Migrating to TAO 6.0#

TAO 6.0 introduces several breaking changes. This guide describes the changes and how to migrate your existing workflows.

Fine-Tuning Microservices OpenAPI Specifications#

Here is a summary of the key changes from TAO 5.5 OpenAPI specification:

  • Unified Resource Scoping with org_name instead of user_id: The most significant change is the shift from using user_id directly in the path to using org_name for scoping resources.

    Endpoints that previously started with /api/v1/users/{user_id}/... for datasets and experiments now follow the pattern /api/v1/orgs/{org_name}/....

    This change is applied consistently across Dataset, Experiment, and now also explicitly for Workspace related endpoints in the new specification.

  • Authentication request update:

    • The login request (POST /api/v1/login) in the latest API now expects ngc_key and ngc_org_name in the request body, whereas the previous API schema expected ngc_api_key. It also allows for an optional enable_telemetry field in the login request.

    • NGC personal keys are required for authentication instead of NGC API keys.

  • New organization-level endpoints: The new API introduces top-level endpoints under /api/v1/orgs/{org_name}/ that were not present in the old specification for supporting NVCF helm chart deployment, such as:

    • POST /api/v1/orgs/{org_name}/super_endpoint

    • GET /api/v1/orgs/{org_name}:gpu_types

  • Workspace endpoint standardization: In the previous API version, the concept of workspaces was handled directly within the dataset and experiment create endpoints, now the API explicitly defines a full suite of CRUD and action endpoints for workspaces under the /api/v1/orgs/{org_name}/workspaces/ path.

  • Parameter and schema refinements:

    • ExperimentRsp includes new fields like model_description, tags, tensorboard_enabled, and base_experiment_metadata. The automl_settings within ExperimentRsp in the new API has automl_hyperparameters instead of automl_add_hyperparameters and automl_remove_hyperparameters found in the old API’s AutoML schema.

    • The DatasetRsp in the new API includes fields like workspace, cloud_file_path, use_for, skip_validation, and authorized_party_nca_id which were not present or structured the same way in the old version.

    • Job actions and statuses have some variations in their enum values between the two versions. For instance, JobSubset and DatasetJob include statuses like Canceling, Pausing, Paused, and Resuming, which are more granular than in the old specification.

    • The structure for listing jobs (e.g., DatasetJobList, ExperimentJobList) in the new API includes pagination_info, which was not explicitly part of the list responses in the old API.

Deprecation Notes#

The following features were deprecated in TAO 5.5, and have been removed in TAO 6.0.

Software Deprecations#

The table below shows the model fine-tuning workflows in TAO 6.0 that replace the deprecated models in TAO 5.5.

Use Case

TAO 5.5

TAO 6.0

Object Detection

  • DetectNet_v2

  • FasterRCNN

  • YOLOv3

  • YOLOv4

  • RetinaNet

  • SSD

  • DSSD

Semantic Segmentation

  • Unet

Instance Segmentation

  • MaskRCNN

The following features are deprecated in TAO 6.0, and will be removed in a future release.

Pretrained models#

The following table shows the pretrained models that are deprecated in TAO 6.0, with the equivalent models that are being maintained going forward.

Deprecated Pretrained Model

Equivalent model in TAO 6.0

TAO Toolkit ODISE 1.1

Mask Grounding DINO
On this page