https://github.com/kubeflow/katib
Tip revision: ed98e17b5faea57b4ad43472159b4b00187f2851 authored by avelichk on 03 September 2020, 20:57:48 UTC
Update doc
Update doc
Tip revision: ed98e17
workflow-design.md
# How Katib v1beta1 tunes hyperparameter automatically in a Kubernetes native way
See the following guides in the Kubeflow documentation:
* [Concepts](https://www.kubeflow.org/docs/components/hyperparameter-tuning/overview/)
in Katib, hyperparameter tuning, and neural architecture search.
* [Getting started with Katib](https://kubeflow.org/docs/components/hyperparameter-tuning/hyperparameter/).
* Detailed guide to [configuring and running a Katib
experiment](https://kubeflow.org/docs/components/hyperparameter-tuning/experiment/).
## Example and Illustration
After install Katib v1beta1, you can run `kubectl apply -f katib/examples/v1beta1/random-example.yaml` to try the first example of Katib.
Then you can get the new `Experiment` as below. Katib concepts will be introduced based on this example.
```yaml
$ kubectl get experiment random-example -n kubeflow -o yaml
apiVersion: kubeflow.org/v1beta1
kind: Experiment
metadata:
...
name: random-example
namespace: kubeflow
...
spec:
algorithm:
algorithmName: random
maxFailedTrialCount: 3
maxTrialCount: 12
metricsCollectorSpec:
collector:
kind: StdOut
objective:
additionalMetricNames:
- Train-accuracy
goal: 0.99
metricStrategies:
- name: Validation-accuracy
value: max
- name: Train-accuracy
value: max
objectiveMetricName: Validation-accuracy
type: maximize
parallelTrialCount: 3
parameters:
- feasibleSpace:
max: "0.03"
min: "0.01"
name: lr
parameterType: double
- feasibleSpace:
max: "5"
min: "2"
name: num-layers
parameterType: int
- feasibleSpace:
list:
- sgd
- adam
- ftrl
name: optimizer
parameterType: categorical
resumePolicy: LongRunning
trialTemplate:
trialParameters:
- description: Learning rate for the training model
name: learningRate
reference: lr
- description: Number of training model layers
name: numberLayers
reference: num-layers
- description: Training model optimizer (sdg, adam or ftrl)
name: optimizer
reference: optimizer
trialSpec:
apiVersion: batch/v1
kind: Job
spec:
template:
spec:
containers:
- command:
- python3
- /opt/mxnet-mnist/mnist.py
- --batch-size=64
- --lr=${trialParameters.learningRate}
- --num-layers=${trialParameters.numberLayers}
- --optimizer=${trialParameters.optimizer}
image: docker.io/kubeflowkatib/mxnet-mnist
name: training-container
restartPolicy: Never
status:
...
```
#### Experiment
When you want to tune hyperparameters for your machine learning model before
training it further, you just need to create an `Experiment` CR like above. To
learn what fields are included in the `Experiment.spec`, see
the detailed guide to [configuring and running a Katib
experiment](https://kubeflow.org/docs/components/hyperparameter-tuning/experiment/).
#### Trial
For each set of hyperparameters, Katib will internally generate a `Trial` CR with the hyperparameters key-value pairs, job manifest string with parameters instantiated and some other fields like below. `Trial` CR is used for internal logic control, and end user can even ignore it.
```yaml
$ kubectl get trial -n kubeflow
NAME TYPE STATUS AGE
random-example-58tbx6xc Succeeded True 14m
random-example-5nkb2gz2 Succeeded True 21m
random-example-88bdbkzr Succeeded True 20m
random-example-9tgjl9nt Succeeded True 17m
random-example-dqzjb2r9 Succeeded True 19m
random-example-gjfdgxxn Succeeded True 20m
random-example-nhrx8tb8 Succeeded True 15m
random-example-nkv76z8z Succeeded True 18m
random-example-pcnmzl76 Succeeded True 21m
random-example-spmk57dw Succeeded True 14m
random-example-tvxz667x Succeeded True 16m
random-example-xpw8wnjc Succeeded True 21m
$ kubectl get trial random-example-gjfdgxxn -o yaml -n kubeflow
apiVersion: kubeflow.org/v1beta1
kind: Trial
metadata:
...
name: random-example-gjfdgxxn
namespace: kubeflow
ownerReferences:
- apiVersion: kubeflow.org/v1beta1
blockOwnerDeletion: true
controller: true
kind: Experiment
name: random-example
uid: 34349cb7-c6af-11ea-90dd-42010a9a0020
...
