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
Repository relating to the CVPR21 paper TimeLens: Event-based Video Frame Interpolation

TimeLens: Event-based Video Frame Interpolation This repository is about the High Speed Event and RGB (HS-ERGB) dataset, used in the 2021 CVPR paper T

Robotics and Perception Group 544 Dec 19, 2022
The goal of the exercises below is to evaluate the candidate knowledge and problem solving expertise regarding the main development focuses for the iFood ML Platform team: MLOps and Feature Store development.

The goal of the exercises below is to evaluate the candidate knowledge and problem solving expertise regarding the main development focuses for the iFood ML Platform team: MLOps and Feature Store dev

George Rocha 0 Feb 03, 2022
Unofficial implementation of Proxy Anchor Loss for Deep Metric Learning

Proxy Anchor Loss for Deep Metric Learning Unofficial pytorch, tensorflow and mxnet implementations of Proxy Anchor Loss for Deep Metric Learning. Not

Geonmo Gu 3 Jun 09, 2021
PyTorch implementation of Decoupling Value and Policy for Generalization in Reinforcement Learning

PyTorch implementation of Decoupling Value and Policy for Generalization in Reinforcement Learning

48 Dec 08, 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
Code repository for the paper: Hierarchical Kinematic Probability Distributions for 3D Human Shape and Pose Estimation from Images in the Wild (ICCV 2021)

Hierarchical Kinematic Probability Distributions for 3D Human Shape and Pose Estimation from Images in the Wild Akash Sengupta, Ignas Budvytis, Robert

Akash Sengupta 149 Dec 14, 2022
This repository attempts to replicate the SqueezeNet architecture and implement the same on an image classification task.

SqueezeNet-Implementation This repository attempts to replicate the SqueezeNet architecture using TensorFlow discussed in the research paper: "Squeeze

Rohan Mathur 3 Dec 13, 2022
A curated list of awesome open source libraries to deploy, monitor, version and scale your machine learning

Awesome production machine learning This repository contains a curated list of awesome open source libraries that will help you deploy, monitor, versi

The Institute for Ethical Machine Learning 12.9k Jan 04, 2023
Real-Time High-Resolution Background Matting

Real-Time High-Resolution Background Matting Official repository for the paper Real-Time High-Resolution Background Matting. Our model requires captur

Peter Lin 6.1k Jan 03, 2023
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
BOVText: A Large-Scale, Multidimensional Multilingual Dataset for Video Text Spotting

BOVText: A Large-Scale, Bilingual Open World Dataset for Video Text Spotting Updated on December 10, 2021 (Release all dataset(2021 videos)) Updated o

weijiawu 47 Dec 26, 2022
Real-time analysis of intracranial neurophysiology recordings.

py_neuromodulation Click this button to run the "Tutorial ML with py_neuro" notebooks: The py_neuromodulation toolbox allows for real time capable pro

Interventional Cognitive Neuromodulation - Neumann Lab Berlin 15 Nov 03, 2022
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

ELECTRA Introduction ELECTRA is a method for self-supervised language representation learning. It can be used to pre-train transformer networks using

Google Research 2.1k Dec 28, 2022
chainladder - Property and Casualty Loss Reserving in Python

chainladder (python) chainladder - Property and Casualty Loss Reserving in Python This package gets inspiration from the popular R ChainLadder package

Casualty Actuarial Society 130 Dec 07, 2022
PyTorch code for training MM-DistillNet for multimodal knowledge distillation

There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge MM-DistillNet is a

51 Dec 20, 2022
Pytorch implementation of our method for regularizing nerual radiance fields for few-shot neural volume rendering.

InfoNeRF: Ray Entropy Minimization for Few-Shot Neural Volume Rendering Pytorch implementation of our method for regularizing nerual radiance fields f

106 Jan 06, 2023
The code for our CVPR paper PISE: Person Image Synthesis and Editing with Decoupled GAN, Project Page, supp.

PISE The code for our CVPR paper PISE: Person Image Synthesis and Editing with Decoupled GAN, Project Page, supp. Requirement conda create -n pise pyt

jinszhang 110 Nov 21, 2022
WSDM2022 Challenge - Large scale temporal graph link prediction

WSDM 2022 Large-scale Temporal Graph Link Prediction - Baseline and Initial Test Set WSDM Cup Website link Link to this challenge This branch offers A

Deep Graph Library 34 Dec 29, 2022
Official PyTorch implementation of "BlendGAN: Implicitly GAN Blending for Arbitrary Stylized Face Generation" (NeurIPS 2021)

BlendGAN: Implicitly GAN Blending for Arbitrary Stylized Face Generation Official PyTorch implementation of the NeurIPS 2021 paper Mingcong Liu, Qiang

onion 462 Dec 29, 2022
LAnguage Model Analysis

LAMA: LAnguage Model Analysis LAMA is a probe for analyzing the factual and commonsense knowledge contained in pretrained language models. The dataset

Meta Research 960 Jan 08, 2023