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 PyTorch implementation for "Low Precision Decentralized Distributed Training with Heterogenous Data"

Low Precision Decentralized Training with Heterogenous Data Official PyTorch implementation for "Low Precision Decentralized Distributed Training with

Aparna Aketi 0 Nov 23, 2021
This is the repository for Learning to Generate Piano Music With Sustain Pedals

SusPedal-Gen This is the official repository of Learning to Generate Piano Music With Sustain Pedals Demo Page Dataset The dataset used in this projec

Joann Ching 12 Sep 02, 2022
Generalizing Gaze Estimation with Outlier-guided Collaborative Adaptation

Generalizing Gaze Estimation with Outlier-guided Collaborative Adaptation Our paper is accepted by ICCV2021. Picture: Overview of the proposed Plug-an

Yunfei Liu 32 Dec 10, 2022
ICS 4u HD project, start before-wards. A curtain shooting game using python.

Touhou-Star-Salvation HDCH ICS 4u HD project, start before-wards. A curtain shooting game using python and pygame. By Jason Li For arts and gameplay,

15 Dec 22, 2022
Spam your friends and famly and when you do your famly will disown you and you will have no friends.

SpamBot9000 Spam your friends and family and when you do your family will disown you and you will have no friends. Terms of Use Disclaimer: Please onl

DJ15 0 Jun 09, 2022
DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction (3DV 2021)

DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction (3DV 2021) This repo is the implementation of DPC. Tested environment Pyth

Dvir Ginzburg 30 Nov 30, 2022
Code for the paper "Zero-shot Natural Language Video Localization" (ICCV2021, Oral).

Zero-shot Natural Language Video Localization (ZSNLVL) by Pseudo-Supervised Video Localization (PSVL) This repository is for Zero-shot Natural Languag

Computer Vision Lab. @ GIST 37 Dec 27, 2022
Official implementation for “Unsupervised Low-Light Image Enhancement via Histogram Equalization Prior”

HEP Unsupervised Low-Light Image Enhancement via Histogram Equalization Prior Implementation Python3 PyTorch=1.0 NVIDIA GPU+CUDA Training process The

FengZhang 34 Dec 04, 2022
Official code for "Mean Shift for Self-Supervised Learning"

MSF Official code for "Mean Shift for Self-Supervised Learning" Requirements Python = 3.7.6 PyTorch = 1.4 torchvision = 0.5.0 faiss-gpu = 1.6.1 In

UMBC Vision 44 Nov 21, 2022
[ICCV 2021] Official Tensorflow Implementation for "Single Image Defocus Deblurring Using Kernel-Sharing Parallel Atrous Convolutions"

KPAC: Kernel-Sharing Parallel Atrous Convolutional block This repository contains the official Tensorflow implementation of the following paper: Singl

Hyeongseok Son 50 Dec 29, 2022
Pre-Training 3D Point Cloud Transformers with Masked Point Modeling

Point-BERT: Pre-Training 3D Point Cloud Transformers with Masked Point Modeling Created by Xumin Yu*, Lulu Tang*, Yongming Rao*, Tiejun Huang, Jie Zho

Lulu Tang 306 Jan 06, 2023
An inofficial PyTorch implementation of PREDATOR based on KPConv.

PREDATOR: Registration of 3D Point Clouds with Low Overlap An inofficial PyTorch implementation of PREDATOR based on KPConv. The code has been tested

ZhuLifa 14 Aug 03, 2022
MobileNetV1-V2,MobileNeXt,GhostNet,AdderNet,ShuffleNetV1-V2,Mobile+ViT etc.

MobileNetV1-V2,MobileNeXt,GhostNet,AdderNet,ShuffleNetV1-V2,Mobile+ViT etc. ⭐⭐⭐⭐⭐

568 Jan 04, 2023
This is the official implementation of VaxNeRF (Voxel-Accelearated NeRF).

VaxNeRF Paper | Google Colab This is the official implementation of VaxNeRF (Voxel-Accelearated NeRF). This codebase is implemented using JAX, buildin

naruya 132 Nov 21, 2022
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Apache MXNet (incubating) for Deep Learning Master Docs License Apache MXNet (incubating) is a deep learning framework designed for both efficiency an

ROCm Software Platform 29 Nov 16, 2022
Revitalizing CNN Attention via Transformers in Self-Supervised Visual Representation Learning

Revitalizing CNN Attention via Transformers in Self-Supervised Visual Representation Learning This repository is the official implementation of CARE.

ChongjianGE 89 Dec 02, 2022
The official implementation of Variable-Length Piano Infilling (VLI).

Variable-Length-Piano-Infilling The official implementation of Variable-Length Piano Infilling (VLI). (paper: Variable-Length Music Score Infilling vi

29 Sep 01, 2022
Project for music generation system based on object tracking and CGAN

Project for music generation system based on object tracking and CGAN The project was inspired by MIDINet: A Convolutional Generative Adversarial Netw

1 Nov 21, 2021
pix2pix in tensorflow.js

pix2pix in tensorflow.js This repo is moved to https://github.com/yining1023/pix2pix_tensorflowjs_lite See a live demo here: https://yining1023.github

Yining Shi 47 Oct 04, 2022
Set of models for classifcation of 3D volumes

Classification models 3D Zoo - Keras and TF.Keras This repository contains 3D variants of popular CNN models for classification like ResNets, DenseNet

69 Dec 28, 2022