Code for the paper "Continual Federated Learning Based on Knowledge Distillation"

• python 3.7+
• cuda 9.0+
• pytorch 1.7+

In the code, we use 0~8 to denote the catagories in text classification tasks: Finance(0), Entertainment(1), Education(2), Fashion(3), Stock(4), Technology(5), Home(6), Sports(7), Social(8).

There is only the Sina2019 dataset which is generated by ourselves in the ./dataset folder.

Limited by the maximium size of uploaded files, only the subsets of the original dataset are included under the path ./dataset . To reproduce the work, you need to download the complete datasets:

• THUCNews: https://thuctc.thunlp.org/
• Sogou News: https://www.sogou.com/labs/resource/cs.php
• Sina2019: provided after the paper is accepted
• NLPIR Weibo Corpus: https://www.nlpir.org/wordpress/download/weibo.7z
• CIFAR-10 & CIFAR-100: https://www.cs.toronto.edu/~kriz/cifar.html
• Caltech-256: https://www.cs.toronto.edu/~kriz/cifar.html

To construct the required tasks, we provide some tools in construct_tasks.py.

1. Run the code like this: python3 federated_run.py --model=TextCNN --paradigm=CFeD --scenario=class-IL ,where the parameter paradigm denotes which method is choosen and scenario.

• positional arguments:
  • model:{TextCNN, TextCNNMultihead, CNN_Cifar10, CNN_Cifar10_multihead, CNN_Cifar100}
  • paradigme:{CFeD, dmc, lwf, ewc, multihead, normal}
  • scenario:{domain, class}

2. Limited by the maximum size of the uploaded file, we decrease the hyper-parameters, like batch_size. To change the hyper-parameters, you can modify the specific model files under ./model, like ./model/TextCNN.py.