A Closer Look at the Training Strategy for Modern Meta-Learning

Enviroment

• Python3
• Pytorch 0.4
• json

Regression

LOO training strategy

Bilevel programming

python ./bilevel_code/regression_loo.py --n_task [NUMBER OF TRAINING TASKS] --train_shot [NUMBER OF SHOTS] --test_shot [NUMBER OF SHOTS]

MAML

python ./protonet_maml_code/train.py --dataset "regression_loo" --model "ReMaml" --method "re_maml" --n_episodes [NUMBER OF TRAINING TASKS] --n_shot [NUMBER OF SHOTS]

S/Q training strategy

Bilevel Programming

python ./bilevel_code/regression_sq.py --n_task [NUMBER OF TRAINING TASKS] --train_shot [NUMBER OF SHOTS] --test_shot [NUMBER OF SHOTS] --query [NUMBER OF QUERIES]

MAML

python ./protonet_maml_code/train.py --dataset "regression" --model "ReMaml" --method "re_maml" --n_episodes [NUMBER OF TRAINING TASKS] --n_shot [NUMBER OF SHOTS] --n_query [NUMBER OF QUERIES]

Classification

mini-ImageNet

• 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.)

ProtoNet

python ./protonet_maml_code/train.py --dataset "miniImagenet" --model "Conv4" --method "protonet" --n_episodes [NUMBER OF TRAINING TASKS] --n_shot [NUMBER OF SHOTS] --n_query [NUMBER OF QUERIES] --gap True

MAML

python ./protonet_maml_code/train.py --dataset "miniImagenet" --model "Conv4" --method "maml" --n_episodes [NUMBER OF TRAINING TASKS] --n_shot [NUMBER OF SHOTS] --n_query [NUMBER OF QUERIES] --gap True

Results

python ./bilevel_code/bilevel_regression_visualization

python ./protonet_maml_code/maml_regression_visualization

References

This code is built on

• Bilevel programming https://github.com/cyvius96/prototypical-network-pytorch
• MAML and ProtoNet https://github.com/wyharveychen/CloserLookFewShot