Build Your First Pipeline
Chain read, split, train, evaluate, a quality gate and a package into one repeatable pipeline, publish it, run it with different parameters and compare the runs.
What you will build
A pipeline that retrains a model, evaluates it on held-out rows, approves it only if it clears a threshold, and packages it. You can run it again with other settings and compare the results.
Before you start
- A labelled dataset in the Data Hub
- A model graph saved in the Model Builder for that dataset (see From Template to Trained Model)
Step 1: Start from a template
Open Pipelines. Choose Templates and pick Train, check and package a model. Choose your dataset and model graph. A draft is created with six steps: Read Dataset, Split, Train Model, Evaluate, Quality Gate and Build Package. Open it.
Step 2: Understand the graph
Switch between the canvas and the outline. Each wire joins an output to a matching input, and the editor will not connect incompatible ports. Click Train Model to see its settings in the panel beside the canvas.
Step 3: Set the gate
Click Quality Gate. The default asks for accuracy of at least 0.9. Change the threshold to something your data can reach, for example 0.8. A gate can also compare against a baseline evaluation instead of a fixed number.
Step 4: Add a parameter
Open Parameters and add a number called min_accuracy with a default of 0.8. Then, in the Quality Gate's settings, use the parameter binding control to take the threshold from min_accuracy. Parameters change per run without editing the pipeline.
Step 5: Check before running
The estimate panel lists which datasets will be read (version, rows, size) and which steps use TensorFlow.js or run your own code. It never guesses times.
Step 6: Publish and run
Press Publish & Run. Publishing makes an immutable, numbered revision, and the run uses it. Watch the timeline: each step starts and succeeds in turn. If the gate rejects the model, the run ends gated and the package is withheld. That is the pipeline working as designed, not a failure.
Step 7: Read the results
Open the run details. Evaluate shows accuracy and the other metrics; the model and the package it produced can be opened in their own modules. If packaging happened, open Deployment → Pipeline packages to download the zip.
Step 8: Run again and compare
Run again with min_accuracy set to 0.95. Then select two runs and choose Compare. You see each run's parameters, metrics and gate verdict side by side, with the better value of accuracy or loss marked.
Step 9: Sweep
Open Sweep and give min_accuracy several values. The pipeline runs once per value, one after another (up to 12), and the runs land in the same history.
Step 10: Safeguards
Open Revision history to see published revisions and a diff between them. If you try to switch projects while a run is in progress, the Studio asks first.
What you learned
Pipelines turn a one-off notebook run into something repeatable: published revisions, parameters, gates and a recorded history.