Federated Re-ID

Federated Learning on Person Re-Identification

Abstract

This project covers my research experiments about Federated Learning on Person Re-Identification. Our primary goal is jointly optimizing performance on seen and unseen domains. One feature of this project is reproducing experimental results presented in Decentralised Person Re-Identification with Selective Knowledge Aggregation [1].

News

  • Dec 6th, 2022. IBN-Net [2] is now available. The source code of IBN-Net for ReID is referenced from reid-strong-baseline.

Source Code

GitHub Link

Usage

  1. Clone the source code by 
    git clone https://github.com/dogsc729/Federated-Learning-on-Re-ID.git
  2. Create two directories, checkpoint and datasets, under the cloned repositories by 
    mkdir checkpoint

    and

    mkdir datasets

  3. You should have four datasets already, including Market1501, MSMT17, DukeMTMC-reID and CUHK03-np-detected.

    Unzip the four datasets and place them in datasets, the structure should be:

     datasets
 |- /Market
 |    |- /bounding_box_test
 |    |- /bounding_box_train
 |    |- /gt_bbox
 |    |- /gt_query
 |- /MSMT17
 |    |- /bounding_box_test
 |    |- /bounding_box_train
 |- /DukeMTMC-reID
 |    |- /bounding_box_test
 |    |- /bounding_box_train
 |- /cuhk03-np-detected
     |- /bounding_box_test
     |- /bounding_box_train

    Note that the structure above including the naming should be exactly the same.

  4. Pre-process the datasets by running python3 ./src/big_data_preprocess.py
  5. Start the training by python3 ./src/federated_train.py. In addition, you can change the settings by adding the arguments below.
    • -s, --scenario: You can change the training scenario by selecting ska for Selective Knowledge Aggregation or fed for classic Federated Learning, The default value is ska.
    • -l, --location: You can change the location of the directory under /checkpoint/. The log file, models, and record of training progress in .png file will be stored here. The default value is the time you start the training.
    • -m, --model: You can change the type of model by selecting attentive for Attentive normalization ResNet50, vanilla for vanilla ResNet50 or ibn for IBN-Net. The default value is attentive.
    • --global_iter: The number of iteration of the global training stage. The default value is 100.
    • --local_epoch: The number of epoch trained on each client model. The default value is 1.
    • lr_feature: The learning rate of the feature extraction layers of the model. The default value is 0.01.
    • lr_classifier: The learning rate of the classifier layers of the model. The default value is 0.1.

    For example, You can run python3 ./src/federated_train.py -s fed -l federated_test -m vanilla --local_epoch 5 to set your experiment on classic Federated Learning scenario, checkpoint location at /checkpoint/federated_test, using vanilla ResNet50 as your model and set the number of local epoch trained for each global round as 5.

Reference

  • [1] Shitong Sun, Guile Wu, Shaogang Gong. Decentralised Person Re-Identification with Selective Knowledge Aggregation. In BMVC, 2021.
  • [2] Xingang Pan, Ping Luo, Jianping Shi, Xiaoou Tang. Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net, In ECCV, 2018.