Repository for the paper : Meta-FDMixup: Cross-Domain Few-Shot Learning Guided byLabeled Target Data

Fu Yuqian

Last update: Nov 18, 2022

Related tags

Deep Learning Meta-FDMixup

Overview

1 Meta-FDMIxup

Repository for the paper :

Meta-FDMixup: Cross-Domain Few-Shot Learning Guided byLabeled Target Data. (ACM MM 2021)

paper

News! the representation video loaded in 2021/10/06 in Bilibili

News! the representation video loaded in 2021/10/10 in Youtube

If you have any questions, feel free to contact me. My email is [email protected].

2 setup and datasets

2.1 setup

A anaconda envs is recommended:

conda create --name py36 python=3.6
conda activate py36
conda install pytorch torchvision -c pytorch
pip3 install scipy>=1.3.2
pip3 install tensorboardX>=1.4
pip3 install h5py>=2.9.0

Then, git clone our repo:

git clone https://github.com/lovelyqian/Meta-FDMixup
cd Meta-FDMixup

2.2 datasets

Totally five datasets inculding miniImagenet, CUB, Cars, Places, and Plantae are used.

Following FWT-repo to download and setup all datasets. (It can be done quickly)
Remember to modify your own dataset dir in the 'options.py'
Under our new setting, we randomly select $num_{target}$ labeled images from the target base set to form the auxiliary set. The splits we used are provided in 'Sources/'.

3 pretrained ckps

We provide several pretrained ckps.

You can download and put them in the 'output/pretrained_ckps/'

3.1 pretrained model trained on the miniImagenet

pretrained_model_399.tar

3.2 full model meta-trained on the target datasets

Since our method is target-set specific, we have to train a model for each target dataset.

Notably, as we stated in the paper, we use the last checkpoint for target dataset, while the best model on the validation set of miniImagenet is used for miniImagenet. Here, we provide the model of 'miniImagenet|CUB' as an example.

full_model_5shot_target_cub_best_eval.tar

4 usage

4.1 network pretraining

python3 network_train.py --stage pretrain  --name pretrain-model --train_aug

If you have downloaded our pretrained_model_399.tar, you can just skip this step.

4.2 pretrained model testing

# test source dataset (miniImagenet)
python network_test.py --ckp_path output/checkpoints/pretrain-model/399.tar --stage pretrain --dataset miniImagenet --n_shot 5 

# test target dataset e.g. cub
python network_test.py --ckp_path output/checkpoints/pretrain-model/399.tar --stage pretrain --dataset cub --n_shot 5

you can test our pretrained_model_399.tar in the same way:

# test source dataset (miniImagenet)
python network_test.py --ckp_path output/pretrained_ckps/pretrained_model_399.tar --stage pretrain --dataset miniImagenet --n_shot 5 


# test target dataset e.g. cub
python network_test.py --ckp_path output/pretrained_ckps/pretrained_model_399.tar --stage pretrain --dataset cub --n_shot 5

4.3 network meta-training

# traget set: CUB
python3 network_train.py --stage metatrain --name metatrain-model-5shot-cub --train_aug --warmup output/checkpoints/pretrain-model/399.tar --target_set cub --n_shot 5

# target set: Cars
python3 network_train.py --stage metatrain --name metatrain-model-5shot-cars --train_aug --warmup output/checkpoints/pretrain-model/399.tar --target_set cars --n_shot 5

# target set: Places
python3 network_train.py --stage metatrain --name metatrain-model-5shot-places --train_aug --warmup output/checkpoints/pretrain-model/399.tar --target_set places --n_shot 5

# target set: Plantae
python3 network_train.py --stage metatrain --name metatrain-model-5shot-plantae --train_aug --warmup output/checkpoints/pretrain-model/399.tar --target_set plantae --n_shot 5

Also, you can use our pretrained_model_399.tar for warmup:

# traget set: CUB
python3 network_train.py --stage metatrain --name metatrain-model-5shot-cub --train_aug --warmup output/pretrained_ckps/pretrained_model_399.tar --target_set cub --n_shot 5

4.4 network testing

To test our provided full models:

# test target dataset (CUB)
python network_test.py --ckp_path output/pretrained_ckps/full_model_5shot_target_cub_399.tar --stage metatrain --dataset cub --n_shot 5 

# test target dataset (Cars)
python network_test.py --ckp_path output/pretrained_ckps/full_model_5shot_target_cars_399.tar --stage metatrain --dataset cars --n_shot 5 

# test target dataset (Places)
python network_test.py --ckp_path output/pretrained_ckps/full_model_5shot_target_places_399.tar --stage metatrain --dataset places --n_shot 5 

# test target dataset (Plantae)
python network_test.py --ckp_path output/pretrained_ckps/full_model_5shot_target_places_399.tar --stage metatrain --dataset plantae --n_shot 5 


# test source dataset (miniImagenet|CUB)
python network_test.py --ckp_path output/pretrained_ckps/full_model_5shot_target_cub_best_eval.tar --stage metatrain --dataset miniImagenet --n_shot 5

To test your models, just modify the 'ckp-path'.

5 citing

If you find our paper or this code useful for your research, please cite us:

@article{fu2021meta,
  title={Meta-FDMixup: Cross-Domain Few-Shot Learning Guided by Labeled Target Data},
  author={Fu, Yuqian and Fu, Yanwei and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2107.11978},
  year={2021}
}

6 Note

Notably, our code is built upon the implementation of FWT-repo.

