This repo contains the reference source code for the paper A Closer Look at Few-shot Classification in International Conference on Learning Representations (ICLR 2019). In this project, we provide a integrated testbed for a detailed empirical study for few-shot classification.
If you find our code useful, please consider citing our work using the bibtex:
@inproceedings{
chen2019closerfewshot,
title={A Closer Look at Few-shot Classification},
author={Chen, Wei-Yu and Liu, Yen-Cheng and Kira, Zsolt and Wang, Yu-Chiang and Huang, Jia-Bin},
booktitle={International Conference on Learning Representations},
year={2019}
}
- Python3
- Pytorch before 0.4 (for newer vesion, please see issue #3 )
- json
- Change directory to
./filelists/CUB
- run
source ./download_CUB.sh
- Change directory to
./filelists/miniImagenet
- run
source ./download_miniImagenet.sh
(WARNING: This would download the 155G ImageNet dataset. You can comment out correponded line 5-6 in download_miniImagenet.sh
if you already have one.)
- Finish preparation for CUB and mini-ImageNet and you are done!
- Change directory to
./filelists/omniglot
- run
source ./download_omniglot.sh
- Finish preparation for omniglot first
- Change directory to
./filelists/emnist
- run
source ./download_emnist.sh
- Require three data split json file: 'base.json', 'val.json', 'novel.json' for each dataset
- The format should follow
{"label_names": ["class0","class1",...], "image_names": ["filepath1","filepath2",...],"image_labels":[l1,l2,l3,...]}
See test.json for reference - Put these file in the same folder and change data_dir['DATASETNAME'] in configs.py to the folder path
Run
python ./train.py --dataset [DATASETNAME] --model [BACKBONENAME] --method [METHODNAME] [--OPTIONARG]
For example, run python ./train.py --dataset miniImagenet --model Conv4 --method baseline --train_aug
Commands below follow this example, and please refer to io_utils.py for additional options.
Save the extracted feature before the classifaction layer to increase test speed. This is not applicable to MAML, but are required for other methods.
Run
python ./save_features.py --dataset miniImagenet --model Conv4 --method baseline --train_aug
Run
python ./test.py --dataset miniImagenet --model Conv4 --method baseline --train_aug
- The test results will be recorded in
./record/results.txt
- For all the pre-computed results, please see
./record/few_shot_exp_figures.xlsx
. This will be helpful for including your own results for a fair comparison.
Our testbed builds upon several existing publicly available code. Specifically, we have modified and integrated the following code into this project:
- Framework, Backbone, Method: Matching Network https://github.com/facebookresearch/low-shot-shrink-hallucinate
- Omniglot dataset, Method: Prototypical Network https://github.com/jakesnell/prototypical-networks
- Method: Relational Network https://github.com/floodsung/LearningToCompare_FSL
- Method: MAML
https://github.com/cbfinn/maml
https://github.com/dragen1860/MAML-Pytorch
https://github.com/katerakelly/pytorch-maml
-
Q1 Why some of my reproduced results for CUB dataset are around 4~5% with you reported result? (#31, #34, #42)
-
A1 Sorry about my reported the results on the paper may run in different epochs or episodes, please see each issue for details.
-
Q2 Why some of my reproduced results for mini-ImageNet dataset are around 1~2% different with your reported results? (#17, #40, #41 #43)
-
A2 Due to random initialization, each training process could lead to different accuracy. Also, each test time could lead to different accuracy.
-
Q3 How do you decided the mean and the standard variation for dataset normalization? (#18, #39)
-
A3 I use the mean and standard variation from ImageNet, but you can use the ones calculated from your own dataset.
-
Q4 Do you have the mini-ImageNet dataset available without downloading the whole ImageNet? (#45 #29)
-
A4 You can use the dataset here https://github.com/oscarknagg/few-shot, but you will need to modify filelists/miniImagenet/write_miniImagenet_filelist.py.