ahmethakanbesel / fedperl

An enhanced PyTorch implementation for MICCAI 2021 paper "FedPerl: Semi-Supervised Peer Learning for Skin Lesion Classification"

FedPerl: Semi-Supervised Peer Learning

Installing Requirements

pip install -r requirements.txt

Preparing Dataset

Converting Images to Numpy Format

There are three examples in dataset folder which show how you can convert images to .npy format. You should create two separate files for labels and images. You can store images and labels in two separate arrays. Label of images[i] should be located at labels[i]. Labels must be integers. If dataset contains string labels you can create a dictionary to replace them with integers.

Create Dataset Class

After creating .npy version of your dataset you need to prepare class for the dataset. You can find example dataset classes under src/data folder. Dataset class must implement abstract Dataset class (src/data/dataset.py). Dataset class contains several functions to return indexes of images for given client or the server. You should look at src/data/isic2019.py for additional implementation details.

Updating settings.py

After implementing your dataset class you need to reserve a name for your dataset. src/modules/settings.py will load the dataset by getting its name from .env file. If you created a new dataset you need to update the if block in src/modules/settings.py

Setting Environment Variables

.env.example files should be copied as .env. There are 5 fields in the .env file.

Field Name	Value
IMG_PATH	Stores the absolute path of .npy file for the images
LBL_PATH	Stores the absolute path of .npy file for the labels
MODEL	Defines which architecture will be used (efficientnet, efficientnet_legacy, densenet)
MODEL_FILE	Stores the absolute path of your model file after finishing the training process. This is used while evaluating the global model.
DATASET	Defines which dataset will be used (brain, ham10000, isic2019)

Running FedPerl

You can run src/train_Perl.py by using command line, or you can set hyperparameters inside that file and run it as a Python file without using command line. Before running the file be sure that folders for storing states, weights, checkpoints and models are created.

Model Evaluation

Evaluate Global Model

You can run src/eval/evaluate_global.py to evaluate the model defined in .env file. It stores the results in a SQLite database src/eval/results.db and an Excel sheet src/eval/All_Scores_Global.xlsx.

Evaluate Clients' Models

You can run src/eval/generate_results.py to evaluate the model defined in .env file. It stores the results in an Excel workbook src/eval/All_Scores.xlsx.

Plotting

Files under src/vis contains various files for plotting data distribution among clients from different aspects.