spec:
metricsCollector:
collector:
kind: StdOut
objective:
additionalMetricNames:
- Train-accuracy
goal: 0.99
metricStrategies:
- name: Validation-accuracy
value: max
- name: Train-accuracy
value: max
objectiveMetricName: Validation-accuracy
type: maximize
parameterAssignments:
- name: lr
value: "0.012171302435678337"
- name: num-layers
value: "3"
- name: optimizer
value: adam
runSpec:
apiVersion: batch/v1
kind: Job
metadata:
name: random-example-gjfdgxxn
namespace: kubeflow
spec:
template:
spec:
containers:
- command:
- python3
- /opt/mxnet-mnist/mnist.py
- --batch-size=64
- --lr=0.012171302435678337
- --num-layers=3
- --optimizer=adam
image: docker.io/kubeflowkatib/mxnet-mnist
name: training-container
restartPolicy: Never
status:
completionTime: "2020-07-15T15:29:00Z"
conditions:
- lastTransitionTime: "2020-07-15T15:25:16Z"
lastUpdateTime: "2020-07-15T15:25:16Z"
message: Trial is created
reason: TrialCreated
status: "True"
type: Created
- lastTransitionTime: "2020-07-15T15:29:00Z"
lastUpdateTime: "2020-07-15T15:29:00Z"
message: Trial is running
reason: TrialRunning
status: "False"
type: Running
- lastTransitionTime: "2020-07-15T15:29:00Z"
lastUpdateTime: "2020-07-15T15:29:00Z"
message: Trial has succeeded
reason: TrialSucceeded
status: "True"
type: Succeeded
observation:
metrics:
- latest: "0.959594"
max: "0.960490"
min: "0.940585"
name: Validation-accuracy
- latest: "0.959022"
max: "0.959188"
min: "0.921658"
name: Train-accuracy
startTime: "2020-07-15T15:25:16Z"
```
#### Suggestion
Katib will internally create a `Suggestion` CR for each `Experiment` CR. `Suggestion` CR includes the hyperparameter algorithm name by `algorithmName` field and how many sets of hyperparameter Katib asks to be generated by `requests` field. The CR also traces all already generated sets of hyperparameter in `status.suggestions`. Same as `Trial`, `Suggestion` CR is used for internal logic control and end user can even ignore it.
```yaml
$ kubectl get suggestion random-example -n kubeflow -o yaml
apiVersion: kubeflow.org/v1beta1
kind: Suggestion
metadata:
...
name: random-example
namespace: kubeflow
ownerReferences:
- apiVersion: kubeflow.org/v1beta1
blockOwnerDeletion: true
controller: true
kind: Experiment
name: random-example
uid: 34349cb7-c6af-11ea-90dd-42010a9a0020
...
spec:
algorithmName: random
requests: 12
status:
suggestionCount: 12
suggestions:
...
- name: random-example-gjfdgxxn
parameterAssignments:
- name: lr
value: "0.012171302435678337"
- name: num-layers
value: "3"
- name: optimizer
value: adam
- name: random-example-88bdbkzr
parameterAssignments:
- name: lr
value: "0.013408352284328112"
- name: num-layers
value: "4"
- name: optimizer
value: ftrl
- name: random-example-dqzjb2r9
parameterAssignments:
- name: lr
value: "0.028873905258692753"
- name: num-layers
value: "3"
- name: optimizer
value: adam
...
```
## What happens after an `Experiment` CR created
When a user created an `Experiment` CR, Katib controllers including experiment controller, trial controller and suggestion controller will work together to achieve hyperparameters tuning for user Machine learning model.
<center>
<img width="100%" alt="image" src="images/katib-workflow.png">
</center>
1. A `Experiment` CR is submitted to Kubernetes API server, Katib experiment mutating and validating webhook will be called to set default value for the `Experiment` CR and validate the CR separately.
2. Experiment controller creates a `Suggestion` CR.
3. Suggestion controller creates the algorithm deployment and service based on the new `Suggestion` CR.
4. When Suggestion controller verifies that the algorithm service is ready, it calls the service to generate `spec.request - len(status.suggestions)` sets of hyperparamters and append them into `status.suggestions`
5. Experiment controller finds that `Suggestion` CR had been updated, then generate each `Trial` for each new hyperparamters set.
6. Trial controller generates job based on `trialSpec` manifest with the new hyperparamters set.
7. Related job controller (Kubernetes batch Job, Kubeflow PyTorchJob or Kubeflow TFJob) generates Pods.
8. Katib Pod mutating webhook is called to inject metrics collector sidecar container to the candidate Pod.
9. During the ML model container runs, metrics collector container in the same Pod tries to collect metrics from it and persists them into Katib DB backend.
10. When the ML model Job ends, Trial controller will update status of the corresponding `Trial` CR.
11. When a `Trial` CR goes to end, Experiment controller will increase `request` field of corresponding
`Suggestion` CR if it is needed, then everything goes to `step 4` again. Of course, if `Trial` CRs meet one of `end` condition (exceeds `maxTrialCount`, `maxFailedTrialCount` or `goal`), Experiment controller will take everything done.