Official PyTorch implementation of the paper Rethinking Generalization in Few-Shot Classification.
🎓 📄 Find our paper: [arXiv] | :bookmark: Reference: [BibTeX]
- September 15, 2022: FewTURE is accepted at NeurIPS 2022! 🔥
- August 30, 2022: Release of our code -- Try out FewTURE! ✨ 💻 ⬅️
- June 15, 2022: Release of our preprint on arXiv
FewTURE is a novel approach to tackle the challenge of supervision collapse introduced by single image-level labels in few-shot learning. Splitting the input samples into patches and encoding these via the help of Vision Transformers allows us to establish semantic correspondences between local regions across images and independent of their respective class. The most informative patch embeddings for the task at hand are then determined as a function of the support set via online optimization at inference time, additionally providing visual interpretability of ‘what matters most’ in the image. We build on recent advances in unsupervised training of networks via masked image modelling to overcome the lack of fine-grained labels and learn the more general statistical structure of the data while avoiding negative image-level annotation influence, aka supervision collapse.
FewTURE achieves strong performance on several few-shot classification benchmarks (Accuracy on unseen test set):
| Dataset | 5-Way 1-Shot | 5-Way 5-Shot |
|---|---|---|
| miniImageNet | 72.40 ± 0.78 | 86.38 ± 0.49 |
| tieredImageNet | 76.32 ± 0.87 | 89.96 ± 0.55 |
| CIFAR-FS | 77.76 ± 0.81 | 88.90 ± 0.59 |
| FC100 | 47.68 ± 0.78 | 63.81 ± 0.75 |
FewTURE learns 'what matters most' at inference time via online optimization:
For detailed instruction how to set up your environment, install required packages and download the datasets, please refer to the installation instructions.
For a glimpse at the documentation of all arguments available for training, please check for self-supervised training:
python train_selfsup_pretrain.py --help
and for meta fine-tuning:
python train_metatrain_FewTURE.py --help
To start the self-supervised pre-training procedure using a ViT-small architecture on one node with 4 GPUs using a total batch size of 512 for 1600 epochs on miniImageNet, run:
export MASTER_ADDR=127.0.0.1
export MASTER_PORT=12345
export WORLD_SIZE=4
echo "Using Master_Addr $MASTER_ADDR:$MASTER_PORT to synchronise, and a world size of $WORLD_SIZE."
echo "Start training..."
torchrun --nproc_per_node 4 train_selfsup_pretrain.py --use_fp16 True --arch vit_small --epochs 1600 --batch_size_per_gpu 128 --teacher_temp 0.07 --warmup_teacher_temp_epochs 30 --norm_last_layer false --dataset miniimagenet --image_size 224 --data_path <path-to-dataset> --saveckp_freq 50 --shared_head true --out_dim 8192 --local_crops_number 10 --global_crops_scale 0.25 1 --local_crops_scale 0.05 0.25 --pred_ratio 0 0.3 --pred_ratio_var 0 0.2
echo "Finished training!"
If you want to instead train a swin Transformer architecture, choose --arch swin_tiny.
Since swin models generally lead to higher memory consumption, consider using multiple nodes
for training (we use 2 in our paper with a batch size of 64 per GPU), or experiment with reducing the overall batch size. We also recommend reducing the number of epochs (e.g. 300 - 800).
Note: Make sure to provide the path to the folder where you stored the datasets via the --data_path <path-to-dataset> argument.
You can use the following links to download the checkpoints of our pretrained models.
Note that all provided models have been trained in a self-supervised manner using ONLY the training split of the denoted few-shot image classification datasets and NO additional data.
| Pretraining Dataset | Architecture | Epochs | Download | ||
|---|---|---|---|---|---|
| miniImageNet | vit-small | 1600 | checkpoint | args | logs |
| swin-tiny | 800 | checkpoint | args | logs | |
| tieredImageNet | vit-small | 800* | checkpoint | args | logs |
| swin-tiny | 800 | checkpoint | args | logs | |
| CIFAR-FS | vit-small | 1600 | checkpoint | args | logs |
| swin-tiny | 800 | checkpoint | args | logs | |
| FC100 | vit-small | 1600 | checkpoint | args | logs |
| swin-tiny | 800 | checkpoint | args | logs | |
*Note: Due to the comparably big size of the tieredImageNet training set, we ran pretraining for only 800 epochs for ViT to reduce the computational load.
To start the meta fine-tuning procedure using a previously pretrained ViT-small architecture using one GPU for 100 epochs on the miniImageNet training dataset using 15 steps to adapt the token importance weights via online-optimisation at inference time, run:
- For a 5-way 5-shot scenario:
python3 train_metatrain_FewTURE.py --num_epochs 100 --data_path <path-to-dataset> --arch vit_small --n_way 5 --k_shot 5 --optim_steps_online 15 --chkpt_epoch 1600 --mdl_checkpoint_path <path-to-checkpoint-of-pretrained-model>
- For a 5-way 1-shot scenario:
python3 train_metatrain_FewTURE.py --num_epochs 100 --data_path <path-to-dataset> --arch vit_small --n_way 5 --k_shot 1 --optim_steps_online 15 --chkpt_epoch 1600 --mdl_checkpoint_path <path-to-checkpoint-of-pretrained-model>
To instead meta fine-tune a pretrained hierarchical swin architecture, use the --arch swin_tiny.
Note: Replace the --data_path <path-to-dataset> and possibly --mdl_checkpoint_path <path-to-checkpoint-of-pretrained-model> with the corresponding paths where you stored the model on your machine!
To evaluate a meta-trained ViT-small architecture on the miniImageNet test dataset, run:
- For a 5-way 5-shot scenario:
python3 eval_FewTURE.py --data_path <path-to-dataset> --arch vit_small --n_way 5 --k_shot 5 --trained_model_type metaft --optim_steps_online 15 --mdl_checkpoint_path <path-to-checkpoint-of-metaft-model>
- For a 5-way 1-shot scenario:
python3 eval_FewTURE.py --data_path <path-to-dataset> --arch vit_small --n_way 5 --k_shot 1 --trained_model_type metaft --optim_steps_online 15 --mdl_checkpoint_path <path-to-checkpoint-of-metaft-model>
This repository is built using components of the iBOT and DINO repositories for pre-training, and the DeepEMD and CloserLookFewShot repositories for loading the few-shot datasets.
This repository is released under the Apache 2.0 license as found in the LICENSE file.
If you find this repository useful, please consider giving us a star ⭐ and cite our work:
@inproceedings{
hillerma2022fewture,
title={Rethinking Generalization in Few-Shot Classification},
author={Markus Hiller and Rongkai Ma and Mehrtash Harandi and Tom Drummond},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=p_g2nHlMus}
}
If you have any questions regarding our work, please feel free to reach out!