We present a regularized self-labeling approach to improve the generalization and robustness properties of fine-tuning.

Related tags

Overview

This repository provides the implementation for the paper "Improved Regularization and Robustness for Fine-tuning in Neural Networks", which will be presented as a poster paper in NeurIPS'21.

In this work, we propose a regularized self-labeling approach that combines regularization and self-training methods for improving the generalization and robustness properties of fine-tuning. Our approach includes two components:

First, we encode layer-wise regularization to penalize the model weights at different layers of the neural net.
Second, we add self-labeling that relabels data points based on current neural net's belief and reweights data points whose confidence is low.

Requirements

To install requirements:

pip install -r requirements.txt

Data Preparation

We use seven image datasets in our paper. We list the link for downloading these datasets and describe how to prepare data to run our code below.

Aircrafts: download and extract into ./data/aircrafts
- remove the class 257.clutter out of the data directory
CUB-200-2011: download and extract into ./data/CUB_200_2011/
Caltech-256: download and extract into ./data/caltech256/
Stanford-Cars: download and extract into ./data/StanfordCars/
Stanford-Dogs: download and extract into ./data/StanfordDogs/
Flowers: download and extract into ./data/flowers/
MIT-Indoor: download and extract into ./data/Indoor/

Our code automatically handles the split of the datasets.

Usage

Our algorithm (RegSL) interpolates between layer-wise regularization and self-labeling. Run the following commands for conducting experiments in this paper.

Fine-tuning ResNet-101 on image classification tasks.

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_indoor.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.136809975858091 --reg_predictor 6.40780158171339 --scale_factor 2.52883770643206\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_aircrafts.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 1.18330556653284 --reg_predictor 5.27713618808711 --scale_factor 1.27679969876201\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_birds.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.204403908747731 --reg_predictor 23.7850606577679 --scale_factor 4.73803591794678\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_caltech.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.0867998872549272 --reg_predictor 9.4552942790218 --scale_factor 1.1785989596144\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_cars.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 1.3340347414257 --reg_predictor 8.26940794089601 --scale_factor 3.47676759842434\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_dogs.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.0561320847651626 --reg_predictor 4.46281825974388 --scale_factor 1.58722606909531\
    --device 1

python train_constraint.py --model ResNet101 \
    --config configs/config_constraint_flower.json \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.131991042311165 --reg_predictor 10.7674132173309 --scale_factor 4.98010215976503\
    --device 1

Fine-tuning ResNet-18 under label noise.

python train_label_noise.py --config configs/config_constraint_indoor.json --model ResNet18 \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 7.80246991703043 --reg_predictor 14.077402847906 \
    --noise_rate 0.2 --train_correct_label --reweight_epoch 5 --reweight_temp 2.0 --correct_epoch 10 --correct_thres 0.9 

python train_label_noise.py --config configs/config_constraint_indoor.json --model ResNet18 \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 8.47139398080791 --reg_predictor 19.0191127114923 \
    --noise_rate 0.4 --train_correct_label --reweight_epoch 5 --reweight_temp 2.0 --correct_epoch 10 --correct_thres 0.9 

python train_label_noise.py --config configs/config_constraint_indoor.json --model ResNet18 \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 10.7576018531961 --reg_predictor 19.8157649727473 \
    --noise_rate 0.6 --train_correct_label --reweight_epoch 5 --reweight_temp 2.0 --correct_epoch 10 --correct_thres 0.9 
    
python train_label_noise.py --config configs/config_constraint_indoor.json --model ResNet18 \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 9.2031662757248 --reg_predictor 6.41568500472423 \
    --noise_rate 0.8 --train_correct_label --reweight_epoch 5 --reweight_temp 1.5 --correct_epoch 10 --correct_thres 0.9

Fine-tuning Vision Transformer on noisy labels.

python train_label_noise.py --config configs/config_constraint_indoor.json \
    --model VisionTransformer --is_vit --img_size 224 --vit_type ViT-B_16 --vit_pretrained_dir pretrained/imagenet21k_ViT-B_16.npz \
    --reg_method none --reg_norm none \
    --lr 0.0001 --device 1 --noise_rate 0.4

python train_label_noise.py --config configs/config_constraint_indoor.json \
    --model VisionTransformer --is_vit --img_size 224 --vit_type ViT-B_16 --vit_pretrained_dir pretrained/imagenet21k_ViT-B_16.npz \
    --reg_method none --reg_norm none \
    --lr 0.0001 --device 1 --noise_rate 0.8

python train_label_noise.py --config configs/config_constraint_indoor.json \
    --model VisionTransformer --is_vit --img_size 224 --vit_type ViT-B_16 --vit_pretrained_dir pretrained/imagenet21k_ViT-B_16.npz \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.7488074175044196 --reg_predictor 9.842955837419588 \
    --train_correct_label --reweight_epoch 24 --correct_epoch 18\
    --lr 0.0001 --device 1 --noise_rate 0.4

python train_label_noise.py --config configs/config_constraint_indoor.json \
    --model VisionTransformer --is_vit --img_size 224 --vit_type ViT-B_16 --vit_pretrained_dir pretrained/imagenet21k_ViT-B_16.npz \
    --reg_method constraint --reg_norm frob \
    --reg_extractor 0.1568903647089986 --reg_predictor 1.407080880079702 \
    --train_correct_label --reweight_epoch 18 --correct_epoch 2\
    --lr 0.0001 --device 1 --noise_rate 0.8

Please follow the instructions in ViT-pytorch to download the pre-trained models.

