使用MistNet在COCO128数据集上协作训练Yolo-v5

本案例介绍如何在MNIST手写数字分类场景中，使用名为MistNet的聚合算法训练联邦学习作业。数据分散在不同的地方（如边缘节点、摄像头等），由于数据隐私和带宽的原因，无法在服务器上聚合。因此，我们不能将所有数据都用于训练。在某些情况下，边缘节点的计算资源有限，甚至没有训练能力。边缘无法从训练过程中获取更新的权重。因此，传统算法（例如，联合平均算法）通常聚合由不同边缘客户端训练的更新权重，在这种情况下无法工作。MistNet 被提议解决这个问题。

MistNet 将 DNN 模型分为两部分，边缘侧的轻量级特征提取器用于从原始数据生成有意义的特征，以及包含云中最多模型层的分类器，用于针对特定任务进行迭代训练。MistNet 实现了可接受的模型效用，同时大大减少了已发布的中间功能造成的隐私泄露。

物体检测实验

假设有两个边缘节点和一个云节点。由于隐私问题，边缘节点上的数据无法迁移到云中。基于此场景，我们将演示mnist示例

安装Sedna

准备数据集

Create data interface for EDGE1_NODE.

mkdir -p /data/1
cd /data/1
wget https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128.zip
unzip coco128.zip -d COCO

Create data interface for EDGE2_NODE.

准备镜像

此示例使用以下映像：
聚合工作器：kubeedge/sedna-example-federated-learning-mistnet-yolo-aggregator:v0.4.0
训练工作器：kubeedge/sedna-example-federated-learning-mistnet-yolo-client:v0.4.0
这些图像是由脚本build_images.sh生成的。

创建联合学习作业

创建用于$EDGE1_NODE和$EDGE2_NODE的数据集

kubectl create -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: Dataset
metadata:
  name: "coco-dataset-1"
spec:
  url: "/data/1/COCO"
  format: "dir"
  nodeName: kubeedge1
EOF

kubectl create -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: Dataset
metadata:
  name: "coco-dataset-2"
spec:
  url: "/data/2/COCO"
  format: "dir"
  nodeName: kubeedge2
EOF

创建模型
在 $EDGE 1_NODE 和 $EDGE 2_NODE 中创建目录 /model 和 /pretrained 。

mkdir -p /model
mkdir -p /pretrained

在$CLOUD_NODE主机上创建目录/model和/pretrained(下载链接在这里)

# on the cloud side
mkdir -p /model
mkdir -p /pretrained
cd /pretrained
wget https://kubeedge.obs.cn-north-1.myhuaweicloud.com/examples/yolov5_coco128_mistnet/yolov5.pth

创建模型

kubectl create -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: Model
metadata:
  name: "yolo-v5-model"
spec:
  url: "/model/yolov5.pth"
  format: "pth"
EOF
 
kubectl create -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: Model
metadata:
  name: "yolo-v5-pretrained-model"
spec:
  url: "/pretrained/yolov5.pth"
  format: "pth"
EOF

使用您的S3用户凭据创建一个密钥。(可选)

开始联邦学习

kubectl create -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: FederatedLearningJob
metadata:
  name: yolo-v5
spec:
  pretrainedModel: # option
    name: "yolo-v5-pretrained-model"
  transmitter: # option
    ws: { } # option, by default
    s3: # optional, but at least one
      aggDataPath: "s3://sedna/fl/aggregation_data"
      credentialName: mysecret
  aggregationWorker:
    model:
      name: "yolo-v5-model"
    template:
      spec:
        nodeName: $CLOUD_NODE
        containers:
          - image: kubeedge/sedna-example-federated-learning-mistnet-yolo-aggregator:v0.4.0
            name: agg-worker
            imagePullPolicy: IfNotPresent
            env: # user defined environments
              - name: "cut_layer"
                value: "4"
              - name: "epsilon"
                value: "100"
              - name: "aggregation_algorithm"
                value: "mistnet"
              - name: "batch_size"
                value: "32"
              - name: "epochs"
                value: "100"
            resources: # user defined resources
              limits:
                memory: 8Gi
  trainingWorkers:
    - dataset:
        name: "coco-dataset-1"
      template:
        spec:
          nodeName: $EDGE1_NODE
          containers:
            - image: kubeedge/sedna-example-federated-learning-mistnet-yolo-client:v0.4.0
              name: train-worker
              imagePullPolicy: IfNotPresent
              args: [ "-i", "1" ]
              env: # user defined environments
                - name: "cut_layer"
                  value: "4"
                - name: "epsilon"
                  value: "100"
                - name: "aggregation_algorithm"
                  value: "mistnet"
                - name: "batch_size"
                  value: "32"
                - name: "learning_rate"
                  value: "0.001"
                - name: "epochs"
                  value: "1"
              resources: # user defined resources
                limits:
                  memory: 2Gi
    - dataset:
        name: "coco-dataset-2"
      template:
        spec:
          nodeName: $EDGE2_NODE
          containers:
            - image: kubeedge/sedna-example-federated-learning-mistnet-yolo-client:v0.4.0
              name: train-worker
              imagePullPolicy: IfNotPresent
              args: [ "-i", "2" ]
              env: # user defined environments
                - name: "cut_layer"
                  value: "4"
                - name: "epsilon"
                  value: "100"
                - name: "aggregation_algorithm"
                  value: "mistnet"
                - name: "batch_size"
                  value: "32"
                - name: "learning_rate"
                  value: "0.001"
                - name: "epochs"
                  value: "1"
              resources: # user defined resources
                limits:
                  memory: 2Gi
EOF