[Official] MetaVers: Meta-Learned Versatile Representations for Personalized Federated Learning

Installation

• Create a virtual environment with conda create

  • conda create -n mvs python=3.8.5

  • conda activate mvs

  • pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

  • pip install learn2learn

  • pip install tensorboardX

  • pip install tensorflow

  • pip install sklearn

---

Fast Run

Results on CIFAR-10

• Run: python trainer.py --dataset cifar10 --rand 0 --seed_list 42

---

Results on CIFAR-100

• Run: python trainer.py --dataset cifar100 --rand 0 --seed_list 42

---

Results on CINIC-10

• Run: python trainer.py --dataset cinic --rand 0 --seed_list 42

---

For Details

--loss: What loss to use, for cross-entropy; ce or tirplet; riplet . Default: hybrid

--d_from: How to set the margin of the parameter used for triplet loss, for fixed margin;margin Default: select

--w: Interval Value $W$. Default: 50

--control: Non-random way, according to the benchmark maximum way. Default: fixed

--seed_list: Receive a list of how to make a random seed (adjust the result randomness)

--method: Whether episodes are created locally in a decentralized setting, for centralized setting; centralized Default: decentralized

--lr: Learning rate for gradient update. Default: 0.001

--optimizer: Whether to use decay for Adam optimizer. When using a weight decay of 1e-3: 1 Default: 0

--evaluation_unit: Evaluate using the validation set every n round. Default: 50

--activated: Number of the activated clients. Default: 5

--embedder: which model to use.[conv4, lenet,resnet18] Default: lenet

--gamma: loss balancing hyper parameter. Default: 0.5

--rand: Fix the randomness of the PyTorch according to the seed.(adjust the result randomness) Default: 1

---

MetaVers: PFL Results Performance Reproduction

PFL Results Performance Reproduction on CIFAR-10

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cifar10 --num_user 50 --optimizer 1 --train_shot 20 --train_query 30 --gamma 0.4

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cifar10 --num_user 100 --optimizer 1 --train_shot 20 --train_query 30 --gamma 0.3

• ```python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cifar10 --num_user 500 --optimizer 1 --train_shot 20 --train_query 30 --gamma 0.2 ``

PFL Results Performance Reproduction on CIFAR-100

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cifar100 --num_user 50 --optimizer 0 --shot_500 3 --train_shot 10 --train_query 20 --gamma 0.6

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cifar100 --num_user 100 --optimizer 0 --shot_500 3 --train_shot 10 --train_query 20 --gamma 0.6

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 500 --control fixed --seed_list 0 21 42 --dataset cifar100 --num_user 50 --optimizer 0 --shot_500 3 --train_query 15 --gamma 0.6

PFL Results Performance Reproduction on CINIC-10

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cinic --num_user 50 --optimizer 0 --train_shot 25 --train_query 35 --gamma 0.5

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cinic --num_user 100 --optimizer 0 --train_shot 25 --train_query 35 --gamma 0.5

• python trainer.py --exp_name reproduction --loss hybrid --d_from select --w 50 --control fixed --seed_list 0 21 42 --dataset cinic --num_user 500 --optimizer 0 --train_shot 20 --train_query 30 --gamma 0.5

---

Additional Experiment

[1] Single Loss Example

• Triplet Loss (only $\mathcal{L}_{T}$) python trainer.py --loss triplet --d_from margin --margin 3.0 --w 50 --dataset cinic --num_user 500 --exp_name triplet --seed_list 0 21 42 --optimizer 0 --train_shot 25 --train_query 35 --gamma 0.5 --lr 0.001 --gpu_number 0
• Cross-entropy Loss (only $\mathcal{L}_{S}$) python trainer.py --loss ce --dataset cinic --num_user 500 --exp_name ce --seed_list 0 21 42 --optimizer 0 --train_shot 20 --train_query 30 --gamma 0.5 --lr 0.001 --gpu_number 0

[2] Lower Way Example

• python trainer.py --loss hybrid --d_from select --w 50 --dataset cifar100 --num_user 100 --exp_name table4_low --low_way 5 --seed_list 0 21 42 --optimizer 0 --train_shot 10 --train_query 20
• Random Way python trainer.py --random_way various --loss hybrid --d_from select --w 50 --dataset cinic --num_user 100 --exp_name table4_random --low_way 3 --seed_list 0 21 42 --optimizer 0 --train_shot 25 --train_query 35

[3] Centralized Learning Example

• python trainer.py --version eccv --loss hybrid --d_from select --w 50 --dataset cifar100 --num_user 50 --exp_name table5_central --method centralized --seed_list 0 21 42 --optimizer 0 --train_shot 10 --train_query 20

---

CINIC-10 Dataset Download

• If the CINIC-10 dataset is not downloaded correctly, you can download it directly from the link and put it in the "'./cinic" folder. https://datashare.is.ed.ac.uk/bitstream/handle/10283/3192/CINIC-10.tar.gz