Fine-tuning ResNet-18 on ChestX-ray14 data set.

Run experiments on ChestX-ray14 in reproduce-chexnet path:

cd reproduce-chexnet

python retrain.py --reg_method None --reg_norm None --device 0

python retrain.py --reg_method constraint --reg_norm frob \
    --reg_extractor 5.728564437344309 --reg_predictor 2.5669480884876905 --scale_factor 1.0340072757925474 \
    --device 0

Citation

If you find this repository useful, consider citing our work titled above.

Acknowledgment

Thanks to the authors of the following repositories for providing their implementation publicly available.

You might also like...

Cartoon-StyleGan2 🙃 : Fine-tuning StyleGAN2 for Cartoon Face Generation

Fine-tuning StyleGAN2 for Cartoon Face Generation

520 Jan 4, 2023

Official codebase for Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World

Legged Robots that Keep on Learning Official codebase for Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World, whic

70 Dec 7, 2022

Fine-tuning StyleGAN2 for Cartoon Face Generation

Cartoon-StyleGAN 🙃 : Fine-tuning StyleGAN2 for Cartoon Face Generation Abstract Recent studies have shown remarkable success in the unsupervised imag

520 Jan 4, 2023

Implementation of the paper "Fine-Tuning Transformers: Vocabulary Transfer"

Transformer-vocabulary-transfer Implementation of the paper "Fine-Tuning Transfo

13 Nov 30, 2022

The coda and data for "Measuring Fine-Grained Domain Relevance of Terms: A Hierarchical Core-Fringe Approach" (ACL '21)

We propose a hierarchical core-fringe learning framework to measure fine-grained domain relevance of terms – the degree that a term is relevant to a broad (e.g., computer science) or narrow (e.g., deep learning) domain.

14 Oct 21, 2022

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation Zhaoyun Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li

25 Dec 16, 2022

The source code for the Cutoff data augmentation approach proposed in this paper: "A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation".

Cutoff: A Simple Data Augmentation Approach for Natural Language This repository contains source code necessary to reproduce the results presented in

49 Dec 22, 2022

Patch Rotation: A Self-Supervised Auxiliary Task for Robustness and Accuracy of Supervised Models

Patch-Rotation(PatchRot) Patch Rotation: A Self-Supervised Auxiliary Task for Robustness and Accuracy of Supervised Models Submitted to Neurips2021 To

4 Jul 12, 2021

The offcial repository for 'CharacterBERT and Self-Teaching for Improving the Robustness of Dense Retrievers on Queries with Typos', SIGIR2022

CharacterBERT-DR The offcial repository for CharacterBERT and Self-Teaching for Improving the Robustness of Dense Retrievers on Queries with Typos, Sh

11 Nov 15, 2022

We present a regularized self-labeling approach to improve the generalization and robustness properties of fine-tuning.

Related tags

Overview

Overview

Requirements

Data Preparation

Usage

Citation

Acknowledgment

You might also like...

Cartoon-StyleGan2 🙃 : Fine-tuning StyleGAN2 for Cartoon Face Generation

Official codebase for Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World

Fine-tuning StyleGAN2 for Cartoon Face Generation

Implementation of the paper "Fine-Tuning Transformers: Vocabulary Transfer"

The coda and data for "Measuring Fine-Grained Domain Relevance of Terms: A Hierarchical Core-Fringe Approach" (ACL '21)

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

The source code for the Cutoff data augmentation approach proposed in this paper: "A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation".

Patch Rotation: A Self-Supervised Auxiliary Task for Robustness and Accuracy of Supervised Models

The offcial repository for 'CharacterBERT and Self-Teaching for Improving the Robustness of Dense Retrievers on Queries with Typos', SIGIR2022

Owner

NEU-StatsML-Research

SUPERVISED-CONTRASTIVE-LEARNING-FOR-PRE-TRAINED-LANGUAGE-MODEL-FINE-TUNING - The Facebook paper about fine tuning RoBERTa with contrastive loss

Official implementation of "Open-set Label Noise Can Improve Robustness Against Inherent Label Noise" (NeurIPS 2021)

[CVPR 2022] Pytorch implementation of "Templates for 3D Object Pose Estimation Revisited: Generalization to New objects and Robustness to Occlusions" paper

PyTorch implementation for Stochastic Fine-grained Labeling of Multi-state Sign Glosses for Continuous Sign Language Recognition.

Example Of Fine-Tuning BERT For Named-Entity Recognition Task And Preparing For Cloud Deployment Using Flask, React, And Docker

Graph Self-Supervised Learning for Optoelectronic Properties of Organic Semiconductors

Black-Box-Tuning - Black-Box Tuning for Language-Model-as-a-Service

Ensemble Knowledge Guided Sub-network Search and Fine-tuning for Filter Pruning

Code for T-Few from "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"

Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning.