Incremental Learning
- JIE PU
Owned by JIE PU
Motivation
Data is continuously generated on the edge side. Traditionally, the data is collected manually and periodically retrained on the cloud to improve the model effect. This method wastes a lot of human resources, and the model update frequency is slow. Incremental learning allows users to continuously monitor the newly generated data and by configuring some triggering rules to determine whether to start training, evaluation, and deployment automatically, and continuously improve the model performance.
Goals
- Automatically retrains, evaluates, and updates models based on the data generated at the edge.
- Support time trigger, sample size trigger, and precision-based trigger.
- Support manual triggering of training, evaluation, and model update.
- support hard sample discovering of unlabeled data, for reducing the manual labeling workload.
- Support lifelong learning that reserves historical knowledge to avoid frequent re-training/ re-fine-tuning, and tackles samples uncovered in historical knowledge base.
Proposal
We propose using Kubernetes Custom Resource Definitions (CRDs) to describe the incremental learning specification/status and a controller to synchronize these updates between edge and cloud.
Use Cases
- Users can create the incremental learning jobs, by providing training scripts, configuring training hyperparameters, providing training datasets, configuring training and deployment triggers.
Design Details
We use the job word to represent the Constantly Iterative Update tasks including the train/eval/deploy task.
There are three stages in a job: train/eval/deploy.
Each stage contains these below states:
- Waiting: wait to trigger satisfied, i.e. wait to train/eval/deploy
- Ready: the corresponding trigger satisfied, now ready to train/eval/deploy
- Starting: the corresponding stage is starting
- Running: the corresponding stage is running
- Failed: the corresponding stage failed
- Completed: the corresponding stage completed
CRD API Group and Version
The IncrementalLearningJob CRD will be namespace-scoped. The tables below summarize the group, kind and API version details for the CRD.
- IncrementalLearningJob
| Field | Description |
|---|---|
| Group | edgeai.io |
| APIVersion | v1alpha1 |
| Kind | IncrementalLearningJob |
Incremental learning CRD
Below is the CustomResourceDefinition yaml for IncrementalLearningJob:
apiVersion: apiextensions.k8s.io/v1
kind: CustomResourceDefinition
metadata:
name: incrementallearningjobs.edgeai.io
spec:
group: edgeai.io
names:
kind: IncrementalLearningJob
plural: incrementallearningjobs
shortNames:
- incrementaljob
- ij
scope: Namespaced
versions:
- name: v1alpha1
subresources:
# status enables the status subresource.
status: {}
served: true
storage: true
schema:
openAPIV3Schema:
type: object
properties:
spec:
type: object
properties:
dataset:
type: object
properties:
name:
type: string
trainProb:
type: number
nodeName:
type: string
outputDir:
type: string
initialModel:
type: object
properties:
name:
type: string
trainSpec:
type: object
properties:
workerSpec:
type: object
properties:
scriptDir:
type: string
scriptBootFile:
type: string
frameworkType:
type: string
frameworkVersion:
type: string
parameters:
type: array
items:
type: object
properties:
key:
type: string
value:
type: string
trigger:
type: object
properties:
checkPeriodSeconds:
type: integer
timer:
type: object
properties:
start:
type: string
end:
type: string
condition:
type: object
properties:
operator:
type: string
enum: [">=",">","=","==","<=","<","ge","gt","eq","le","lt"]
threshold:
type: number
metric:
type: string
evalSpec:
type: object
properties:
workerSpec:
type: object
properties:
scriptDir:
type: string
scriptBootFile:
type: string
frameworkType:
type: string
frameworkVersion:
type: string
parameters:
type: array
items:
type: object
properties:
key:
type: string
value:
type: string
deploySpec:
type: object
properties:
model:
type: object
properties:
name:
type: string
trigger:
type: object
properties:
checkPeriodSeconds:
type: integer
timer:
type: object
properties:
start:
type: string
end:
type: string
condition:
type: object
properties:
operator:
type: string
enum: [">=",">","=","==","<=","<","ge","gt","eq","le","lt"]
threshold:
type: number
metric:
type: string
status:
type: object
properties:
conditions:
type: array
items:
type: object
properties:
type:
type: string
status:
type: string
lastProbeTime:
type: string
format: date-time
lastTransitionTime:
type: string
format: date-time
reason:
type: string
message:
type: string
data:
type: string
stage:
type: string
startTime:
type: string
format: date-time
completionTime:
type: string
format: date-time
active:
type: integer
succeeded:
type: integer
failed:
type: integer
additionalPrinterColumns:
- name: stage
type: string
description: The stage of the incremental learning job
jsonPath: ".status.conditions[-1].stage"
- name: status
type: string
description: The status of the incremental learning job
jsonPath: ".status.conditions[-1].type"
- name: Age
type: date
jsonPath: .metadata.creationTimestamp
Incremental learning job type definition
// +genclient
// +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object
// IncrementalLearningJob defines the incremental learing job crd
type IncrementalLearningJob struct {
metav1.TypeMeta `json:",inline"`
metav1.ObjectMeta `json:"metadata"`
Spec IncrementalLearningJobSpec `json:"spec"`
Status IncrementalLearningJobStatus `json:"status"`
}
//IncrementalLearningJobSpec describes the details configuration of incremental learning job
type IncrementalLearningJobSpec struct {
Dataset IncrementalDataset `json:"dataset"`
OutputDir string `json:"outputDir"`
NodeName string `json:"nodeName"`
InitialModel InitialModel `json:"initialModel"`
TrainSpec TrainSpec `json:"trainSpec"`
EvalSpec EvalSpec `json:"evalSpec"`
DeploySpec DeploySpec `json:"deploySpec"`
}
// TrainSpec describes the train worker
type TrainSpec struct {
WorkerSpec WorkerSpec `json:"workerSpec"`
Trigger Trigger `json:"trigger"`
}
// EvalSpec describes the train worker
type EvalSpec struct {
WorkerSpec WorkerSpec `json:"workerSpec"`
}
// DeploySpec describes the deploy model to be updated
type DeploySpec struct {
Model DeployModel `json:"model"`
Trigger Trigger `json:"trigger"`
}
// WorkerSpec describes the details to run the worker
type WorkerSpec struct {
ScriptDir string `json:"scriptDir"`
ScriptBootFile string `json:"scriptBootFile"`
FrameworkType string `json:"frameworkType"`
FrameworkVersion string `json:"frameworkVersion"`
Parameters []ParaSpec `json:"parameters"`
}
type Trigger struct {
CheckPeriodSeconds int `json:"checkPeriodSeconds,omitempty"`
Timer *Timer `json:"timer,omitempty"`
Condition Condition `json:"condition"`
}
type Timer struct {
Start string `json:"start"`
End string `json:"end"`
}
type Condition struct {
Operator string `json:"operator"`
Threshold float64 `json:"threshold"`
Metric string `json:"metric"`
}
type IncrementalDataset struct {
Name string `json:"name"`
TrainProb float64 `json:"trainProb"`
}
type InitialModel struct {
Name string `json:"name"`
}
type DeployModel struct {
Name string `json:"name"`
}
// +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object
// IncrementalLearningJobList is a list of incremental learning jobs.
type IncrementalLearningJobList struct {
metav1.TypeMeta `json:",inline"`
metav1.ListMeta `json:"metadata"`
Items []IncrementalLearningJob `json:"items"`
}
// IncrementalLearningJobStatus represents the current state of a incremental learning job
type IncrementalLearningJobStatus struct {
// The latest available observations of a incrementl job's current state.
// +optional
Conditions []IncrementalLearningJobCondition `json:"conditions,omitempty"`
// Represents time when the job was acknowledged by the job controller.
// It is not guaranteed to be set in happens-before order across separate operations.
// It is represented in RFC3339 form and is in UTC.
// +optional
StartTime *metav1.Time `json:"startTime,omitempty"`
// Represents time when the job was completed. It is not guaranteed to
// be set in happens-before order across separate operations.
// It is represented in RFC3339 form and is in UTC.
// +optional
CompletionTime *metav1.Time `json:"completionTime,omitempty"`
}
type IncrementalStageConditionType string
// These are valid stage conditions of a job.
const (
IncrementalStageCondWaiting IncrementalStageConditionType = "Waiting"
IncrementalStageCondReady IncrementalStageConditionType = "Ready"
IncrementalStageCondStarting IncrementalStageConditionType = "Starting"
IncrementalStageCondRunning IncrementalStageConditionType = "Running"
IncrementalStageCondCompleted IncrementalStageConditionType = "Completed"
IncrementalStageCondFailed IncrementalStageConditionType = "Failed"
)
// IncrementalLearningJobCondition describes current state of a job.
type IncrementalLearningJobCondition struct {
// Type of job condition, Complete or Failed.
Type IncrementalStageConditionType `json:"type"`
// Status of the condition, one of True, False, Unknown.
Status v1.ConditionStatus `json:"status"`
// Stage of the condition
Stage IncrementalLearningJobStage `json:"stage"`
// Last time the condition was checked.
// +optional
LastProbeTime metav1.Time `json:"lastProbeTime,omitempty"`
// Last time the condition transit from one status to another.
// +optional
LastTransitionTime metav1.Time `json:"lastTransitionTime,omitempty"`
// (brief) reason for the condition's last transition.
// +optional
Reason string `json:"reason,omitempty"`
// Human readable message indicating details about last transition.
