Skip to content
This repository was archived by the owner on Mar 24, 2026. It is now read-only.

README.md

Latest commit

 

History

History
45 lines (37 loc) · 4.06 KB

README.md

File metadata and controls

45 lines (37 loc) · 4.06 KB
Raw

Task Builder

Task builder consists of Python utilities to build tasks based on a TFF model and task configuration.

A task encapulates instructions on how to train or evaluate the model, including training artifacts such as:

  • a server-side training plan Plan
  • a device-side training plan ClientOnlyPlan
  • Initial model weights checkpoint in bytes

Task Builder Service

The task builder service provides two APIs for different use cases:

Build Task Group API

  • URL: /taskbuilder/v1:build-task-group
  • Method: POST
  • Description: Creates tasks and artifacts based on SavedModel and TaskConfig input in BuildTaskRequest.
  • Request: The request body should contain binary data in protobuf format (application/x-protobuf) of BuildTaskRequest
  • Response: The API returns responses in protobuf format (application/x-protobuf) of BuildTaskResponse

Build Artifacts API

  • URL: /taskbuilder/v1:build-artifacts
  • Method: POST
  • Description: Creates artifacts only without tasks.
  • Request and response align with data format of :build-task-group.
  • ArtifactBuilding has to be set in TaskConfig of the request, which point to GCS paths where created artifacts should to uploaded to.

Start a local Task Builder Service for Test

bazel run //python/taskbuilder:task_builder

Task Builder Client

For early testing, we provide a simple client that interacts with task builder services.

Workflow 1: create tasks and artifacts together

Run the Python client: bazel run //python/taskbuilder:task_builder_client -- --saved_model=gs://<resource_uri> --task_config=gs://<resource_uri> --task_builder_server=https://<task_builder_server>

  • --saved_model: a GCS path to a TensorFlow SavedModel. You can find a sample here
  • --task_config: a GCS path to a TaskConfig in text protobuf. You can find a sample here
  • --task_builder_server: the task builder server endpoint. Local host (http://localhost:5000) will be used as default if not provided.
  • --population_name: an updated population name for the task. If provided, the original population_name in TaskConfig will be overriden. This is helpful if you want to reuse the same --task_config but with a different population name for testing purpose.
  • --impersonate_service_account: a service account to impersonate which has permission to access the task builder server endpoint.
  • --api_key: an API key which has permission to access the task builder API.

Workflow 2: create artifacts only

Run the Python client: bazel run //python/taskbuilder:task_builder_client -- --saved_model=gs://<resource_uri> --task_config=gs://<resource_uri> --task_builder_server=https://<task_builder_server> --build_artifact_only=true

  • --build_artifact_only: the option to skip task creation and build artifacts only. You can find a sample here
  • --skip_flex_ops_check: the option to skip flex ops check in Android TensorflowLite library.
  • All other options are same as above. The generated task will uploaded to the GCS path you specified here. You can find some output samples here