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
Mouse Brain in the Model Zoo

Deep Neural Mouse Brain Modeling This is the repository for the ongoing deep neural mouse modeling project, an attempt to characterize the representat

Colin Conwell 15 Aug 22, 2022
This project is a loose implementation of paper "Algorithmic Financial Trading with Deep Convolutional Neural Networks: Time Series to Image Conversion Approach"

Stock Market Buy/Sell/Hold prediction Using convolutional Neural Network This repo is an attempt to implement the research paper titled "Algorithmic F

Asutosh Nayak 136 Dec 28, 2022
A privacy-focused, intelligent security camera system.

Self-Hosted Home Security Camera System A privacy-focused, intelligent security camera system. Features: Multi-camera support w/ minimal configuration

Scott Barnes 175 Jan 01, 2023
Pytorch implementation of Make-A-Scene: Scene-Based Text-to-Image Generation with Human Priors

Make-A-Scene - PyTorch Pytorch implementation (inofficial) of Make-A-Scene: Scene-Based Text-to-Image Generation with Human Priors (https://arxiv.org/

Casual GAN Papers 259 Dec 28, 2022
PyTorch implementation of UNet++ (Nested U-Net).

PyTorch implementation of UNet++ (Nested U-Net) This repository contains code for a image segmentation model based on UNet++: A Nested U-Net Architect

4ui_iurz1 642 Jan 04, 2023
Code for paper PairRE: Knowledge Graph Embeddings via Paired Relation Vectors.

PairRE Code for paper PairRE: Knowledge Graph Embeddings via Paired Relation Vectors. This implementation of PairRE for Open Graph Benchmak datasets (

Alipay 65 Dec 19, 2022
A small library for doing fluid simulation with neural networks.

Neural Fluid Fields This is a small library for doing fluid simulation with neural fields. Check out our review paper, Neural Fields in Visual Computi

Towaki 23 Jun 23, 2022
Implementation of CVPR 2020 Dual Super-Resolution Learning for Semantic Segmentation

Dual super-resolution learning for semantic segmentation 2021-01-02 Subpixel Update Happy new year! The 2020-12-29 update of SISR with subpixel conv p

Sam 79 Nov 24, 2022
tf2onnx - Convert TensorFlow, Keras and Tflite models to ONNX.

tf2onnx converts TensorFlow (tf-1.x or tf-2.x), tf.keras and tflite models to ONNX via command line or python api.

Open Neural Network Exchange 1.8k Jan 08, 2023
Black-Box-Tuning - Black-Box Tuning for Language-Model-as-a-Service

Black-Box-Tuning Source code for paper "Black-Box Tuning for Language-Model-as-a-Service". Being busy recently, the code in this repo and this tutoria

Tianxiang Sun 149 Jan 04, 2023
D2LV: A Data-Driven and Local-Verification Approach for Image Copy Detection

Facebook AI Image Similarity Challenge: Matching Track —— Team: imgFp This is the source code of our 3rd place solution to matching track of Image Sim

16 Dec 25, 2022
ManiSkill-Learn is a framework for training agents on SAPIEN Open-Source Manipulation Skill Challenge (ManiSkill Challenge), a large-scale learning-from-demonstrations benchmark for object manipulation.

ManiSkill-Learn ManiSkill-Learn is a framework for training agents on SAPIEN Open-Source Manipulation Skill Challenge, a large-scale learning-from-dem

Hao Su's Lab, UCSD 48 Dec 30, 2022
Code for "NeRS: Neural Reflectance Surfaces for Sparse-View 3D Reconstruction in the Wild," in NeurIPS 2021

Code for Neural Reflectance Surfaces (NeRS) [arXiv] [Project Page] [Colab Demo] [Bibtex] This repo contains the code for NeRS: Neural Reflectance Surf

Jason Y. Zhang 234 Dec 30, 2022
Creating Multi Task Models With Keras

Creating Multi Task Models With Keras About The Project! I used the keras and Tensorflow Library, To build a Deep Learning Neural Network to Creating

Srajan Chourasia 4 Nov 28, 2022
Pixel-wise segmentation on VOC2012 dataset using pytorch.

PiWiSe Pixel-wise segmentation on the VOC2012 dataset using pytorch. FCN SegNet PSPNet UNet RefineNet For a more complete implementation of segmentati

Bodo Kaiser 378 Dec 30, 2022
Train DeepLab for Semantic Image Segmentation

Train DeepLab for Semantic Image Segmentation Martin Kersner, [email protected]

Martin Kersner 172 Dec 14, 2022
Yolact-keras实例分割模型在keras当中的实现

Yolact-keras实例分割模型在keras当中的实现 目录 性能情况 Performance 所需环境 Environment 文件下载 Download 训练步骤 How2train 预测步骤 How2predict 评估步骤 How2eval 参考资料 Reference 性能情况 训练数

Bubbliiiing 11 Dec 26, 2022
AI-generated-characters for Learning and Wellbeing

AI-generated-characters for Learning and Wellbeing Click here for the full project page. This repository contains the source code for the paper AI-gen

MIT Media Lab 214 Jan 01, 2023
C3DPO - Canonical 3D Pose Networks for Non-rigid Structure From Motion.

C3DPO: Canonical 3D Pose Networks for Non-Rigid Structure From Motion By: David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, Andrea Vedal

Meta Research 309 Dec 16, 2022
Implementation for Curriculum DeepSDF

Curriculum-DeepSDF This repository is an implementation for Curriculum DeepSDF. Full paper is available here. Preparation Please follow original setti

Haidong Zhu 69 Dec 29, 2022