// +optional
Message string `json:"message,omitempty"`
// The json data related to this condition
// +optional
Data string `json:"data,omitempty"`
}
// IncrementalLearningJobStage is a label for the stage of a job at the current time.
type IncrementalLearningJobStage string
const (
IncrementalLearningJobTrain IncrementalLearningJobStage = "Train"
IncrementalLearningJobEval IncrementalLearningJobStage = "Eval"
IncrementalLearningJobDeploy IncrementalLearningJobStage = "Deploy"
)
Validation
Open API v3 Schema based validation can be used to guard against bad requests. Invalid values for fields (example string value for a boolean field etc) can be validated using this.
Here is a list of validations we need to support :
- The
datasetspecified in the crd should exist in k8s. - The
modelspecified in the crd should exist in k8s. - The edgenode name specified in the crd should exist in k8s.
Incremental learning job sample
apiVersion: edgeai.io/v1alpha1
kind: IncrementalLearningJob
metadata:
name: helmet-detection-demo
spec:
initialModel:
name: "initial-model"
dataset:
name: "incremental-dataset"
trainProb: 0.8
trainSpec:
workerSpec:
scriptDir: "/model_train/yolov3_algorithms/"
scriptBootFile: "train.py"
frameworkType: "tensorflow"
frameworkVersion: "1.18"
parameters:
- key: "batch_size"
value: "32"
- key: "learning_rate"
value: "0.001"
- key: "max_epochs"
value: "100"
trigger:
checkPeriodSeconds: 60
timer:
start: 02:00
end: 04:00
condition:
operator: ">"
threshold: 500
metric: num_of_samples
evalSpec:
workerSpec:
scriptDir: "/model_train/yolov3_algorithms/"
scriptBootFile: "eval.py"
frameworkType: "tensorflow"
frameworkVersion: "1.18"
deploySpec:
model:
name: "deploy-model"
trigger:
condition:
operator: ">"
threshold: 0.1
metric: precision_delta
nodeName: edge1
outputDir: "/helmet-detection/"
Controller Design
The incremental learning controller starts three separate goroutines called upstream, downstream and incrementallearningjobcontroller.
These are not separate controllers as such but named here for clarity.
- incremental learning: watch the updates of incremental-learning job crds, and create the workers depending on the state machine.
- downstream: synchronize the incremental-learning-job updates from the cloud to the edge node.
- upstream: synchronize the incremental-learning-job updates from the edge to the cloud node.
Incremental Learning Controller
The incremental-learning controller watches for the updates of incremental-learning jobs and the corresponding pods against the K8S API server.
Updates are categorized below along with the possible actions:
| Update Type | Action |
|---|---|
| New Incremental-learning-job Created | Wait to train trigger satisfied |
| Incremental-learning-job Deleted | NA. These workers will be deleted by k8s gc. |
| The Status of Incremental-learning-job Updated | Create the train/eval worker if it's ready. |
| The corresponding pod created/running/completed/failed | Update the status of incremental-learning job. |
Downstream Controller
The downstream controller watches for the incremental-learning job updates against the K8S API server.
Updates are categorized below along with the possible actions that the downstream controller can take:
| Update Type | Action |
|---|---|
| New Incremental-learning-job Created | Sends the job information to LCs. |
| Incremental-learning-job Deleted | The controller sends the delete event to LCs. |
Upstream Controller
The upstream controller watches for the incremental-learning job updates from the edge node and applies these updates against the API server in the cloud.
Updates are categorized below along with the possible actions that the upstream controller can take:
| Update Type | Action |
|---|---|
| Incremental-learning-job Reported State Updated | The controller appends the reported status of the job by LC in the cloud. |
Details of api between GC(cloud) and LC(edge)
GC(downstream controller) syncs the job info to LC:
// POST <namespace>/incrementallearningjobs/<job-name> // body same to the job crd of k8s api, omitted here.LC uploads the job status which reported by the worker to GC(upstream controller):
// POST <namespace>/incrementallearningjobs/<job-name>/status // WorkerMessage defines the message from that the training worker. It will send to GC. type WorkerMessage struct { Phase string `json:"phase"` Status string `json:"status"` Output *WorkerOutput `json:"output"` } // type WorkerOutput struct { Models []*Model `json:"models"` TaskInfo *TaskInfo `json:"taskInfo"` } // Model defines the model information type Model struct { Format string `json:"format"` URL string `json:"url"` // Including the metrics, e.g. precision/recall Metrics map[string]float64 `json:"metrics"` } // TaskInfo defines the task information type TaskInfo struct { // Current training round CurrentRound int `json:"currentRound"` UpdateTime string `json:"updateTime"` }
The flows of incremental learning job
- Flow of the
trainstage
- Flow of the
evalstage
- Flow of the
deploystage
Workers Communication
No need to communicate between workers.