Robust Fine-tuning of Zero-shot Models via Variance Reduction

This repository contains code for the Variance Reduction Fine-tuning (VRF).

TLDR: We propose a sample-wise ensembling technique for fine-tuning zero-shot models that can simultaneously attain the best ID and OOD accuracy.

Install dependencies

conda env create --file environment.yml
conda activate vrf

Add directory to PYTHONPATH:

cd VRF.public
export PYTHONPATH="$PYTHONPATH:$PWD"

Prepare Datasets

Prepare ImageNet

• (Skip if already downloaded) Download the dataset from the official website and extract to ${ANY_PATH_YOU_PREFER}/imagenet/images.

• Create data directory and create symbolic links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir imagenet
ln -s ${ANY_PATH_YOU_PREFER}/imagenet/images imagenet

Prepare ImageNet-V2

• (Skip if already downloaded) Download the dataset:

  cd ${ANY_PATH_YOU_PREFER}
  wget https://s3-us-west-2.amazonaws.com/imagenetv2public/imagenetv2-matched-frequency.tar.gz
  tar -xvf imagenetv2-matched-frequency.tar.gz
  rm imagenetv2-matched-frequency.tar.gz

• Create data directory and create symbolic links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir ImageNetV2-matched-frequency/
ln -s ${ANY_PATH_YOU_PREFER}/ImageNetV2-matched-frequency/ ImageNetV2-matched-frequency/

Prepare ImageNet Sketch

• (Skip if already downloaded) Download the dataset from Google Drive

• Create data directory and create symbolict links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir sketch/
ln -s ${ANY_PATH_YOU_PREFER}/imagenet-sketch/images/ sketch/

Prepare ImageNet A

• (Skip if already downloaded):

wget https://people.eecs.berkeley.edu/~hendrycks/imagenet-a.tar
tar -xvzf imagenet-a.tar
rm imagenet-a.tar

• Create data directory and create symbolict links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir imagenet-a/
ln -s ${ANY_PATH_YOU_PREFER}/imagenet-adversarial/imagenet-a imagenet-a/

Prepare ImageNet R

• (Skip if already downloaded):

wget https://people.eecs.berkeley.edu/~hendrycks/imagenet-r.tar
tar -xvzf imagenet-r.tar
rm imagenet-r.tar

• Create data directory and create symbolict links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir imagenet-r 
ln -s ${ANY_PATH_YOU_PREFER}/imagenet-rendition/imagenet-r/ imagenet-r/

Prepare ObjectNet

• (Skip if already downloaded):

wget https://objectnet.dev/downloads/objectnet-1.0.zip
unzip objectnet-1.0.zip
rm objectnet-1.0.zip

• Create data directory and create symbolict links.

export DATA_LOCATION=~/data
cd $DATA_LOCATION
mkdir objectnet-1.0
ln -s ${ANY_PATH_YOU_PREFER}/data3/data2/objectnet-1.0/ objectnet-1.0

Download pre-trained models

• Download the available zero-shot and fine-tuned models (CLIP ViT-B/16) from WiSE-FT to the directory ./models/vit-16/. The directory structure should look like:

models/
   |–– vit-16/ 
      |–– zeroshot.pt 
      |–– checkpoint_10.pt

Run baselines

• Compare the performance using CLIP ViT-B/16 of (1) zero-shot, (2) fine-tuned (E2E-FT), (3) WSE (WiSE-FT), and (4) OSE.

python src/baselines.py --load=models/vit-16/zeroshot.pt,models/vit-16/checkpoint_10.pt

Run our VRF

• Step 1: Identification and Step 2: Distance Calculation

python src/knn_distance.py --load=models/vit-16/zeroshot.pt,models/vit-16/checkpoint_10.pt

• Step 3: Sample-Wise Ensembling

python src/main.py --load=models/vit-16/zeroshot.pt,models/vit-16/checkpoint_10.pt