(MobiCom'24) Delta: A Cloud-assisted Data Enrichment Framework for On-Device Continual Learning

Welcome to the repository of Delta, an efficient and effective data enrichment framework designed for on-device continual learning (CL).

The code repository is currently undergoing updates as we are re-organizing the code for improved clarity.

Table of Contents

1. Repository Structure & Description
2. Requirements
3. Datasets & Preparation
4. Run Commands
5. Acknowledgments and Note

1. Repository Structure & Description

├──root/Experiments/MobiCom24_Delta     # Root path of the repository
    ├──Agents                           # Files for different continual learning algorithms
        ├──base.py                      # Abstract class for algorithms
        ├──fskd.py                      # Few-shot CL with knowledge distillation
        ├──fspr.py                      # Few-shot CL with parameter freezing
        ├──fsro.py                      # Few-shot CL with robust optimization
        ├──fed_cl.py                    # Federated CL algorithm
        ├──test.py                      # CL with data enrichment algorithms (vanilla, random and delta) 
        ├──delta_class.py               # Implementation of delta operations (device-side softmatching and cloud-side sampling)
    ├──Buffer                           # Files related to buffer management
        ├──buffer.py                    
        ├──name_match.py                # Name-to-function mappings
        ├──random_retrieve.py           # Random retrieval methods
        ├──reservoir_update.py          # Random update methods
    ├──Data                             # Files for create the data stream objects of different datasets
        ├──RawData                      # Raw data files and corresponding preprocessing scripts for each task
            ├──cifar-10-C               # CIFAR-10-C dataset
            ├──har                      # HHAR, UCI, Motion, Shoaib datasets
                ├──1.preprocess.py      # Preprocessing script
            ├──textclassification       # XGLUE dataset
                ├──1.preprocess.py      # Preprocessing script
        ├──CloudData                    # Cloud-side data for each task, including public raw data and processed directory dataset
            ├──cifar-10-C               
            ├──har                      
            ├──textclassification       
        ├──cloud.py                     # File for cloud-side operations to generate directory dataset
        ├──continumm.py                 # Data stream object creation
        ├──name_match.py                # Name-to-function mappings
        ├──utils.py 
    ├──Experiment                       # Files for running specified agents (algorithms) multiple times 
        ├──run.py                       
    ├──Figures                          # Directory for saving figures of experimental results 
    ├──Log                              # Directory for saving final and intermediate results during experiments
    ├──Models                           # Files for backbone models and corresponding pre-training process on cloud server
        ├──Pretrain                     # Directory for saving pre-trained model weights
        ├──HAR_model.py                 # DCNN model for HAR task (requires pretraining)
        ├──resnet.py                    # ResNet model for image tasks (uses pretrained weights from PyTorch)
        ├──speechmodel.py               # VGG model for audio tasks (uses pretrained weights from PyTorch)
        ├──pretrain.py                  # Methods for loading models with pretrained weights
    ├──2-cloud_pretrain.py              # Cloud-side model pretraining
    ├──3-cloud_preprocess.py            # Cloud-side data processing
    ├──4-main.py                        # On-device continual learning with specified configurations
    ├──5-plot.py                        # Script for plotting the experimental results

2. Requirements

Ensure you have the following dependencies installed:

3. Datasets

3.1 Downloading Raw Data

All raw data files can be downloaded from Google Drive. Alternatively, you can download each dataset from the following sources:

• Image Classification: CIFAR-10-C
• Human Activity Recognition: HHAR, UCI, Motion, Shoaib
• Text CLassification: Microsoft XGLUE
• Audio Recognition: Google Speech Commands

3.2 Preprocessing Raw Data

3.3 Preparing Data for Each Context

4. Run Commands

1. Preprocess Raw Data: Run 1-preprocess.py for each dataset repository in Data/RawData/.
2. Cloud-side Pretraining: Execute 2-cloud_pretrain.py to pretrain models on cloud server. (Note that ResNet and Transformers can directly load pre-trained weights provided by PyTorch)
3. Cloud-side Data Processing: Run 3-cloud_preprocess.py to process public data on the cloud server and generate directory weights/cluster centers.
4. On-Device Continual Learning: Run 4-main.py for each task with specific commands and configurations provided in Scripts/run_main.sh, and save results in Log/.
5. Plot Experimental Results: Run 5-plot.py to output and visualize the experimental results, and save figures in Figures/.

5. Acknowledgments and Note

Our code is built upon the repositories of online-continual-learning and Miro. We extend our sincere gratitude to their foundational work.

If you have any problems, please feel free to contact us.