This repository is used to evaluate how federated learning can be improved using AutoML techniques.

⚠️ Please use Python 3.8 to run this project, PyTorch pip packaging is currently not working with Python 3.9 (pytorch/pytorch#47116).

⚠️ Due to a Windows bug do not upgrade the current numpy and stay with version 1.19.3 (https://github.com/numpy/numpy/wiki/FMod-Bug-on-Windows)

Create a new virtual environment and install the requirements in requirements.txt

Add the repository root to your python path e.g. on windows set PYTHONPATH=%cd%

Run the sacred experiments in mlmi/experiments

e.g.

run fedavg with the MNIST dataset python mlmi/experiments/fedavg.py with mnist

run fedavg with the FEMNIST dataset python mlmi/experiments/fedavg.py

Reptile is a meta-learning algorithm. In this project, we apply it to the federated learning context. Our implementation is based on the code from the Nichol 2018 paper.

To run our federated version of the Reptile algorithm on either the FEMNIST or Omniglot dataset, execute python mlmi/experiments/reptile_federated.py. All relevant parameters can be specified using a sacred experiment configuration in reptile_federated.py.

There are three additional runnable files in mlmi/reptile that we created for validation purposes. All of them use only the Omniglot dataset:

• run_reptile_original.py runs the original experimental setting from the Nichol paper, including the same data loading. Minor but important adaptations have been made: (1) command-line options not relevant to the federated learning setting have been removed and (2) the optimizer state is reset after every task. The latter was not the case in the original source.
• run_reptile_original_dataloading_federated.py runs an experiment with the TensorFlow model from the Nichol paper, but with a federated data loading design of the Omniglot data.
• run_reptile_federated.py runs the full adaptation of Reptile to the federated learning setting

All experiments can be specified with the same command-line options:

--seed:             random seed (default 0)
--classes:          number of classes per inner task (default 5)
--shots:            number of examples per class (default 5)
--inner-batch:      inner batch size (default 10)
--inner-iters:      inner iterations (default 5)
--learning-rate:    Adam step size (default 1e-3)
--meta-step:        meta-training step size (default 1)
--meta-step-final:  meta-training step size by the end (default 0)
--meta-batch:       meta-training batch size (default 5)
--meta-iters:       meta-training iterations (default 3000)
--eval-iters:       eval inner iterations (default 50)
--eval-samples:     evaluation samples (default 100)
--eval-interval:    train steps per eval (default 10)
--sgd:              use gradient descent instead of Adam as inner optimizer

In addition, run_reptile_original_dataloading_federated.py and run_reptile_federated.py take the following options:

--train-clients:    number of training clients (default 10000)
--test-clients:     number of test clients (default 1000)

# transductive 1-shot 5-way Omniglot.
python -u run_omniglot.py --shots 1 --inner-batch 10 --inner-iters 5 --meta-step 1 --meta-batch 5 --meta-iters 100000 --eval-batch 5 --eval-iters 50 --learning-rate 0.001 --meta-step-final 0 --train-shots 10 --checkpoint ckpt_o15t --transductive

# 5-shot 5-way Omniglot.
python -u run_omniglot.py --train-shots 10 --inner-batch 10 --inner-iters 5 --learning-rate 0.001 --meta-step 1 --meta-step-final 0 --meta-batch 5 --meta-iters 100000 --eval-batch 5 --eval-iters 50 --checkpoint ckpt_o55

# 1-shot 5-way Omniglot.
python -u run_omniglot.py --shots 1 --inner-batch 10 --inner-iters 5 --meta-step 1 --meta-batch 5 --meta-iters 100000 --eval-batch 5 --eval-iters 50 --learning-rate 0.001 --meta-step-final 0 --train-shots 10 --checkpoint ckpt_o15

# 1-shot 20-way Omniglot.
python -u run_omniglot.py --shots 1 --classes 20 --inner-batch 20 --inner-iters 10 --meta-step 1 --meta-batch 5 --meta-iters 200000 --eval-batch 10 --eval-iters 50 --learning-rate 0.0005 --meta-step-final 0 --train-shots 10 --checkpoint ckpt_o120

# 5-shot 20-way Omniglot.
python -u run_omniglot.py --classes 20 --inner-batch 20 --inner-iters 10 --meta-step 1 --meta-batch 5 --meta-iters 200000 --eval-batch 10 --eval-iters 50 --learning-rate 0.0005 --meta-step-final 0 --train-shots 10 --checkpoint ckpt_o520

The project uses Tensorboard to log results.

1. Start Tensorboard tensorboard --logdir run
2. View results in Browser at http://localhost:6006