This repository contains the TensorFlow and PyTorch implementation for the CVPR 2019 Paper "Meta-Transfer Learning for Few-Shot Learning" by Qianru Sun,* Yaoyao Liu,* Tat-Seng Chua, and Bernt Schiele (*=equal contribution).
If you have any questions on this repository or the related paper, feel free to create an issue or send me an email.
Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. The key idea is to leverage a large number of similar few-shot tasks in order to learn how to adapt a base-learner to a new task for which only a few labeled samples are available. As deep neural networks (DNNs) tend to overfit using a few samples only, meta-learning typically uses shallow neural networks (SNNs), thus limiting its effectiveness. In this paper we propose a novel few-shot learning method called meta-transfer learning (MTL) which learns to adapt a deep NN for few shot learning tasks. Specifically, meta refers to training multiple tasks, and transfer is achieved by learning scaling and shifting functions of DNN weights for each task. We conduct experiments using (5-class, 1-shot) and (5-class, 5-shot) recognition tasks on two challenging few-shot learning benchmarks: ππππImageNet and Fewshot-CIFAR100.
Figure: Meta-Transfer Learning. (a) Parameter-level fine-tuning (FT) is a conventional meta-training operation, e.g. in MAML. Its update works for all neuron parameters, π and π. (b) Our neuron-level scaling and shifting (SS) operations in meta-transfer learning. They reduce the number of learning parameters and avoid overfitting problems. In addition, they keep large-scale trained parameters (in yellow) frozen, preventing βcatastrophic forgettingβ.
Please see README.md files in the corresponding folders:
- TensorFlow: [Document]
- PyTorch: [Document]
Directly download processed images: [Download Page]
The ππππImageNet dataset was proposed by Vinyals et al. for few-shot learning evaluation. Its complexity is high due to the use of ImageNet images but requires fewer resources and infrastructure than running on the full ImageNet dataset. In total, there are 100 classes with 600 samples of 84Γ84 color images per class. These 100 classes are divided into 64, 16, and 20 classes respectively for sampling tasks for meta-training, meta-validation, and meta-test. To generate this dataset from ImageNet, you may use the repository ππππImageNet tools.
Fewshot-CIFAR100 (FC100) is based on the popular object classification dataset CIFAR100. The splits were proposed by TADAM. It offers a more challenging scenario with lower image resolution and more challenging meta-training/test splits that are separated according to object super-classes. It contains 100 object classes and each class has 600 samples of 32 Γ 32 color images. The 100 classes belong to 20 super-classes. Meta-training data are from 60 classes belonging to 12 super-classes. Meta-validation and meta-test sets contain 20 classes belonging to 4 super-classes, respectively.
The π‘πππππImageNet dataset is a larger subset of ILSVRC-12 with 608 classes (779,165 images) grouped into 34 higher-level nodes in the ImageNet human-curated hierarchy. To generate this dataset from ImageNet, you may use the repository π‘πππππImageNet dataset: π‘πππππImageNet tools.
| (%) | ππππ 1-shot | ππππ 5-shot | FC100 1-shot | FC100 5-shot |
|---|---|---|---|---|
MTL Paper |
60.2 Β± 1.8 |
74.3 Β± 0.9 |
43.6 Β± 1.8 |
55.4 Β± 0.9 |
TensorFlow |
60.8 Β± 1.8 |
74.3 Β± 0.9 |
44.3 Β± 1.8 |
56.8 Β± 1.0 |
- The performance for the PyTorch version is under checking.
Please cite our paper if it is helpful to your work:
@inproceedings{SunLCS2019MTL,
author = {Qianru Sun and
Yaoyao Liu and
Tat{-}Seng Chua and
Bernt Schiele},
title = {Meta-Transfer Learning for Few-Shot Learning},
booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR}
2019, Long Beach, CA, USA, June 16-20, 2019},
pages = {403--412},
publisher = {Computer Vision Foundation / {IEEE}},
year = {2019}
}Our implementations use the source code from the following repositories and users: