Review pipeline resources and training configuration

Understand the generated plan#

The pipeline detail page separates planning context from executable resources. During creation, planning uses the dataset and instruction to select a task-compatible architecture; review that planned resource after planning completes. Resource availability varies by data type and task, so a pipeline may not show every resource listed below.

• Architecture defines the model family and task-compatible output.
• Transforms prepare training and evaluation inputs without changing the original dataset.
• Optimizer and Loss control how supported trainable models learn.
• Metrics define the measurements calculated during training or evaluation.
• Training Configuration contains settings such as epochs, batch size, input handling, and other task-compatible options.

Review and edit resources#

1. Read the plan and task cards

   Confirm that the generated task, target, modality, and evaluation goal match your instruction.

2. Inspect the planned architecture

   Open the architecture resource selected during planning and review its fields. Use Knowledge Base after planning only when choosing a replacement compatible with the task, inputs, and labels.

3. Review transforms

   Check training and evaluation transforms for input size, normalization, augmentation, and any modality-specific processing.

4. Review optimizer and loss

   For trainable architectures, confirm that the optimization and loss choices match the intended task and label structure.

5. Review metrics

   Keep metrics that answer the evaluation question. Add a metric resource only when it is supported by the task and is needed for a decision.

6. Review Training Configuration

   Check the initial batch size, epochs, and other exposed settings. Use a conservative batch size when input resolution or model size is high.

7. Save and reread the plan

   Save each valid change and confirm the resource cards show the intended values before running.

Add or replace a resource safely#

1. Choose the correct resource action

   Edit an existing resource for a small supported change. Use Add Resource for an additional compatible metric or resource type. Use the architecture swap control for a different model family.

2. Keep fields internally consistent

   Nested fields can depend on one another. Add or remove only fields understood by the selected resource.

3. Recheck downstream resources

   After changing architecture or input handling, verify transforms, loss, metrics, and training configuration again.

Update an idle pipeline#

1. Open Update

   From the pipeline detail page, choose Update only when no planning, preparation, training, evaluation, or recovery action is active.

2. Review editable fields

   Confirm the attached dataset, pipeline name, available backend profile, and image explainability choice before changing anything.

3. Change instructions only before the first run

   Revise the training instruction while the pipeline remains not started. After execution begins, use a clone or new lineage when the prediction goal must change.

4. Regenerate when the plan must be rebuilt

   While the pipeline is still not started, choose Regenerate to rebuild the generated plan from the updated definition, then reread every resource.

5. Save with Update

   Submit once and confirm the detail page reflects the new configuration before selecting Run.