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 PyTorch Implementation of Neural IMage Assessment

NIMA: Neural IMage Assessment This is a PyTorch implementation of the paper NIMA: Neural IMage Assessment (accepted at IEEE Transactions on Image Proc

yunxiaos 418 Dec 29, 2022
(CVPR 2022) A minimalistic mapless end-to-end stack for joint perception, prediction, planning and control for self driving.

LAV Learning from All Vehicles Dian Chen, Philipp Krähenbühl CVPR 2022 (also arXiV 2203.11934) This repo contains code for paper Learning from all veh

Dian Chen 300 Dec 15, 2022
Evaluation Pipeline for our ECCV2020: Journey Towards Tiny Perceptual Super-Resolution.

Journey Towards Tiny Perceptual Super-Resolution Test code for our ECCV2020 paper: https://arxiv.org/abs/2007.04356 Our x4 upscaling pre-trained model

Royson 6 Mar 30, 2022
Shuwa Gesture Toolkit is a framework that detects and classifies arbitrary gestures in short videos

Shuwa Gesture Toolkit is a framework that detects and classifies arbitrary gestures in short videos

Google 89 Dec 22, 2022
This is an official implementation for "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" on Object Detection and Instance Segmentation.

Swin Transformer for Object Detection This repo contains the supported code and configuration files to reproduce object detection results of Swin Tran

Swin Transformer 1.4k Dec 30, 2022
Implementation of the paper "Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning"

Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning This is the implementation of the paper "Self-Promoted Prototype Refinement

Kai Zhu 78 Dec 02, 2022
Proof of concept GnuCash Webinterface

Proof of Concept GnuCash Webinterface This may one day be a something truly great. Milestones [ ] Browse accounts and view transactions [ ] Record sim

Josh 14 Dec 28, 2022
The project is an official implementation of our CVPR2019 paper "Deep High-Resolution Representation Learning for Human Pose Estimation"

Deep High-Resolution Representation Learning for Human Pose Estimation (CVPR 2019) News [2020/07/05] A very nice blog from Towards Data Science introd

Leo Xiao 3.9k Jan 05, 2023
TorchXRayVision: A library of chest X-ray datasets and models.

torchxrayvision A library for chest X-ray datasets and models. Including pre-trained models. ( 🎬 promo video about the project) Motivation: While the

Machine Learning and Medicine Lab 575 Jan 08, 2023
Distributed Asynchronous Hyperparameter Optimization in Python

Hyperopt: Distributed Hyperparameter Optimization Hyperopt is a Python library for serial and parallel optimization over awkward search spaces, which

6.5k Jan 01, 2023
The code for paper "Contrastive Spatio-Temporal Pretext Learning for Self-supervised Video Representation" which is accepted by AAAI 2022

Contrastive Spatio Temporal Pretext Learning for Self-supervised Video Representation (AAAI 2022) The code for paper "Contrastive Spatio-Temporal Pret

8 Jun 30, 2022
Simple and Distributed Machine Learning

Synapse Machine Learning SynapseML (previously MMLSpark) is an open source library to simplify the creation of scalable machine learning pipelines. Sy

Microsoft 3.9k Dec 30, 2022
Ganilla - Official Pytorch implementation of GANILLA

GANILLA We provide PyTorch implementation for: GANILLA: Generative Adversarial Networks for Image to Illustration Translation. Paper Arxiv Updates (Fe

Samet Hi 462 Dec 05, 2022
Fast sparse deep learning on CPUs

SPARSEDNN **If you want to use this repo, please send me an email: [email pro

Ziheng Wang 44 Nov 30, 2022
Bringing sanity to world of messed-up data

Sanitize sanitize is a Python module for making sure various things (e.g. HTML) are safe to use. It was originally written by Mark Pilgrim and is dist

Alireza Savand 63 Oct 26, 2021
PyTorch implementation of paper “Unbiased Scene Graph Generation from Biased Training”

A new codebase for popular Scene Graph Generation methods (2020). Visualization & Scene Graph Extraction on custom images/datasets are provided. It's also a PyTorch implementation of paper “Unbiased

Kaihua Tang 824 Jan 03, 2023
Official Implementation of "Transformers Can Do Bayesian Inference"

Official Code for the Paper "Transformers Can Do Bayesian Inference" We train Transformers to do Bayesian Prediction on novel datasets for a large var

AutoML-Freiburg-Hannover 103 Dec 25, 2022
Official and maintained implementation of the paper "OSS-Net: Memory Efficient High Resolution Semantic Segmentation of 3D Medical Data" [BMVC 2021].

OSS-Net: Memory Efficient High Resolution Semantic Segmentation of 3D Medical Data Christoph Reich, Tim Prangemeier, Özdemir Cetin & Heinz Koeppl | Pr

Christoph Reich 23 Sep 21, 2022
🦕 NanoSaur is a little tracked robot ROS2 enabled, made for an NVIDIA Jetson Nano

🦕 nanosaur NanoSaur is a little tracked robot ROS2 enabled, made for an NVIDIA Jetson Nano Website: nanosaur.ai Do you need an help? Discord For tech

NanoSaur 162 Dec 09, 2022
The most simple and minimalistic navigation dashboard.

Navigation This project follows a goal to have simple and lightweight dashboard with different links. I use it to have my own self-hosted service dash

Yaroslav 23 Dec 23, 2022