This repository contains the code for the paper
"Simultaneous Perturbation Method for Multi-Task Weight Optimization in One-Shot Meta-Learning".
This code has been tested on Ubuntu 16.04 with Python 3.8 and PyTorch 1.8.
To install the required dependencies:
pip install -r requirements.txt
To reproduce the results on benchmarks described in our article, use the following scripts. To vary types of the experiments, change the parameters of the scripts responsible for benchmark dataset, shot and way (e.g. miniImageNet 1-shot 5-way or CIFAR-FS 5-shot 2-way).
Multi-task modification (MTM) for Model-Agnostic Meta-Learning (MAML) (Finn et al., 2017).
Multi-task modifications for MAML are trained on top of baseline MAML model which has to be trained beforehand.
To train MAML (reproduced) on miniImageNet 1-shot 2-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name reproduced-miniimagenet \
--dataset miniimagenet \
--num-ways 2 \
--num-shots 1 \
--num-steps 5 \
--num-epochs 300 \
--use-cuda \
--output-folder ./results
To train MAML MTM SPSA-Track on miniImageNet 1-shot 2-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name mini-imagenet-mtm-spsa-track \
--load "./results/reproduced-miniimagenet/model.th" \
--dataset miniimagenet \
--num-ways 2 \
--num-shots 1 \
--num-steps 5 \
--task-weighting spsa-track \
--normalize-spsa-weights-after 100 \
--num-epochs 40 \
--use-cuda \
--output-folder ./results
To train MAML (reproduced) on tieredImageNet 1-shot 2-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name reproduced-tieredimagenet \
--dataset tieredimagenet \
--num-ways 2 \
--num-shots 1 \
--num-steps 5 \
--num-epochs 300 \
--use-cuda \
--output-folder ./results
To train MAML MTM SPSA on tieredImageNet 1-shot 2-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name tiered-imagenet-mtm-spsa \
--load "./results/reproduced-tieredimagenet/model.th" \
--dataset tieredimagenet \
--num-ways 2 \
--num-shots 1 \
--num-steps 5 \
--task-weighting spsa-delta \
--normalize-spsa-weights-after 100 \
--num-epochs 40 \
--use-cuda \
--output-folder ./results
To train MAML (reproduced) on FC100 5-shot 5-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name reproduced-fc100 \
--dataset fc100 \
--num-ways 5 \
--num-shots 5 \
--num-steps 5 \
--num-epochs 300 \
--use-cuda \
--output-folder ./results
To train MAML MTM SPSA-Coarse on FC100 5-shot 5-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name fc100-mtm-spsa-coarse \
--load "./results/reproduced-fc100/model.th" \
--dataset fc100 \
--num-ways 5 \
--num-shots 5 \
--num-steps 5 \
--task-weighting spsa-per-coarse-class \
--num-epochs 40 \
--use-cuda \
--output-folder ./results
To train MAML (reproduced) on CIFAR-FS 1-shot 5-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name reproduced-cifar \
--dataset cifarfs \
--num-ways 5 \
--num-shots 1 \
--num-steps 5 \
--num-epochs 600 \
--use-cuda \
--output-folder ./results
To train MAML MTM Inner First-Order on CIFAR-FS 1-shot 5-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name cifar-mtm-inner-first-order \
--load "./results/reproduced-cifar/model.th" \
--dataset cifarfs \
--num-ways 5 \
--num-shots 1 \
--num-steps 5 \
--task-weighting gradient-novel-loss \
--use-inner-optimizer \
--num-epochs 40 \
--use-cuda \
--output-folder ./results
To train MAML MTM Backprop on CIFAR-FS 1-shot 5-way benchmark, run:
python maml/train.py ./datasets/ \
--run-name cifar-mtm-backprop \
--load "./results/reproduced-cifar-5shot-5way/model.th" \
--dataset cifarfs \
--num-ways 5 \
--num-shots 1 \
--num-steps 5 \
--task-weighting gradient-novel-loss \
--num-epochs 40 \
--use-cuda \
--output-folder ./results
To test any of the above-described benchmarks, run:
python maml/test.py ./results/path-to-config/config.json --num-steps 10 --use-cuda
For instance, to test MAML MTM SPSA-Track on miniImageNet 1-shot 2-way benchmark, run:
python maml/test.py ./results/mini-imagenet-mtm-spsa-track/config.json --num-steps 10 --use-cuda
Multi-task modification (MTM) for Prototypical Networks (ProtoNet) (Snell et al., 2017).
To train ProtoNet MTM (2 tasks) SPSA-Track with ResNet-12 backbone on miniImageNet 1-shot 5-way benchmark, run:
python protonet/train.py \
--dataset miniImageNet \
--network ResNet12 \
--tracking \
--task-number 2 \
--train-shot 1 \
--train-way 5 \
--val-shot 1 \
--val-way 5
To test ProtoNet MTM SPSA-Track with ResNet-12 backbone on miniImageNet 1-shot 5-way benchmark, run:
python protonet/test.py --dataset miniImageNet --network ResNet12 --shot 1 --way 5
To train ProtoNet MTM Backprop with 64-64-64-64 backbone on CIFAR-FS 1-shot 2-way benchmark, run:
python protonet/train.py \
--dataset CIFAR_FS \
--train-weights \
--train-weights-layer \
--train-shot 1 \
--train-way 2 \
--val-shot 1 \
--val-way 2
To test ProtoNet MTM Backprop with 64-64-64-64 backbone on CIFAR-FS 1-shot 5-way benchmark, run:
python protonet/test.py --dataset CIFAR_FS --shot 1 --way 2
To train ProtoNet MTM Inner First-Order with 64-64-64-64 backbone on FC100 10-shot 5-way benchmark, run:
python protonet/train.py \
--dataset FC100 \
--train-weights \
--train-weights-opt \
--train-shot 10 \
--train-way 5 \
--val-shot 10 \
--val-way 5
To test ProtoNet MTM Inner First-Order with 64-64-64-64 backbone on FC100 10-shot 5-way benchmark, run:
python protonet/test.py --dataset FC100 --shot 10 --way 5
To train ProtoNet MTM SPSA with 64-64-64-64 backbone on tieredImageNet 5-shot 2-way benchmark, run:
python protonet/train.py \
--dataset tieredImageNet \
--train-shot 5 \
--train-way 2 \
--val-shot 5 \
--val-way 2
To test ProtoNet MTM SPSA with 64-64-64-64 backbone on tieredImageNet 5-shot 2-way benchmark, run:
python protonet/test.py --dataset tieredImageNet --shot 5 --way 2
Our code uses some dataloaders from Torchmeta.
Code in maml folder is based on the extended implementation from Torchmeta and pytorch-maml. The code has been updated so that baseline scores more closely follow those of the original MAML paper.
Code in protonet folder is based on the implementation from MetaOptNet. All .py files in this folder except for dataloaders.py and optimize.py were adopted from this implementation and modified afterwards. A copy of Apache License, Version 2.0 is available in protonet folder.
If you want to cite our paper, you can use the following BibTeX entry:
@article{boiarov2021multi,
title={Simultaneous Perturbation Method for Multi-Task Weight Optimization in One-Shot Meta-Learning},
author={Boiarov, Andrei and Khabarlak, Kostiantyn and Yastrebov, Igor},
journal={arXiv preprint arXiv:2110.13188},
year={2021}
}