Visual Representation Learning with Self-Supervised Attention for Low-Label High-Data Regime

Created by Prarthana Bhattacharyya.

Disclaimer: This is not an official product and is meant to be a proof-of-concept and for academic/educational use only.

This repository contains the PyTorch implementation for the paper Visual Representation Learning with Self-Supervised Attention for Low-Label High-Data Regime, to be presented at ICASSP-2022.

Self-supervision has shown outstanding results for natural language processing, and more recently, for image recognition. Simultaneously, vision transformers and its variants have emerged as a promising and scalable alternative to convolutions on various computer vision tasks. In this paper, we are the first to question if self-supervised vision transformers (SSL-ViTs) can be adapted to two important computer vision tasks in the low-label, high-data regime: few-shot image classification and zero-shot image retrieval. The motivation is to reduce the number of manual annotations required to train a visual embedder, and to produce generalizable, semantically meaningful and robust embeddings.

Results

• SSL-ViT + few-shot image classification:

• Qualitative analysis for base-classes chosen by supervised CNN and SSL-ViT for few-shot distribution calibration:

• SSL-ViT + zero-shot image retrieval:

Pretraining Self-Supervised ViT

• Run DINO with ViT-small network on a single node with 4 GPUs for 100 epochs with the following command.

cd dino/

python -m torch.distributed.launch --nproc_per_node=4 main_dino.py --arch vit_small --data_path /path/to/imagenet/train --output_dir /path/to/saving_dir

• For mini-ImageNet pretraining, we use the classes listed in: ssl-vit-fewshot/data/ImageNetSSLTrainingSplit_mini.txt For tiered-ImageNet pretraining, we use the classes listed in: ssl-vit-fewshot/data/ImageNetSSLTrainingSplit_tiered.txt
• For CUB-200, Cars-196 and SOP, we use the pretrained model from:

import torch
vits16 = torch.hub.load('facebookresearch/dino:main', 'dino_vits16')

Visual Representation Learning with Self-Supervised ViT for Low-Label High-Data Regime

Dataset Preparation

Please follow the instruction in FRN for few-shot image classification and RevisitDML for zero-shot image retrieval to download the datasets and put the corresponding datasets in ssl-vit-fewshot/data and DIML/data folder.

Training and Evaluation for few-shot image classification

• The first step is to extract features for base and novel classes using the pretrained SSL-ViT.
• get_dino_miniimagenet_feats.ipynb extracts SSL-ViT features for the base and novel classes.
• Change the hyper-parameter data_path to use CUB or tiered-ImageNet.
• The SSL-ViT checkpoints for the various datasets are provided below (Note: this has only been trained without labels). We also provide the extracted features which need to be stored in ssl-vit-fewshot/dino_features_data/.

• For n-way-k-shot evaluation, we provide miniimagenet_evaluate_dinoDC.ipynb.

Training and Evaluation for zero-shot image retrieval

• To train the baseline CNN models, run the scripts in DIML/scripts/baselines. The checkpoints are saved in Training_Results folder. For example:

cd DIML/
CUDA_VISIBLE_DEVICES=0 ./script/baselines/cub_runs.sh

• To train the supervised ViT and self-supervised ViT:

cp -r ssl-vit-retrieval/architectures/* DIML/ssl-vit-retrieval/architectures/

CUDA_VISIBLE_DEVICES=0 ./script/baselines/cub_runs.sh --arch vits
CUDA_VISIBLE_DEVICES=0 ./script/baselines/cub_runs.sh --arch dino

• To test the models, first edit the checkpoint paths in test_diml.py, then run

CUDA_VISIBLE_DEVICES=0 ./scripts/diml/test_diml.sh cub200

Acknowledgement

The code is based on:

• DINO
• Few-Shot Distribution Calibration
• DIML