Robust fine-tuning of zero-shot models

Related tags

Deep Learningwise-ft
Overview

Robust fine-tuning of zero-shot models

This repository contains code for the paper Robust fine-tuning of zero-shot models by Mitchell Wortsman*, Gabriel Ilharco*, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, Ludwig Schmidt.

Abstract

Large pre-trained models such as CLIP offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning approaches substantially improve accuracy in-distribution, they also reduce out-of-distribution robustness. We address this tension by introducing a simple and effective method for improving robustness: ensembling the weights of the zero-shot and fine-tuned models. Compared to standard fine-tuning, the resulting weight-space ensembles provide large accuracy improvements out-of-distribution, while matching or improving in-distribution accuracy. On ImageNet and five derived distribution shifts, weight-space ensembles improve out-of-distribution accuracy by 2 to 10 percentage points while increasing in-distribution accuracy by nearly 1 percentage point relative to standard fine-tuning. These improvements come at no additional computational cost during fine-tuning or inference.

Summary figure

figure1

Compared to standard fine-tuning, weight-space ensembles for fine-tuning (WiSE-FT) improve out-of-distribution (OOD) accuracy without decreasing in-distribution (ID) performance. Top left: Zero-shot CLIP models exhibit high effective robustness and moderate in-distribution accuracy, while standard fine-tuning (end-to-end or with a linear classifier) attains higher ID accuracy and less effective robustness. Top right: Our method linearly interpolates between the zero-shot and fine-tuned models with a mixing coefficient alpha in [0,1]. Bottom: On five distribution shifts derived from ImageNet (ImageNetV2, ImageNet-R, ImageNet Sketch, ObjectNet, and ImageNet-A), WiSE-FT improves average OOD accuracy by 8.7 percentage points (pp) when fine-tuning end-to-end (+2.1 pp when fine-tuning a linear classifier) while maintaining ID accuracy.

Code

Overview

WiSE-FT can be implemented in a few lines of code in addition to standard fine-tuning, as shown below. See src/wise_ft.py for more details.

# Load models
zeroshot = ImageClassifier.load(zeroshot_checkpoint)
finetuned = ImageClassifier.load(finetuned_checkpoint)
theta_0 = zeroshot.state_dict()
theta_1 = finetuned.state_dict()

# make sure checkpoints are compatible
assert set(theta_0.keys()) == set(theta_1.keys())

# interpolate between checkpoints with mixing coefficient alpha
theta = {
    key: (1-alpha) * theta_0[key] + alpha * theta_1[key]
    for key in theta_0.keys()
}

# update the model acccording to the new weights
finetuned.load_state_dict(theta)

# evaluate
evaluate(finetuned, args)

Install dependencies

conda env create
conda activate wiseft

Add directory to PYTHONPATH:

cd wise-ft
export PYTHONPATH="$PYTHONPATH:$PWD"

Download data

When necessary, please refer to datasets.md for instructions on how to download datasets.

Run WiSE-FT

Sample command when zeroshot and fine-tuned models are available:

python src/wise_ft.py   \
    --eval-datasets=ImageNet,ImageNetV2,ImageNetR,ImageNetA,ImageNetSketch  \
    --load=models/zeroshot.pt,models/finetuned.pt  \
    --results-db=results.jsonl  \
    --save=models/wiseft  \
    --data-location=~/data \
    --alpha 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Sample command for running WiSE-FT from scratch using ViT-B/32:

python src/wise_ft.py   \
    --train-dataset=ImageNet  \
    --epochs=10  \
    --lr=0.00003  \
    --batch-size=512  \
    --cache-dir=cache  \
    --model=ViT-B/32  \
    --eval-datasets=ImageNet,ImageNetV2,ImageNetR,ImageNetA,ImageNetSketch  \
    --template=openai_imagenet_template  \
    --results-db=results.jsonl  \
    --save=models/wiseft/ViTB32  \
    --data-location=~/data \
    --alpha 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Note: the flag --freeze-encoder controls whether only a linear classifier is fine-tuned, or if all weights are fine-tuned (end-to-end).

Plotting results

Sample command for generating a scatter plot:

python src/scatter_plot.py  \
    --eval-datasets=ImageNetV2,ImageNetR,ImageNetA,ImageNetSketch  \
    --results-db=results.jsonl  \
    --save plots

We show samples of expected behavior below when running the commands above using ViT-B/16 (models can be downloaded here):

ImageNet-Sketch         ImageNet-A

ImageNet-R         ImageNetV2

ObjectNet

Citing

If you found this repository useful, please consider citing:

@article{wortsman2021robust,
  title={Robust fine-tuning of zero-shot models},
  author={Wortsman, Mitchell and Ilharco, Gabriel and Kim, Jong Wook and Li, Mike and Kornblith, Simon and Roelofs, Rebecca and Gontijo-Lopes, Raphael and Hajishirzi, Hannaneh and Farhadi, Ali and Namkoong, Hongseok and Schmidt, Ludwig},
  journal={arXiv preprint arXiv:2109.01903},
  note={\url{https://arxiv.org/abs/2109.01903}},
  year={2021}
}
The 1st place solution of track2 (Vehicle Re-Identification) in the NVIDIA AI City Challenge at CVPR 2021 Workshop.

