Skip to content

Latest commit

 

History

History

learning_by_ignoring

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 

Data Reweighting for Pre-training

Introduction

Pretraining/finetuning frameworks are getting popularized with the recent advancement in self-supervised learning. However, pretraining data are oftentimes from a different distribution than finetuning data, which could potentially cause negative transfer. To overcome this issue, Learning by Ignoring combines a data reweighting shceme with pretraining/finetuning frameworks to automatically suppress the weight of pretraining data samples that cause negative transfer while increase the weight of others. The similar idea was also proposed in Meta-Learning to Improve Pre-TrainingMeta-learning (NeurIPS 2021).

Dataset

OfficeHome dataset can be downloaded from here

Environment

Our code is developed/tested on:

  • Python 3.8.10
  • pytorch 1.10
  • torchvision 1.11

Scripts

Baseline:

python main.py --gpu=0 --source_domain=Cl --target_domain=Ar --lam=7e-3 --baseline

Learning by Ignoring:

python main.py --gpu=0 --source_domain=Cl --target_domain=Ar --lam=7e-3

Run all experiments:

bash run.sh

Results

We present the result of Learning by Ignoring on the OfficeHome datset.

Cl-Ar Ar-Pr Pr-Rw Rw-Cl
Baseline 65.63% 87.35% 77.88% 68.00%
LBI (Ours) 66.87% 88.88% 78.77% 70.17%