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
CT-Net: Channel Tensorization Network for Video Classification

[ICLR2021] CT-Net: Channel Tensorization Network for Video Classification @inproceedings{ li2021ctnet, title={{\{}CT{\}}-Net: Channel Tensorization Ne

33 Nov 15, 2022
Algorithm to texture 3D reconstructions from multi-view stereo images

MVS-Texturing Welcome to our project that textures 3D reconstructions from images. This project focuses on 3D reconstructions generated using structur

Nils Moehrle 766 Jan 04, 2023
Regularized Frank-Wolfe for Dense CRFs: Generalizing Mean Field and Beyond

CRF - Conditional Random Fields A library for dense conditional random fields (CRFs). This is the official accompanying code for the paper Regularized

Đ.Khuê Lê-Huu 21 Nov 26, 2022
Generic Foreground Segmentation in Images

Pixel Objectness The following repository contains pretrained model for pixel objectness. Please visit our project page for the paper and visual resul

Suyog Jain 157 Nov 21, 2022
Official implementation of NeurIPS 2021 paper "Contextual Similarity Aggregation with Self-attention for Visual Re-ranking"

CSA: Contextual Similarity Aggregation with Self-attention for Visual Re-ranking PyTorch training code for CSA (Contextual Similarity Aggregation). We

Hui Wu 19 Oct 21, 2022
Consumer Fairness in Recommender Systems: Contextualizing Definitions and Mitigations

Consumer Fairness in Recommender Systems: Contextualizing Definitions and Mitigations This is the repository for the paper Consumer Fairness in Recomm

7 Nov 30, 2022
⚡️Optimizing einsum functions in NumPy, Tensorflow, Dask, and more with contraction order optimization.

Optimized Einsum Optimized Einsum: A tensor contraction order optimizer Optimized einsum can significantly reduce the overall execution time of einsum

Daniel Smith 653 Dec 30, 2022
The official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness.

This repository is the official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness. Requirements pip install -r requi

Jie Ren 17 Dec 12, 2022
The official PyTorch implementation of the paper: *Xili Dai, Xiaojun Yuan, Haigang Gong, Yi Ma. "Fully Convolutional Line Parsing." *.

F-Clip — Fully Convolutional Line Parsing This repository contains the official PyTorch implementation of the paper: *Xili Dai, Xiaojun Yuan, Haigang

Xili Dai 115 Dec 28, 2022
Real-time multi-object tracker using YOLO v5 and deep sort

This repository contains a two-stage-tracker. The detections generated by YOLOv5, a family of object detection architectures and models pretrained on the COCO dataset, are passed to a Deep Sort algor

Mike 3.6k Jan 05, 2023
Repository for the Bias Benchmark for QA dataset.

BBQ Repository for the Bias Benchmark for QA dataset. Authors: Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Tho

ML² AT CILVR 18 Nov 18, 2022
Phylogeny Partners

Phylogeny-Partners Two states models Instalation You may need to install the cython, networkx, numpy, scipy package: pip install cython, networkx, num

1 Sep 19, 2022
Opinionated code formatter, just like Python's black code formatter but for Beancount

beancount-black Opinionated code formatter, just like Python's black code formatter but for Beancount Try it out online here Features MIT licensed - b

Launch Platform 16 Oct 11, 2022
a short visualisation script for pyvideo data

PyVideo Speakers A CLI that visualises repeat speakers from events listed in https://github.com/pyvideo/data Not terribly efficient, but you know. Ins

Katie McLaughlin 3 Nov 24, 2021
reimpliment of DFANet: Deep Feature Aggregation for Real-Time Semantic Segmentation

DFANet This repo is an unofficial pytorch implementation of DFANet:Deep Feature Aggregation for Real-Time Semantic Segmentation log 2019.4.16 After 48

shen hui xiang 248 Oct 21, 2022
Machine Learning toolbox for Humans

Reproducible Experiment Platform (REP) REP is ipython-based environment for conducting data-driven research in a consistent and reproducible way. Main

Yandex 662 Nov 20, 2022
Audio2Face - Audio To Face With Python

Audio2Face Discription We create a project that transforms audio to blendshape w

FACEGOOD 724 Dec 26, 2022
Deployment of PyTorch chatbot with Flask

Chatbot Deployment with Flask and JavaScript In this tutorial we deploy the chatbot I created in this tutorial with Flask and JavaScript. This gives 2

Patrick Loeber (Python Engineer) 107 Dec 29, 2022
Geneva is an artificial intelligence tool that defeats censorship by exploiting bugs in censors

Geneva is an artificial intelligence tool that defeats censorship by exploiting bugs in censors

Kevin Bock 1.5k Jan 06, 2023
Official implementation of Long-Short Transformer in PyTorch.

Long-Short Transformer (Transformer-LS) This repository hosts the code and models for the paper: Long-Short Transformer: Efficient Transformers for La

NVIDIA Corporation 198 Dec 29, 2022