dahyun-kang / renet

[ICCV'21] Official PyTorch implementation of Relational Embedding for Few-Shot Classification

Relational Embedding for Few-Shot Classification
(ICCV 2021)

Dahyun Kang, Heeseung Kwon, Juhong Min, Minsu Cho

[paper], [project hompage]

✔️ Requirements

• Ubuntu 16.04
• Python 3.7
• CUDA 11.0
• PyTorch 1.7.1

⚙️ Conda environmnet installation

conda env create --name renet_iccv21 --file environment.yml
conda activate renet_iccv21

📚 Datasets

cd datasets
bash download_miniimagenet.sh
bash download_cub.sh
bash download_cifar_fs.sh
bash download_tieredimagenet.sh

🌳 Authors' checkpoints

cd checkpoints
bash download_checkpoints_renet.sh

The file structure should be as follows:

renet/
├── datasets/
├── model/
├── scripts/
├── checkpoints/
│   ├── cifar_fs/
│   ├── cub/
│   ├── miniimagenet/
│   └── tieredimagenet/
train.py
test.py
README.md
environment.yml

📌 Quick start: testing scripts

To test in the 5-way K-shot setting:

bash scripts/test/{dataset_name}_5wKs.sh

For example, to test ReNet on the miniImagenet dataset in the 5-way 1-shot setting:

bash scripts/test/miniimagenet_5w1s.sh

🔥 Training scripts

To train in the 5-way K-shot setting:

bash scripts/train/{dataset_name}_5wKs.sh

For example, to train ReNet on the CUB dataset in the 5-way 1-shot setting:

bash scripts/train/cub_5w1s.sh

Training & testing a 5-way 1-shot model on the CUB dataset using a TitanRTX 3090 GPU takes 41m 30s.

🎨 Few-shot classification results

Experimental results on few-shot classification datasets with ResNet-12 backbone. We report average results with 2,000 randomly sampled episodes.

datasets	miniImageNet		tieredImageNet
setups	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
accuracy	67.60	82.58	71.61	85.28

datasets	CUB-200-2011		CIFAR-FS
setups	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
accuracy	79.49	91.11	74.51	86.60

🔍 Related repos

Our project references the codes in the following repos:

• Zhang et al., DeepEMD.
• Ye et al., FEAT
• Wang et al., Non-local neural networks
• Ramachandran et al., Stand-alone self-attention
• Huang et al., DCCNet
• Yang et al., VCN

💌 Acknowledgement

We adopted the main code bases from DeepEMD, and we really appreciate it 😃. We also sincerely thank all the ICCV reviewers, especially R#2, for valuable suggestions.

📜 Citing RENet

If you find our code or paper useful to your research work, please consider citing our work using the following bibtex:

@inproceedings{kang2021renet,
    author   = {Kang, Dahyun and Kwon, Heeseung and Min, Juhong and Cho, Minsu},
    title    = {Relational Embedding for Few-Shot Classification},
    booktitle= {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year     = {2021}
}

renet's Issues

Testing Output Network

Hi,
I'd like to give an image input from testset, in order to visualize the result (the corrisponding label).
How can I do this? Is there a command?
Thanks

About the Eq.(4) in the paper

Hi, I'm a beginner.
But I was confused, the implementation of eq.(4) in the code seems to take F as input, not Z (in the paper).

About GAP baseline

Paper results

Very interesting work!

Hi, I am very interested in your work. I would like to ask how to replicate the results in table 3 in the main paper? More specifically, I want to study the effect of the two modules by switching on/off them as in table 3

.

Inductive or transductive？

In the n-way k-shot( k>1) setting, the attended features of the support set are computed by summing k attended features, which are influenced by the query set. I wonder whether it is an inductive or transductive setting in few-shot learning.

A question

Is there an easy way to save checkpoints for Colab user?

The question is very simple: how can I save the training checkpoints in the guide and then resume them?

Conv3d or Conv4d in SCR

Hi, thanks for releasing the organized code.

I find that in the code SCR is implemented with Conv3d, while in the paper it is Conv4d. Does this matter?

the miniimagenet data set you shared can't be accessed

Hello, the miniimagenet data set you shared can't be accessed. Could you share it again? Thank you very much。

Could you please provide the code of RENet applied to Grad-cam?

Hi~
The support set and query set visualized by Grad-cam is really beautiful in this article.
Could you please provide the code of RENet applied to Grad-cam?
I'd like to try that!

GAP baseline

could you please show me the code about how to replace the attention process #8
i tried many times but i got failed