AICITY2021_Track2_DMT The 1st place solution of track2 (Vehicle Re-Identification) in the NVIDIA AI City Challenge at CVPR 2021 Workshop. Introduction

Hao Luo 91 Dec 21, 2022
This repo provides function call to track multi-objects in videos

Custom Object Tracking Introduction This repo provides function call to track multi-objects in videos with a given trained object detection model and

Jeff Lo 51 Nov 22, 2022
Simple PyTorch hierarchical models.

A python package adding basic hierarchal networks in pytorch for classification tasks. It implements a simple hierarchal network structure based on feed-backward outputs.

Rajiv Sarvepalli 5 Mar 06, 2022
This repository collects 100 papers related to negative sampling methods.

Negative-Sampling-Paper This repository collects 100 papers related to negative sampling methods, covering multiple research fields such as Recommenda

RUCAIBox 119 Dec 29, 2022
Repository for XLM-T, a framework for evaluating multilingual language models on Twitter data

This is the XLM-T repository, which includes data, code and pre-trained multilingual language models for Twitter. XLM-T - A Multilingual Language Mode

Cardiff NLP 112 Dec 27, 2022
Tensorflow implementation of MIRNet for Low-light image enhancement

MIRNet Tensorflow implementation of the MIRNet architecture as proposed by Learning Enriched Features for Real Image Restoration and Enhancement. Lanu

Soumik Rakshit 91 Jan 06, 2023
GluonMM is a library of transformer models for computer vision and multi-modality research

GluonMM is a library of transformer models for computer vision and multi-modality research. It contains reference implementations of widely adopted baseline models and also research work from Amazon

42 Dec 02, 2022
Official Pytorch Implementation of Adversarial Instance Augmentation for Building Change Detection in Remote Sensing Images.

IAug_CDNet Official Implementation of Adversarial Instance Augmentation for Building Change Detection in Remote Sensing Images. Overview We propose a

53 Dec 02, 2022
A PyTorch implementation: "LASAFT-Net-v2: Listen, Attend and Separate by Attentively aggregating Frequency Transformation"

LASAFT-Net-v2 Listen, Attend and Separate by Attentively aggregating Frequency Transformation Woosung Choi, Yeong-Seok Jeong, Jinsung Kim, Jaehwa Chun

Woosung Choi 29 Jun 04, 2022
[ICCV2021] 3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds

3DVG-Transformer This repository is for the ICCV 2021 paper "3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds" Our method "3DV

22 Dec 11, 2022
Main repository for the HackBio'2021 Virtual Internship Experience for #Team-Greider ❤️

Hello 🤟 #Team-Greider The team of 20 people for HackBio'2021 Virtual Bioinformatics Internship 💝 🖨️ 👨‍💻 HackBio: https://thehackbio.com 💬 Ask us

Siddhant Sharma 7 Oct 20, 2022
Implementation of Shape Generation and Completion Through Point-Voxel Diffusion

Shape Generation and Completion Through Point-Voxel Diffusion Project | Paper Implementation of Shape Generation and Completion Through Point-Voxel Di

Linqi Zhou 103 Dec 29, 2022
A knowledge base construction engine for richly formatted data

Fonduer is a Python package and framework for building knowledge base construction (KBC) applications from richly formatted data. Note that Fonduer is

HazyResearch 386 Dec 05, 2022
A simple, fast, and efficient object detector without FPN

You Only Look One-level Feature (YOLOF), CVPR2021 A simple, fast, and efficient object detector without FPN. This repo provides an implementation for

789 Jan 09, 2023
Deep Illuminator is a data augmentation tool designed for image relighting. It can be used to easily and efficiently generate a wide range of illumination variants of a single image.

Deep Illuminator Deep Illuminator is a data augmentation tool designed for image relighting. It can be used to easily and efficiently generate a wide

George Chogovadze 52 Nov 29, 2022
A deep learning tabular classification architecture inspired by TabTransformer with integrated gated multilayer perceptron.

The GatedTabTransformer. A deep learning tabular classification architecture inspired by TabTransformer with integrated gated multilayer perceptron. C

Radi Cho 60 Dec 15, 2022
Net2net - Network-to-Network Translation with Conditional Invertible Neural Networks

Net2Net Code accompanying the NeurIPS 2020 oral paper Network-to-Network Translation with Conditional Invertible Neural Networks Robin Rombach*, Patri

CompVis Heidelberg 206 Dec 20, 2022
code for EMNLP 2019 paper Text Summarization with Pretrained Encoders

PreSumm This code is for EMNLP 2019 paper Text Summarization with Pretrained Encoders Updates Jan 22 2020: Now you can Summarize Raw Text Input!. Swit

Yang Liu 1.2k Dec 28, 2022
CenterFace(size of 7.3MB) is a practical anchor-free face detection and alignment method for edge devices.

CenterFace Introduce CenterFace(size of 7.3MB) is a practical anchor-free face detection and alignment method for edge devices. Recent Update 2019.09.

StarClouds 1.2k Dec 21, 2022
Conflict-aware Inference of Python Compatible Runtime Environments with Domain Knowledge Graph, ICSE 2022

PyCRE Conflict-aware Inference of Python Compatible Runtime Environments with Domain Knowledge Graph, ICSE 2022 Dependencies This project is developed

<a href=[email protected]"> 7 May 06, 2022