You might also like...

PyTorch code for the paper "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" (CVPR2021)

PyTorch code for the paper "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" (CVPR2021) This repo presents PyTorch implementation of M

79 Dec 19, 2022

Implementation of "Meta-rPPG: Remote Heart Rate Estimation Using a Transductive Meta-Learner"

Meta-rPPG: Remote Heart Rate Estimation Using a Transductive Meta-Learner This repository is the official implementation of Meta-rPPG: Remote Heart Ra

137 Dec 13, 2022

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation [arxiv] This is the official repository for CDTrans: Cross-domain Transformer for

238 Dec 22, 2022

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

[ICCV2021] TransReID: Transformer-based Object Re-Identification [pdf] The official repository for TransReID: Transformer-based Object Re-Identificati

569 Dec 30, 2022

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021)

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021, official Pytorch implementatio

247 Dec 25, 2022

Pytorch implementation of the paper "Optimization as a Model for Few-Shot Learning"

Optimization as a Model for Few-Shot Learning This repo provides a Pytorch implementation for the Optimization as a Model for Few-Shot Learning paper.

238 Jan 4, 2023

Implementation of the paper "Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning"

Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning This is the implementation of the paper "Self-Promoted Prototype Refinement

78 Dec 2, 2022

This repository contains the code for using the H3DS dataset introduced in H3D-Net: Few-Shot High-Fidelity 3D Head Reconstruction

H3DS Dataset This repository contains the code for using the H3DS dataset introduced in H3D-Net: Few-Shot High-Fidelity 3D Head Reconstruction Access

72 Dec 10, 2022

EMNLP 2021 Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections

Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections Ruiqi Zhong, Kristy Lee*, Zheng Zhang*, Dan Klein EMN

42 Nov 3, 2022

Comments

Partitioning of classes in the auxiliary dataset

Hello, thank you for the code. When reappearing, I found that the label_base_xxx_X. json file provided by sources in your Github seems to have one class less (for example, the base class of CUB divided in FWT is 100 classes. But the label_base_cub_5.json you provided seems to have only 99 classes (0-98)). Data partitioning in FWT label_base_cub_5. json in sources

opened by MAGIC-XINE 3
How to eval baseline?

Hello, I am a beginner in this field ！May I know how you tested the performance of baseline by modifying the code? It is the s-base mentioned in the article.(☆▽☆)

opened by SmartNight-cc 1
Test about the choice of features

This is excellent work, and thanks for providing the code. I have a question in this paper, that is, in the test stage, why do you choose domain-irrelevant features to attain accuracy instead of domain-specific features? I hope that through your explanation, my confusion can be resolved. Thank you！

opened by Henkwu 1
Evaluation on other datasets

Excellent work and thank you for your release! I have one question to ask you: since the setting of your task is cross-domain few-shot learning, have you evaluated your method on the other four datasets (CropDisease, EuroSAT, ISIC and ChestX, the benchmark is set by [1]), whose domain shifts are larger than the other datasets (Places, Cars, Plantae and CUB)?

[1] Guo Y, Codella N C, Karlinsky L, et al. A broader study of cross-domain few-shot learning. ECCV 2020.

opened by nuistji 1

Repository for the paper : Meta-FDMixup: Cross-Domain Few-Shot Learning Guided byLabeled Target Data

Related tags

Overview

1 Meta-FDMIxup

2 setup and datasets

2.1 setup

2.2 datasets

3 pretrained ckps

3.1 pretrained model trained on the miniImagenet

3.2 full model meta-trained on the target datasets

4 usage

4.1 network pretraining

4.2 pretrained model testing

4.3 network meta-training

4.4 network testing

5 citing

6 Note

You might also like...

PyTorch code for the paper "Curriculum Graph Co-Teaching for Multi-target Domain Adaptation" (CVPR2021)

Implementation of "Meta-rPPG: Remote Heart Rate Estimation Using a Transductive Meta-Learner"

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021)

Pytorch implementation of the paper "Optimization as a Model for Few-Shot Learning"

Implementation of the paper "Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning"

This repository contains the code for using the H3DS dataset introduced in H3D-Net: Few-Shot High-Fidelity 3D Head Reconstruction

EMNLP 2021 Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections

Comments

Partitioning of classes in the auxiliary dataset

How to eval baseline?

Test about the choice of features

Evaluation on other datasets

Owner

Fu Yuqian

Code and data of the ACL 2021 paper: Few-Shot Text Ranking with Meta Adapted Synthetic Weak Supervision

The implementation of PEMP in paper "Prior-Enhanced Few-Shot Segmentation with Meta-Prototypes"

Few-shot Relation Extraction via Bayesian Meta-learning on Relation Graphs

This is the code of NeurIPS'21 paper "Towards Enabling Meta-Learning from Target Models".

Code for T-Few from "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"

PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation (TPAMI).

Few-NERD: Not Only a Few-shot NER Dataset

This repository is the code of the paper "Sparse Spatial Transformers for Few-Shot Learning".

Implementation of Cross Transformer for spatially-aware few-shot transfer, in Pytorch

Pytorch Implementation for CVPR2018 Paper: Learning to Compare: Relation Network for Few-Shot Learning