This repository contains the public release of the Python implementation of
Learning to Affiliate: Mutual Centralized Learning for Few-shot Classification (CVPR 2022)
If you use this code or find this work useful for your research, please cite:
@InProceedings{Liu_2022_CVPR,
author = {Liu, Yang and Zhang, Weifeng and Xiang, Chao and Zheng, Tu and Cai, Deng and He, Xiaofei},
title = {Learning To Affiliate: Mutual Centralized Learning for Few-Shot Classification},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {14411-14420}
}
The following packages are required to run the scripts:
Some comparing methods may require additional packages to run (e.g, OpenCV in DeepEMD and qpth, cvxpy in MetaOptNet).
The dataset should be placed in dir "./data/miniImagenet" (or "./data/tieredimagenet") with the following format:
MCL
├── data
│ ├── miniImagenet
│ │ ├── train
│ │ │ ├──n01532829
│ │ │ ├──────n0153282900000987.png
│ │ ├── val
│ │ │ ├──
│ │ │ ├──────
│ │ ├── test
│ │ │ ├──
│ │ │ ├──────
The general mini-/tieredimagenet dataset can be downloaded from DeepEMD. The fine-grained datasets (i.e., CUB, meta-iNat and tiered meta-iNat) can be downloaded from FRN.
The train/test configs, tensorboard log and saved checkpoints are saved in the following format:
Dataset_Method_NwayKshot_Backbone_Accuracy (e.g., miniImagenet_MEL_katz_N5K1_R12_67.509)
├── tensorboard_log_date
│ ├──events.out.tfevents
├── predictions.txt (evaluation acc)
├── config.yaml
├── ebest_Nway_Kshot.txt (validation best epoch .txt)
├── ebest_Nway_Kshot.pth (validation best epocg .pth)
Download the snapshot files from Google Drive and extract it into the snapshots/ folder.
For example, MCL-Katz 5-way 1-shot ResNet12 VanillaFCN GPU 0
python experiments/run_evaluator.py \
--cfg ./snapshots/ResNet-12/MEL_katz/VanillaFCN/miniImagenet_MEL_katz_N5K1_R12_67.509/MEL_katz_N5K1_R12.yaml \
-c ./snapshots/ResNet-12/MEL_katz/VanillaFCN/miniImagenet_MEL_katz_N5K1_R12_67.509/ebest_5way_1shot.pth \
--device 0
We provide three pretraining config files motivated by FRN, DN4 and Linear Classifier. For example, FRN pretrainer on miniimagenet ResNet12 is performed by:
python experiments/run_pre.py \
--cfg ./configs/miniImagenet/pretrainer/FRN_pre.yaml
--device 0
The tensorboard log and pretrained model is saved in snapshots/ResNet-12/pretrainer/.
For Conv-4 experiments, we directly train the model from scratch. Just select any of config files from snapshots folder to the configs directory, e.g.,
cp ./snapshots/Conv-4/MEL_katz/VanillaFCN/miniImagenet_MEL_katz_N5K1_Conv4_55.747/MEL_katz_N5K1_Conv4.yaml ./configs/miniImagenet/Conv-4/
sh ./fast_train_test.sh ./configs/miniImagenet/Conv-4/MEL_katz_N5K1_Conv4.yaml 0
For ResNet-12 experiments, we first select the config files by analogous:
cp ./snapshots/ResNet-12/**/xxx.yaml ./configs/miniImagenet/ResNet-12/
Then we manually create the target checkpoint folders and copy (or soft link) the pretrained-model (e.g., snapshots/ResNet-12/pretrainer/miniImagenet_FRN_pre/miniimagenet-e0_pre.pth) to it:
mkdir ./checkpoint/xxx/
cp ./snapshots/ResNet-12/pretrainer/miniImagenet_FRN_pre/miniimagenet-e0_pre.pth ./checkpoint/xxx/
sh ./fast_train_test.sh ./configs/miniImagenet/ResNet-12/xxx.yaml 0
where xxx is the prefix of .yaml file and 0 indicates the GPU device number.
Experimental results on few-shot learning datasets with ResNet-12/Conv-4 backbone. We report average results with 10,000 randomly sampled episodes for both 1-shot and 5-shot evaluation.
The centrality plugins experiments follow the pre-train + evaluation setting proposed in Baseline that directly evaluates the methods without meta-training. We simply run experiments/run_evaluator.py on the pre-trained models that gives the following results:
We reimplement many of state-of-the-art methods in our unified framework according to their original implementations, i.e., DN4, DeepEMD, FRN, DSN, MetaOptNet, R2D2, CTX, Baseline, Baseline++, NegativeMargin, ProtoNet, RelationNet, MatchingNet.
It is easy to implement your own method in our unified framework by adding the similar your-own-method.py in the directory ./modules/query/ like the others.
We have tried our best to upload the correct snapshots on the google drive. However, since the config structures had been changed a lot in development (e.g., we introduce train/val/test n_way k_shot options instead of a single n_way k_shot in case some methods using the larger shot training), some of the .yaml configs in uploaded snapshots may have conflicts during loading.
If you encounter any issues or have questions about using the code, feel free to contact me lyng_95@zju.edu.cn