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
ColossalAI-Benchmark - Performance benchmarking with ColossalAI

Benchmark for Tuning Accuracy and Efficiency Overview The benchmark includes our

HPC-AI Tech 31 Oct 07, 2022
This project aims to segment 4 common retinal lesions from Fundus Images.

This project aims to segment 4 common retinal lesions from Fundus Images.

Husam Nujaim 1 Oct 10, 2021
Codes for NeurIPS 2021 paper "Adversarial Neuron Pruning Purifies Backdoored Deep Models"

Adversarial Neuron Pruning Purifies Backdoored Deep Models Code for NeurIPS 2021 "Adversarial Neuron Pruning Purifies Backdoored Deep Models" by Dongx

Dongxian Wu 31 Dec 11, 2022
TensorFlow implementation of PHM (Parameterization of Hypercomplex Multiplication)

Parameterization of Hypercomplex Multiplications (PHM) This repository contains the TensorFlow implementation of PHM (Parameterization of Hypercomplex

Aston Zhang 9 Oct 26, 2022
Semi-Supervised Learning with Ladder Networks in Keras. Get 98% test accuracy on MNIST with just 100 labeled examples !

Semi-Supervised Learning with Ladder Networks in Keras This is an implementation of Ladder Network in Keras. Ladder network is a model for semi-superv

Divam Gupta 101 Sep 07, 2022
[NeurIPS 2021] "Drawing Robust Scratch Tickets: Subnetworks with Inborn Robustness Are Found within Randomly Initialized Networks" by Yonggan Fu, Qixuan Yu, Yang Zhang, Shang Wu, Xu Ouyang, David Cox, Yingyan Lin

Drawing Robust Scratch Tickets: Subnetworks with Inborn Robustness Are Found within Randomly Initialized Networks Yonggan Fu, Qixuan Yu, Yang Zhang, S

12 Dec 11, 2022
Portfolio Optimization and Quantitative Strategic Asset Allocation in Python

Riskfolio-Lib Quantitative Strategic Asset Allocation, Easy for Everyone. Description Riskfolio-Lib is a library for making quantitative strategic ass

Riskfolio 1.7k Jan 07, 2023
Fuwa-http - The http client implementation for the fuwa eco-system

Fuwa HTTP The HTTP client implementation for the fuwa eco-system Example import

Fuwa 2 Feb 16, 2022
J.A.R.V.I.S is an AI virtual assistant made in python.

J.A.R.V.I.S is an AI virtual assistant made in python. Running JARVIS Without Python To run JARVIS without python: 1. Head over to our installation pa

somePythonProgrammer 16 Dec 29, 2022
Data-depth-inference - Data depth inference with python

Welcome! This readme will guide you through the use of the code in this reposito

Marco 3 Feb 08, 2022
DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generative Transformers

DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generative Transformers Authors: Jaemin Cho, Abhay Zala, and Mohit Bansal (

Jaemin Cho 98 Dec 15, 2022
Implement of "Training deep neural networks via direct loss minimization" in PyTorch for 0-1 loss

This is the implementation of "Training deep neural networks via direct loss minimization" published at ICML 2016 in PyTorch. The implementation targe

Cuong Nguyen 1 Jan 18, 2022
BASH - Biomechanical Animated Skinned Human

We developed a method animating a statistical 3D human model for biomechanical analysis to increase accessibility for non-experts, like patients, athletes, or designers.

Machine Learning and Data Analytics Lab FAU 66 Nov 19, 2022
library for nonlinear optimization, wrapping many algorithms for global and local, constrained or unconstrained, optimization

NLopt is a library for nonlinear local and global optimization, for functions with and without gradient information. It is designed as a simple, unifi

Steven G. Johnson 1.4k Dec 25, 2022
A Python type explainer!

typesplainer A Python typehint explainer! Available as a cli, as a website, as a vscode extension, as a vim extension Usage First, install the package

Typesplainer 79 Dec 01, 2022
Extreme Rotation Estimation using Dense Correlation Volumes

Extreme Rotation Estimation using Dense Correlation Volumes This repository contains a PyTorch implementation of the paper: Extreme Rotation Estimatio

Ruojin Cai 29 Nov 18, 2022
In-place Parallel Super Scalar Samplesort (IPS⁴o)

In-place Parallel Super Scalar Samplesort (IPS⁴o) This is the implementation of the algorithm IPS⁴o presented in the paper Engineering In-place (Share

82 Dec 22, 2022
DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.

DeepSpeed+Megatron trained the world's most powerful language model: MT-530B DeepSpeed is hiring, come join us! DeepSpeed is a deep learning optimizat

Microsoft 8.4k Dec 28, 2022
Deep ViT Features as Dense Visual Descriptors

dino-vit-features [paper] [project page] Official implementation of the paper "Deep ViT Features as Dense Visual Descriptors". We demonstrate the effe

Shir Amir 113 Dec 24, 2022
Official Pytorch and JAX implementation of "Efficient-VDVAE: Less is more"

The Official Pytorch and JAX implementation of "Efficient-VDVAE: Less is more" Arxiv preprint Louay Hazami   ·   Rayhane Mama   ·   Ragavan Thurairatn

Rayhane Mama 144 Dec 23, 2022