This repo is for the paper 'Detach and Unite: A Simple Meta-Transfer for Few-Shot Learning'

The following packages are required to run the scripts:

• PyTorch 1.8
• Python 3.7

For Omniglot, the data split settings is proposed by [1,2]. For mini-ImageNet, the data split is proposed by [3].

[1] Vinyals O, Blundell C, Lillicrap T, et al. Matching networks for one shot learning[J]. Advances in neural information processing systems, 2016, 29.

[2] Lake B M, Salakhutdinov R, Tenenbaum J B. Human-level concept learning through probabilistic program induction[J]. Science, 2015, 350(6266): 1332-1338.

[3] Ravi S, Larochelle H. Optimization as a model for few-shot learning[C]//International conference on learning representations. 2017.

There are five parts in the code:

• backbones: It contains the backbone network (Conv4 and ResNet-12) for the experiments.
• configs: Training settings.
• datasets: Dataloader and Samplers of different datasets.
• models: Baseline network and model register.
• utils: Help tools.

• Pre-training

python met.py --config train.yaml --gpu 0 --seed 1 --mode train --mannr whole

• Train MET

python met.py --config train.yaml --gpu 0 --seed 1 --mode train --mannr episode

• Train A-MET

python amet.py --config train_amet.yaml --gpu 0 --seed 1 --mode train --mannr episode

Our project references the codes in the following repos.

• meta-baseline

• matching networks