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
a practicable framework used in Deep Learning. So far UDL only provide DCFNet implementation for the ICCV paper (Dynamic Cross Feature Fusion for Remote Sensing Pansharpening)

UDL UDL is a practicable framework used in Deep Learning (computer vision). Benchmark codes, results and models are available in UDL, please contact @

Xiao Wu 11 Sep 30, 2022
Omniverse sample scripts - A guide for developing with Python scripts on NVIDIA Ominverse

Omniverse sample scripts ここでは、NVIDIA Omniverse ( https://www.nvidia.com/ja-jp/om

ft-lab (Yutaka Yoshisaka) 37 Nov 17, 2022
Code for KDD'20 "Generative Pre-Training of Graph Neural Networks"

GPT-GNN: Generative Pre-Training of Graph Neural Networks GPT-GNN is a pre-training framework to initialize GNNs by generative pre-training. It can be

Ziniu Hu 346 Dec 19, 2022
Code for our paper "Multi-scale Guided Attention for Medical Image Segmentation"

Medical Image Segmentation with Guided Attention This repository contains the code of our paper: "'Multi-scale self-guided attention for medical image

Ashish Sinha 394 Dec 28, 2022
Official implementation of the ICML2021 paper "Elastic Graph Neural Networks"

ElasticGNN This repository includes the official implementation of ElasticGNN in the paper "Elastic Graph Neural Networks" [ICML 2021]. Xiaorui Liu, W

liuxiaorui 34 Dec 04, 2022
Medical Image Segmentation using Squeeze-and-Expansion Transformers

Medical Image Segmentation using Squeeze-and-Expansion Transformers Introduction This repository contains the code of the IJCAI'2021 paper 'Medical Im

askerlee 172 Dec 20, 2022
MILK: Machine Learning Toolkit

MILK: MACHINE LEARNING TOOLKIT Machine Learning in Python Milk is a machine learning toolkit in Python. Its focus is on supervised classification with

Luis Pedro Coelho 610 Dec 14, 2022
Google Landmark Recogntion and Retrieval 2021 Solutions

Google Landmark Recogntion and Retrieval 2021 Solutions In this repository you can find solution and code for Google Landmark Recognition 2021 and Goo

Vadim Timakin 5 Nov 25, 2022
An Open-Source Toolkit for Prompt-Learning.

An Open-Source Framework for Prompt-learning. Overview • Installation • How To Use • Docs • Paper • Citation • What's New? Nov 2021: Now we have relea

THUNLP 2.3k Jan 07, 2023
A Pytorch Implementation of Source Data-free Domain Adaptation for a Faster R-CNN

A Pytorch Implementation of Source Data-free Domain Adaptation for a Faster R-CNN Please follow Faster R-CNN and DAF to complete the environment confi

2 Jan 12, 2022
BTC-Generator - BTC Generator With Python

Что такое BTC-Generator? Это генератор чеков всеми любимого @BTC_BANKER_BOT Для

DoomGod 3 Aug 24, 2022
🛠 All-in-one web-based IDE specialized for machine learning and data science.

All-in-one web-based development environment for machine learning Getting Started • Features & Screenshots • Support • Report a Bug • FAQ • Known Issu

Machine Learning Tooling 2.9k Jan 09, 2023
[ICLR 2021] Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization

Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization Kaidi Cao, Yining Chen, Junwei Lu, Nikos Arechiga, Adrien Gaidon, Tengyu Ma

Kaidi Cao 29 Oct 20, 2022
TSIT: A Simple and Versatile Framework for Image-to-Image Translation

TSIT: A Simple and Versatile Framework for Image-to-Image Translation This repository provides the official PyTorch implementation for the following p

Liming Jiang 255 Nov 23, 2022
High-performance moving least squares material point method (MLS-MPM) solver.

High-Performance MLS-MPM Solver with Cutting and Coupling (CPIC) (MIT License) A Moving Least Squares Material Point Method with Displacement Disconti

Yuanming Hu 2.2k Dec 31, 2022
这是一个facenet-pytorch的库,可以用于训练自己的人脸识别模型。

Facenet:人脸识别模型在Pytorch当中的实现 目录 性能情况 Performance 所需环境 Environment 注意事项 Attention 文件下载 Download 预测步骤 How2predict 训练步骤 How2train 参考资料 Reference 性能情况 训练数据

Bubbliiiing 210 Jan 06, 2023
StyleGAN - Official TensorFlow Implementation

StyleGAN — Official TensorFlow Implementation Picture: These people are not real – they were produced by our generator that allows control over differ

NVIDIA Research Projects 13.1k Jan 09, 2023
Neural style transfer in PyTorch.

style-transfer-pytorch An implementation of neural style transfer (A Neural Algorithm of Artistic Style) in PyTorch, supporting CPUs and Nvidia GPUs.

Katherine Crowson 395 Jan 06, 2023
Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy" (ICLR 2022 Spotlight)

About Code release for Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy (ICLR 2022 Spotlight)

THUML @ Tsinghua University 221 Dec 31, 2022
A Multi-modal Model Chinese Spell Checker Released on ACL2021.

ReaLiSe ReaLiSe is a multi-modal Chinese spell checking model. This the office code for the paper Read, Listen, and See: Leveraging Multimodal Informa

DaDa 106 Dec 29, 2022