Clustering for Transductive Few-shot Learning
This repository contains the code for our paper entitled:
Transductive Few-Shot Learning: Clustering is All You Need?
Imtiaz Masud Ziko, Malik Boudiaf, Jose Dolz, Eric Granger and Ismail Ben Ayed, ArXiv 2021
We adapt several clustering methods to transductive inference in few-shot learning tasks. The clustering part works on a feature extractor initially trained over the base-class data. Using standard training on the base classes, without resorting to complex meta-learning and episodic-training strategies, our regularized-clustering approaches outperform state-of-the-art few-shot methods by significant margins, across various models, settings and data sets. Surprisingly, we found that even standard clustering procedures (e.g., K-means), which correspond to particular, non-regularized cases of our general model, already achieve competitive performances in comparison to the state-of-the-art in transductive few-shot learning. These surprising results point to the limitations of the current few-shot benchmarks, and question the viability of a large body of convoluted few-shot learning techniques in the recent literature.
- Python 3.6+
- Pytorch 1.0+
You can download the dataset from here
You can download the Tiered-ImageNet from here. After downloading and unziping this dataset run the following script to generate split files.
python src/utils/tieredImagenet.py --data path-to-tiered --split split/tiered/
Download and unpack the CUB 200-2011 from here After downloading and unziping this dataset run the following script to generate split files.
python src/utils/cub.py --data path-to-cub --split split/cub/
You can download our pretrained network models on base classes by running:
cd ./src
python download_models.py
Alternatively to train the network on the base classes from scratch remove the "--evaluate " options in the following script. The scripts to test SLK-shot:
sh run_SLK.sh
We get the following results in different few-shot benchmarks:
With WRN network:
| Methods | 1-shot | 5-shot |
|---|---|---|
| ProtoNet (Snell et al., NeurIPS 2017) | 62.60 | 79.97 |
| CC+rot (Gidaris et al., ICCV 2019) | 62.93 | 79.87 |
| MatchingNet (Vinyals et al., NeurIPS 2016) | 64.03 | 76.32 |
| FEAT (Ye et al., CVPR 2020) | 65.10 | 81.11 |
| Transductive fine-tuning (Dhillon et al., ICLR 2020) | 65.73 | 78.40 |
| SimpleShot (Wang et al., ArXiv 2019) | 65.87 | 82.09 |
| SIB (Hu et al., ICLR 2020) | 70.0 | 79.2 |
| BD-CSPN (Liu et al., ECCV 2020) | 70.31 | 81.89 |
| LaplacianShot (Ziko et al., ICML 2020) | 73.44 | 83.93 |
| K-means | 73.80 | 84.62 |
| K-modes | 74.78 | 84.45 |
| SLK-Means | 74.75 | 84.61 |
| SLK-MS | 75.17 | 84.28 |
With WRN network:
| Methods | 1-shot | 5-shot |
|---|---|---|
| CC+rot (Gidaris et al., ICCV 2019) | 70.53 | 84.98 |
| FEAT (Ye et al., CVPR 2020) | 70.41 | 84.38 |
| Transductive fine-tuning (Dhillon et al., ICLR 2020) | 73.34 | 85.50 |
| SimpleShot (Wang et al., ArXiv 2019) | 70.90 | 85.76 |
| BD-CSPN (Liu et al., ECCV 2020) | 78.74 | 86.92 |
| LaplacianShot (Ziko et al., ICML 2020) | 78.80 | 87.72 |
| K-means | 79.78 | 87.23 |
| K-modes | 80.67 | 87.23 |
| SLK-Means | 80.55 | 87.57 |
| SLK-MS | 81.13 | 87.69 |
With ResNet-18 network
| Methods | 1-shot | 5-shot |
|---|---|---|
| MatchingNet (Vinyals et al., NeurIPS 2016) | 73.49 | 84.45 |
| MAML (Finn et al., ICML 2017) | 68.42 | 83.47 |
| ProtoNet (Snell et al., NeurIPS 2017) | 72.99 | 86.64 |
| RelationNet (Sung et al., CVPR 2018) | 68.58 | 84.05 |
| Chen (Chen et al., ICLR 2019) | 67.02 | 83.58 |
| SimpleShot (Wang et al., ArXiv 2019) | 70.28 | 86.37 |
| LaplacianShot (Ziko et al., ICML 2020) | 79.93 | 88.59 |
| K-means | 80.30 | 88.51 |
| K-modes | 81.73 | 88.58 |
| SLK-Means | 81.40 | 88.61 |
| SLK-MS | 81.88 | 88.55 |