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
YuNetのPythonでのONNX、TensorFlow-Lite推論サンプル

YuNet-ONNX-TFLite-Sample YuNetのPythonでのONNX、TensorFlow-Lite推論サンプルです。 TensorFlow-LiteモデルはPINTO0309/PINTO_model_zoo/144_YuNetのものを使用しています。 Requirement Op

KazuhitoTakahashi 8 Nov 17, 2021
PyTorch implementation of "A Full-Band and Sub-Band Fusion Model for Real-Time Single-Channel Speech Enhancement."

FullSubNet This Git repository for the official PyTorch implementation of "A Full-Band and Sub-Band Fusion Model for Real-Time Single-Channel Speech E

郝翔 357 Jan 04, 2023
Pixel-Perfect Structure-from-Motion with Featuremetric Refinement (ICCV 2021, Oral)

Pixel-Perfect Structure-from-Motion (ICCV 2021 Oral) We introduce a framework that improves the accuracy of Structure-from-Motion by refining keypoint

Computer Vision and Geometry Lab 831 Dec 29, 2022
Sentinel-1 vessel detection model used in the xView3 challenge

sar_vessel_detect Code for the AI2 Skylight team's submission in the xView3 competition (https://iuu.xview.us) for vessel detection in Sentinel-1 SAR

AI2 6 Sep 10, 2022
[ICLR 2021] Is Attention Better Than Matrix Decomposition?

Enjoy-Hamburger 🍔 Official implementation of Hamburger, Is Attention Better Than Matrix Decomposition? (ICLR 2021) Under construction. Introduction T

Gsunshine 271 Dec 29, 2022
Camview - A CLI-tool used to stream CCTV online footage based on URL params

CamView A CLI-tool used to stream CCTV online footage based on URL params Get St

Finn Lancaster 54 Dec 09, 2022
PEPit is a package enabling computer-assisted worst-case analyses of first-order optimization methods.

PEPit: Performance Estimation in Python This open source Python library provides a generic way to use PEP framework in Python. Performance estimation

Baptiste 53 Nov 16, 2022
Official implementation of "Learning Forward Dynamics Model and Informed Trajectory Sampler for Safe Quadruped Navigation" (RSS 2022)

Intro Official implementation of "Learning Forward Dynamics Model and Informed Trajectory Sampler for Safe Quadruped Navigation" Robotics:Science and

Yunho Kim 21 Dec 07, 2022
Official implementation of the network presented in the paper "M4Depth: A motion-based approach for monocular depth estimation on video sequences"

M4Depth This is the reference TensorFlow implementation for training and testing depth estimation models using the method described in M4Depth: A moti

Michaël Fonder 76 Jan 03, 2023
A Pytorch implementation of "LegoNet: Efficient Convolutional Neural Networks with Lego Filters" (ICML 2019).

LegoNet This code is the implementation of ICML2019 paper LegoNet: Efficient Convolutional Neural Networks with Lego Filters Run python train.py You c

YangZhaohui 140 Sep 26, 2022
Camera ready code repo for the NeuRIPS 2021 paper: "Impression learning: Online representation learning with synaptic plasticity".

Impression-Learning-Camera-Ready Camera ready code repo for the NeuRIPS 2021 paper: "Impression learning: Online representation learning with synaptic

2 Feb 09, 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
BEGAN in PyTorch

BEGAN in PyTorch This project is still in progress. If you are looking for the working code, use BEGAN-tensorflow. Requirements Python 2.7 Pillow tqdm

Taehoon Kim 260 Dec 07, 2022
Official implementation of VQ-Diffusion

Vector Quantized Diffusion Model for Text-to-Image Synthesis Overview This is the official repo for the paper: [Vector Quantized Diffusion Model for T

Microsoft 592 Jan 03, 2023
RealTime Emotion Recognizer for Machine Learning Study Jam's demo

Emotion recognizer Table of contents Clone project Dataset Install dependencies Main program Demo 1. Clone project git clone https://github.com/GDSC20

Google Developer Student Club - UIT 1 Oct 05, 2021
TensorFlow GNN is a library to build Graph Neural Networks on the TensorFlow platform.

TensorFlow GNN This is an early (alpha) release to get community feedback. It's under active development and we may break API compatibility in the fut

889 Dec 30, 2022
A new version of the CIDACS-RL linkage tool suitable to a cluster computing environment.

Fully Distributed CIDACS-RL The CIDACS-RL is a brazillian record linkage tool suitable to integrate large amount of data with high accuracy. However,

Robespierre Pita 5 Nov 04, 2022
Very large and sparse networks appear often in the wild and present unique algorithmic opportunities and challenges for the practitioner

Sparse network learning with snlpy Very large and sparse networks appear often in the wild and present unique algorithmic opportunities and challenges

Andrew Stolman 1 Apr 30, 2021
A simple consistency training framework for semi-supervised image semantic segmentation

PseudoSeg: Designing Pseudo Labels for Semantic Segmentation PseudoSeg is a simple consistency training framework for semi-supervised image semantic s

Google Interns 143 Dec 13, 2022
Official code of paper: MovingFashion: a Benchmark for the Video-to-Shop Challenge

SEAM Match-RCNN Official code of MovingFashion: a Benchmark for the Video-to-Shop Challenge paper Installation Requirements: Pytorch 1.5.1 or more rec

HumaticsLAB 31 Oct 10, 2022