sedna/examples/lifelong_learning/robot_dog_delivery

• Using Lifelong Learning in Campus Robot Delivery Scenario
• • Introduction
  • System Architecture
  • • Install Sedna
    • Prepare Dataset
    • Prepare Images
    • Create Lifelong Learning Job
    • • Preparations
      • Create s3 storage resources
      • Create Initial Dataset Resource
      • Create robot-dog-delivery Lifelong Learning Job
    • Check Lifelong Learning Status
    • Effect Display

Using Lifelong Learning in Campus Robot Delivery Scenario

Introduction

This example introduces how to use Sedna lifelong learning to implement the lifelong learning delivery task of the robot in the campus.

This demo mainly shows:

1. Lifelong learning unseen task recognition algorithm (prediction) can identify and collect unseen task samples.
2. Lifelong learning unseen task recognition algorithm (detection) can trigger alarms to remind robot admin that emergency response is required.
3. Lifelong learning unseen task training algorithm improves the inference accuracy of known categories and the ability to identify new categories.

System Architecture

Install Sedna

Follow the Sedna installation document to install Sedna.

Prepare Dataset

Step1: download data.txt
and upload it to a specfied s3 directory. In this example, the directory is s3://kubeedge/sedna-robo/sedna_data.

Step2: download and decompress sedna_data.zip,
then upload the training images in it to the s3 directory where data.txt is stored.

Prepare Images

This example uses these images:

1. training worker: swr.cn-south-1.myhuaweicloud.com/sedna/sedna-robo:v0.1.1
2. evaluate worker: swr.cn-south-1.myhuaweicloud.com/sedna/sedna-robo:v0.1.1
3. inference worker: swr.cn-south-1.myhuaweicloud.com/sedna/sedna-robo-infer:v0.1.1

These images are generated by the script build_images.sh.

Create Lifelong Learning Job

Preparations

Before this, user must provide an s3 directory for storage which is denoted as "s3_prefix" in this example, by setting environment variable s3_prefix.

s3_prefix=$s3_prefix
cloud_image=swr.cn-south-1.myhuaweicloud.com/sedna/sedna-robo:v0.1.1
edge_image=swr.cn-south-1.myhuaweicloud.com/sedna/sedna-robo-infer:v0.1.1
data_url=$s3_prefix/sedna_data/data.txt

DATA_NODE=sedna-mini-control-plane 
TRAIN_NODE=sedna-mini-control-plane
EVAL_NODE=sedna-mini-control-plane
INFER_NODE=sedna-mini-control-plane
OUTPUT=$s3_prefix/lifelonglearningjob/output
job_name=robo-demo

Create s3 storage resources

Before this, users must generate the S3_ENDPOINT, ACCESS_KEY_ID and SECRET_ACCESS_KEY of their own s3 accounts and set environment
variables S3_ENDPOINT, ACCESS_KEY_ID and SECRET_ACCESS_KEY.

action=${1:-create}

kubectl $action -f - <<EOF
apiVersion: v1
stringData:
  ACCESS_KEY_ID: $ACCESS_KEY_ID
  SECRET_ACCESS_KEY: $SECRET_ACCESS_KEY
kind: Secret
metadata:
  name: my
  annotations:
    s3-endpoint: $S3_ENDPOINT
type: Opaque
EOF

Create Initial Dataset Resource

kubectl $action -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: Dataset
metadata:
  name: lifelong-dataset-robo
spec:
  url: "$data_url"
  format: "txt"
  nodeName: "$DATA_NODE"
  credentialName: my
EOF

Create robot-dog-delivery Lifelong Learning Job

action=${1:-create}

kubectl $action -f - <<EOF
apiVersion: sedna.io/v1alpha1
kind: LifelongLearningJob
metadata:
  name: $job_name
spec:
  dataset:
    name: "lifelong-dataset-robo"
    trainProb: 0.8
  trainSpec:
    template:
      spec:
        nodeName: $TRAIN_NODE
        dnsPolicy: ClusterFirstWithHostNet
        containers:
          - image: $cloud_image
            name:  train-worker
            imagePullPolicy: IfNotPresent
            args: ["sedna_train.py"]
            env:
              - name: "train_ratio"
                value: "0.9"
              - name: "BACKEND_TYPE"
                value: "PYTORCH"
            resources:
              limits:
                cpu: 6
                memory: 6Gi
              requests:
                cpu: 3
                memory: 3Gi
            volumeMounts:
            - mountPath: /dev/shm
              name: cache-volume
        volumes:
        - emptyDir:
            medium: Memory
            sizeLimit: 128Mi
          name: cache-volume
    trigger:
      checkPeriodSeconds: 30
      timer:
        start: 00:00
        end: 24:00
      condition:
        operator: ">"
        threshold: 100
        metric: num_of_samples
  evalSpec:
    template:
      spec:
        nodeName: $EVAL_NODE
        dnsPolicy: ClusterFirstWithHostNet
        containers:
          - image: $cloud_image
            name:  eval-worker
            imagePullPolicy: IfNotPresent
            args: ["sedna_evaluate.py"]
            env:
              - name: "operator"
                value: "<"
              - name: "model_threshold"  # Threshold for filtering deploy models
                value: "0"
              - name: "BACKEND_TYPE"
                value: "PYTORCH"
            resources:
              limits:
                cpu: 6
                memory: 6Gi
              requests:
                cpu: 3
                memory: 3Gi
  deploySpec:
    template:
      spec:
        nodeName: $INFER_NODE
        dnsPolicy: ClusterFirstWithHostNet
        hostNetwork: true
        containers:
        - image: $edge_image
          name:  infer-worker
          imagePullPolicy: IfNotPresent
          env:
            - name: "BIG_MODEL_IP"
              value: "http://94.74.91.114"
            - name: "BIG_MODEL_PORT"
              value: "30001"
            - name: "RUN_FILE"
              value: "integration_main.py"
          resources:  # user defined resources
            limits:
              cpu: 6
              memory: 6Gi
            requests:
              cpu: 3
              memory: 3Gi

  credentialName: my
  outputDir: $OUTPUT/$job_name
EOF

Check Lifelong Learning Status

kubectl get lifelonglearningjob

Effect Display