Home

Awesome

Bayesian_MQDA

《Shallow Bayesian Meta Learning for Real World Few-shot Recognition》

ICCV 2021 Version: https://openaccess.thecvf.com/content/ICCV2021/papers/Zhang_Shallow_Bayesian_Meta_Learning_for_Real-World_Few-Shot_Recognition_ICCV_2021_paper.pdf

0. Outline

Preparation

Experiments

References

1. Preparation

1.1 download datasets

e.g., as for mini-Imagenet, please download mini-Imagenet and put it at ./data/mini and run proc_image.py to preprocess generate train/val/test datasets. (This process method is based on maml).

For CUB_200_2011, please reference the instruction at ${Project_dir}/data/cub/source/readme.md

For Cars, please reference the instruction at ${Project_dir}/data/cars/source/readme.md

1.2 download models

name format: $net_domain-$net_arch.pkl. e.g. mini-conv4.pkl.

network backbones: conv4, resnet18, wrn_28_20

download link. passport: 5kxg

save models at $project_dir/encoder/model_parameters

1.3 extract features

python extract_features.py --encoder 'conv4/resnet18/wrn' --dataset_train 'mini/tiered/cifarfs' --dataset_inference 'mini/tiered/cifarfs/cub/cars'

2. Experiments

the following "log name" is created after the training process

2.1 Single-Domain Few Shot Learning

e.g.: 5Way-5Shot, using encoder conv4 on miniImagenet dataset, using MetaQDA_MAP version

python train.py --n_way 5 --k_spt 5 --net_arch conv4 --net_domain mini --strategy map
python test.py -l_n "log name"

2.2 Cross-Domain Few Shot Testing

e.g.: testing trained models "log name" on cub dataset

python test.py -l_n "log name" -x_d cub

2.3 Multi-Domain Few Shot Learning: Meta-dataset

  1. Installation Meta-Dataset Follow the the "User instructions" in the Meta-Dataset repository for "Installation" and "Downloading and converting datasets".

  2. Download features download all the extracted features from HERE and put it in the ${Project_dir}/data/meta-dataset/

python train_urt_mqda.py
python test_urt_mqda.py -l_n "log name"

2.4 Few-shot Class Incremental Learning

python train_mqda_incremental.py
python test_mqda_map_incremental.py -l_n "log name"

2.5 Compute Calibration Error

python compute_ece.py -l_n "log name"

3. References