nikamanthab/Federated_Pneumonia

Federated Learning Strategies for Image Classification

Exploring FL aggregation and robustness on distributed model training.

Data-set: https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia

Instructions to run:

• Install the requirements - torch, pandas, wandb, flask
• Clone the repo in multiple systems/ run multiple instances with changed input parameters
• Download the dataset and form the following folder structure: x-ray/(train or test)/(NORMAL or PNEUMONIA)
• Select the aggregator (fedavg, geometric median, comed) and the model architecture
• Start the server and start the client instances with appropriate parameters like node_name and train_csv
• CNN Architecture can be altered in the models.py and modelloader.py
• Execute the run.py file in the program folder

Aggregation Algorithms:

• FedAvg - weighted federated averaging (Torch CUDA supported)
• Geometric median - robust aggregation (Torch CUDA supported)
• COMED - Coordinate wise median (Numpy CPU support only)

Model Architectures:

• TwoLayerNet - Basic two layer CNN architecture
• ResNeXt50 - PyTorch implementation of resnext with added FCN layers.
• ResNet18 - PyTorch implementation of resnet with added FCN layers.

Hardware Support:

• Supports GPU in main branch
• For simulation with PySyft refer syft branch (supports only CPU)

Example:

    Server System / Terminal 0:
    cd src/aggregator
    export FLASK_APP=server.py
    python server.py
    
    System 1 / Terminal 1:
    python run.py --node_name node_0 --train_csv ../../csv/train_0.csv
    
    System 2 / Terminal 2:
    python run.py --node_name node_1 --train_sv ../../csv/train_1.csv

Command to run:

SERVER:
    
    usage: server.py [-h] [--architecture ARCHITECTURE]
             [--test_batch_size TEST_BATCH_SIZE] [--agg_epochs AGG_EPOCHS]
             [--device DEVICE] [--momentum MOMENTUM]
             [--log_interval LOG_INTERVAL] [--image_height IMAGE_HEIGHT]
             [--image_width IMAGE_WIDTH] [--aggregator AGGREGATOR]
             [--agg_iterations AGG_ITERATIONS]
             [--agg_optimizer AGG_OPTIMIZER] [--agg_optim_lr AGG_OPTIM_LR]
             [--agg_optim_momentum AGG_OPTIM_MOMENTUM] [--wandb WANDB]
             [--num_of_nodes NUM_OF_NODES] [--test_csv TEST_CSV]
             [--data_location DATA_LOCATION]
             [--aggregated_model_location AGGREGATED_MODEL_LOCATION]
             [--labels LABELS]

    Server module.

    optional arguments:
      -h, --help            show this help message and exit
      --architecture ARCHITECTURE
                            TwoLayerNet, ResNeXt50, ResNet18
      --test_batch_size TEST_BATCH_SIZE
      --agg_epochs AGG_EPOCHS
      --device DEVICE
      --momentum MOMENTUM
      --log_interval LOG_INTERVAL
      --image_height IMAGE_HEIGHT
      --image_width IMAGE_WIDTH
      --aggregator AGGREGATOR
                            fedavg, codem, geomed
      --agg_iterations AGG_ITERATIONS
      --agg_optimizer AGG_OPTIMIZER
                            Adam, SGD supported
      --agg_optim_lr AGG_OPTIM_LR
      --agg_optim_momentum AGG_OPTIM_MOMENTUM
      --wandb WANDB
      --num_of_nodes NUM_OF_NODES
      --test_csv TEST_CSV
      --data_location DATA_LOCATION
      --aggregated_model_location AGGREGATED_MODEL_LOCATION
      --labels LABELS
CLIENT:
    cd src/client
    python run.py

    usage: run.py [-h] [--node_name NODE_NAME] [--agg_ip AGG_IP]
              [--agg_port AGG_PORT] [--train_batch_size TRAIN_BATCH_SIZE]
              [--epochs EPOCHS] [--lr LR] [--device DEVICE]
              [--momentum MOMENTUM] [--log_interval LOG_INTERVAL]
              [--train_csv TRAIN_CSV] [--data_location DATA_LOCATION]
              [--wandb WANDB] [--model_location MODEL_LOCATION]
              [--labels LABELS]

    Client module.

    optional arguments:
      -h, --help            show this help message and exit
      --node_name NODE_NAME
      --agg_ip AGG_IP
      --agg_port AGG_PORT
      --train_batch_size TRAIN_BATCH_SIZE
      --epochs EPOCHS
      --lr LR
      --device DEVICE
      --momentum MOMENTUM
      --log_interval LOG_INTERVAL
      --train_csv TRAIN_CSV
      --data_location DATA_LOCATION
      --wandb WANDB
      --model_location MODEL_LOCATION
      --labels LABELS

Scores

Model	Number of nodes	Distribution	Epochs before aggregation	Aggregation_iter	Accuracy
Two layer CNN	2 nodes	0.5, 0.5	4 epochs	5 iters	0.8477