIST Training MIST TRAIN_FILE=/your/path/to/train.json VALID_FILE=/your/path/to/valid.json OUTPUT_DIR=/your/path/to/save_checkpoints CACHE_DIR=/your/p

7 Nov 22, 2022
Employee-Managment - Company employee registration software in the face recognition system

Employee-Managment Company employee registration software in the face recognitio

Alireza Kiaeipour 7 Jul 10, 2022
Machine Learning From Scratch. Bare bones NumPy implementations of machine learning models and algorithms with a focus on accessibility. Aims to cover everything from linear regression to deep learning.

Machine Learning From Scratch About Python implementations of some of the fundamental Machine Learning models and algorithms from scratch. The purpose

Erik Linder-Norén 21.8k Jan 09, 2023
Code and data (Incidents Dataset) for ECCV 2020 Paper "Detecting natural disasters, damage, and incidents in the wild".

Incidents Dataset See the following pages for more details: Project page: IncidentsDataset.csail.mit.edu. ECCV 2020 Paper "Detecting natural disasters

Ethan Weber 67 Dec 27, 2022
Code for WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models.

WECHSEL Code for WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models. arXiv: https://arx

Institute of Computational Perception 45 Dec 29, 2022
Fast, flexible and fun neural networks.

Brainstorm Discontinuation Notice Brainstorm is no longer being maintained, so we recommend using one of the many other,available frameworks, such as

IDSIA 1.3k Nov 21, 2022
Jiminy Cricket Environment (NeurIPS 2021)

Jiminy Cricket This is the repository for "What Would Jiminy Cricket Do? Towards Agents That Behave Morally" by Dan Hendrycks*, Mantas Mazeika*, Andy

Dan Hendrycks 15 Aug 29, 2022
A package to predict protein inter-residue geometries from sequence data

trRosetta This package is a part of trRosetta protein structure prediction protocol developed in: Improved protein structure prediction using predicte

Ivan Anishchenko 185 Jan 07, 2023
PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)

About PyTorch 1.2.0 Now the master branch supports PyTorch 1.2.0 by default. Due to the serious version problem (especially torch.utils.data.dataloade

Sanghyun Son 2.1k Dec 27, 2022
Source code to accompany Defunctland's video "FASTPASS: A Complicated Legacy"

Shapeland Simulator Source code to accompany Defunctland's video "FASTPASS: A Complicated Legacy" Download the video at https://www.youtube.com/watch?

TouringPlans.com 70 Dec 14, 2022