Counterfactual Generation Framework for Few-Shot Learning

By Dang Zhuohang, Luo Minnan, Jia Chengyou, Yan Caixia, Chang Xiaojun and Zheng Qinghua.

This repository is an implementation of the paper "Counterfactual Generation Framework for Few-Shot Learning".

** Apologize for the dirty code. We are working on organizing the code.

Requirements

pip install -r requirements.txt

Prepare Datasets

---

To be updated.

Training

MiniImageNet:

python -u main.py --phase pretrain_encoder --gpu 1 --save-path "./experiments/" --train-shot 5 --val-shot 1 --train-query 15 --val-query 15 --head FuseCosNet --network ResNet12_inv --dataset miniImageNet --z_disentangle --zd_beta 6.0 --zd_beta_annealing --add_noise 0.2 --temperature 500 --feature_size 640 --generative_model vae --latent_size 64 --attSize 171

TieredImageNet:

python -u main.py --phase pretrain_encoder --gpu 1 --save-path "./experiments/" --train-shot 1 --val-shot 1 --train-query 15 --val-query 15 --head FuseCosNet --network ResNet12_inv --dataset tieredImageNet --z_disentangle --zd_beta 6.0 --zd_beta_annealing --add_noise 0.2 --temperature 500 --feature_size 640 --generative_model vae --attSize 641 --latent_size 64

CIFAR-FS:

python -u main.py --phase pretrain_encoder --gpu 0 --save-path "./experiments/" --train-shot 5 --val-shot 1 --train-query 15 --val-query 15 --head FuseCosNet --network ResNet12_inv --dataset CIFAR-FS --z_disentangle --zd_beta 6.0 --zd_beta_annealing --add_noise 0.2 --temperature 500 --feature_size 640 --generative_model vae --attSize 164 --latent_size 64

Main Results

Dataset	Backbone	5w1s	5w5s
MiniImageNet	ResNet12	80.12%	86.13%
TieredImageNet	ResNet12	77.60%	87.38%
CIFAR-FS	ResNet12	87.20%	89.99%

License

To be updated