FedAvg Optimization On LQR Control Of A 1-Dimensional Point Mass "Robot"

Implementation of the vanilla federated learning paper in setting of LQR control of a 1-dimensional point mass "robot" shown as Fig 1: Communication-Efficient Learning of Deep Networks from Decentralized Data and Personalized Federated Learning of Driver Prediction Models for Autonomous Driving.

Experiments are produced on synthetic non-iid dataset.

Since the purpose of these experiments are to illustrate the performance of the federated learning paradigm, only simple models such as DynamicsModel(LQR controller and unit-mass system) are used.

Requirments

Install all the packages from requirments.txt

• Python3.7
• Pytorch
• Torchvision

Data

Execute the following code to generate a synthetic dataset involving 3 robots in the ./data directory.

python3 src/generate_data_main.py

Running the experiments

The baseline experiment trains the model on the default data of robot 1.

• To run the baseline experiment with synthetic dataset on DynamicsModel using CPU:

python3 src/baseline_main.py --lr=0.01 --local_ep=300 --local_bs=10

• Or to run it on GPU (eg: if gpu:0 is available):

python3 baseline_main.py --lr=0.01 --local_ep=300 --local_bs=10 --gpu=0

---

Federated experiment involves training a global model using many local models.

• To run the federated experiment with synthetic dataset on DynamicsModel:

python3 federated_main.py --lr=0.01 --local_ep=50 --local_bs=10 --gpu=0

You can change the default values of other parameters to simulate different conditions. Refer to the options section.

Options

The default values for various paramters parsed to the experiment are given in options.py. Details are given some of those parameters:

• --dataset: Default: 'synthetic dataset'.
• --model: Default: 'DynamicsModel'.
• --gpu: Default: None (runs on CPU). Can also be set to the specific gpu id.
• --epochs: Number of rounds of training.
• --lr: Learning rate set to 0.01 by default.
• --verbose: Detailed log outputs. Activated by default, set to 0 to deactivate.
• --seed: Random Seed. Default set to 1.

Federated Parameters

• --iid: Distribution of data amongst users. Default set to IID. Set to 0 for non-IID.
• --num_users:Number of robots. Default is 3.
• --frac: Fraction of robots to be used for federated updates. Default is 1.0.
• --local_ep: Number of local training epochs in each robots. Default is 10.
• --local_bs: Batch size of local updates in each robots. Default is 10.
• --unequal: Used in non-iid setting. Option to split the data amongst users equally or unequally. Default set to 0 for equal splits. Set to 1 for unequal splits.

Results on Synthetic Dataset

Baseline Experiment:

The experiment involves training a single model in the conventional way.

Parameters:

• Optimizer: : Adma
• Learning Rate: 0.01

Fig 2: Training LQR Loss after training for 300 epochs:

---

Federated Experiment:

The experiment involves training a global model in the federated setting.

Federated parameters (default values):

• Fraction of users (C): 1.0
• Local Batch size (B): 10
• Local Epochs (E): 50
• Optimizer : Adma
• Learning Rate : 0.01
  

Fig 3: Traing Loss after training for 30 global epochs with:

Reference

Papers:

• Communication-Efficient Learning of Deep Networks from Decentralized Data
• Personalized Federated Learning of Driver Prediction Models for Autonomous Driving

Source Code:

• Federated-Learning (PyTorch)