The PyTorch implementation of "Self-paced Meta-learning" (SepMeta). The curriculum in SepMeta is effectively integrated as a regularization term into the objective so as to the meta-learner can measure the hardness of tasks adaptively (or dynamically), according to what the model has already learned (i.e., the computed task-level losses).

• Python 3.5
• PyTorch >= 1.2
• TorchVision >= 0.2
• tqdm

• Training set: 64 classes (600 images per class)
• Val set: 16 classes
• Test set: 20 classes

• Training set: 351 classes (600 images per class)
• Val set: 97 classes
• Test set: 160 classes

After downloading the dataset, please create a new folder named "images" under the folder "miniimagenet" or "tieredimagenet", and put all images in this folder. The provided data loader will read images from the "images" folder by default. Of course, it is also OK to change the read path. For example, for the miniimagenet dataset, please change the line 10 of "./dataloader/mini_imagenet.py" as the path of the downloaded images.

SepMeta is an end-to-end method, you can dirrectly perform the following script to train and test a specific meta-learner.

python train_fsl.py --model_class ProtoNet --backbone_class Res12 --dataset MiniImageNet --max_epoch 100 --episodes_per_epoch 300 --temperature 40 --shot 1 --eval_shot 1 --step_size 20 --lr 0.001 --percent 0.1 --inc 0.005 --gpu 6

You can using the following script to test your trained model using tasks sampled from test set.

python test_fsl.py --shot 5 --eval_shot 5 --num_test_episodes 3000 --test_model .your_trained_model_path --gpu 5

Our implementations use the source code from the following repository:

• Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions, CVPR2020

If you have any questions about this implementation, please do not hesitate to contact with me.