Polyaxon Optimization Engine - Iterative Optimization Specification

Iterative

This is part of our commercial offering.

V1Iterative

polyaxon._flow.matrix.iterative.V1Iterative()

To build a custom optimization algorithm, this interface lets you create an iterative process for creating suggestions and training your model based on those suggestions

The iterative process expect a user defined a tuner that will generate the suggestions for running the component.

Args:
- kind: str, should be equal iterative
- max_iterations: int
- params: List[Dict[str, params]]
- concurrency: int, optional
- seed: int, optional
- tuner: V1Tuner, optional
- early_stopping: List[EarlyStopping], optional

YAML usage

matrix:
  kind: iterative
  concurrency:
  params:
  maxIterations:
  seed:
  tuner:
  earlyStopping:

Python usage

from polyaxon.schemas import (
    V1Iterative,
    V1HpLogSpace,
    V1HpUniform,
    V1FailureEarlyStopping,
    V1MetricEarlyStopping,
    V1Tuner,
)
matrix = V1Iterative(
  max_iterations=20,
  concurrency=2,
  seed=23,
  params={"param1": V1HpLogSpace(...), "param2": V1HpUniform(...), ... },
  early_stopping=[V1FailureEarlyStopping(...), V1MetricEarlyStopping(...)],
  tuner=V1Tuner(hub_ref="org/my-suggestion-component")
)

Fields

kind

The kind signals to the CLI, client, and other tools that this matrix is an iterative process.

If you are using the python client to create the mapping, this field is not required and is set by default.

matrix:
  kind: iterative

concurrency

An optional value to set the number of concurrent operations.

This value only makes sense if less or equal to the total number of possible runs.

matrix:
  kind: iterative
  concurrency: 2

For more details about concurrency management, please check the concurrency section.

params

A dictionary of key -> value generator to generate the parameters.

To learn about all possible params generators.

The parameters generated will be validated against the component’s inputs/outputs definition to check that the values can be passed and have valid types.

matrix:
  kind: iterative
  params:
    param1:
       kind: ...
       value: ...
    param2:
       kind: ...
       value: ...

maxIterations

Maximum number of iterations to run the process of -> suggestions -> training ->\

matrix:
  kind: iterative
  maxIterations: 5

seed

Since this algorithm uses random generators, if you want to control the seed for the random generator, you can pass a seed.

matrix:
 kind: iterative
 seed: 523

earlyStopping

A list of early stopping conditions to check for terminating all operations managed by the pipeline. If one of the early stopping conditions is met, a signal will be sent to terminate all running and pending operations.

matrix:
  kind: iterative
  earlyStopping: ...

For more details please check the early stopping section.

tuner

The tuner reference definition (with a component hub reference) to use. The component contains the logic for creating new suggestions.

matrix:
  kind: iterative
  tuner:
    hubRef: acme/suggestion-logic:v1

Example

In this example the iterative process will try run 5 iterations generating new experiments based on the search space defined in the params subsection.

version: 1.1
kind: operation
matrix:
  kind: iterative
  concurrency: 10
  maxIterations: 5
  tuner:
    hubRef: my-suggestion-component
  params:
    lr:
      kind: logspace
      value: 0.01:0.1:5
    dropout:
      kind: choice
      value: [0.2, 0.5]
   activation:
      kind: pchoice
      value: [[elu, 0.1], [relu, 0.2], [sigmoid, 0.7]]
   early_stopping:
     - metric: accuracy
       value: 0.9
       optimization: maximize
     - metric: loss
       value: 0.05
       optimization: minimize
component:
  inputs:
    - name: batch_size
      type: int
      isOptional: true
      value: 128
    - name: lr
      type: float
    - name: dropout
      type: float
    - name: activation
      type: str
  container:
    image: image:latest
    command: [python3, train.py]
    args: [
        "--batch-size={{ batch_size }}",
        "--lr={{ lr }}",
        "--dropout={{ dropout }}",
        "--activation={{ activation }}"
    ]

Automation Iterative