Build your own workflow

Steps

1. 1

   Step 1: Start from a blank canvas

   On the Workflow page, with your project open, open Templates, choose Blank workflow and select Start blank. Your task and data references stay; Undo brings the previous modules back.

   The other templates are “Single step (no loop)” (Propose → Implement → Validate → Train & tune once), “Single loop with your own advice” (the research loop Quick start opens; Propose model reads your advice) and “Single research loop” (the same loop without Propose model's advice input connected, so its advice field shows a warning). Starting from one of them is often quicker.

   You should see: an empty canvas, with the module list on the left.

2. 2

   Step 2: Add modules

   The list on the left is in running order. Propose model is marked Start here; Interpret results and Literature review are marked Optional context: they feed extra information into the main chain. Expand a module to see what it receives, what it sends and what it connects to. Hover over one of its routes for a preview; click, press Enter or drag it onto the canvas to add it.

   Add Propose model, Implement, Validate and Train & tune. In a project bound to a task package, inputs the run prepares for you (such as Storage folder and Training settings) are connected automatically.

   You should see: four cards on the canvas: Propose model, Implement, Validate, Train & tune.

3. 3

   Step 3: Connect the modules

   Select an output (or drag from it); the inputs that accept it light up; select one to connect. The + beside an output lists modules that accept it; choosing one adds it to the right, already connected.

   A refused connection names both types, for example “Can't connect: Propose model sends Model plan; this input receives Design brief.” Socket colour and shape show the type; Connector types at the bottom left of the canvas is the legend. Optional inputs you have not connected fold into a “+N optional” row.

   Train & tune needs two wires: the Code verdict from Validate and the Model plan from Propose model. Each socket is described on its reference page.

   You should see: Propose model's Model plan feeds Implement, Implement's Model code feeds Validate, and so on.

4. 4

   Step 4: Check the workflow

   Select Check the workflow on the canvas. A token travels the wires in running order, then the research loop's feedback wire if there is one. At the first break it stops: that socket gets a soft red halo and one sentence, for example “Train & tune can't start: Code verdict is not connected.” Connect it and check again.

   This check reads only the wires on the canvas. Save runs the full check, which is the one that counts.

   You should see: a token travels every wire to the end with no red halo.

5. 5

   Step 5: (Optional) Add a research loop

   A loop needs a Propose model that reads Results analysis. The Propose model from step 2 uses the Start here route, which has no such input, and an existing module's route cannot be changed. The quickest way is the “Single loop with your own advice” template (the one Quick start opens). To build it by hand, on a blank canvas add Interpret results first, then use the + beside its Results analysis output and choose Propose model: it is added with its “from Interpret results” route, already connected. Add and connect Implement, Validate and Train & tune as in steps 2 and 3. Then select Add research loop in the Research loop panel. It needs exactly one Train & tune and one Interpret results on the canvas. Rounds and stop conditions are set as in tutorial 2.

   You should see: a feedback wire from Train & tune back to Interpret results.

6. 6

   Step 6: Save a named version

   Select Save, read the exact Python changes, enter a Version name (for example “cnn-baseline-v1”) and select Save version. If the file changed on disk while you were editing, the page says This workflow changed on disk; choose Reapply my changes onto latest to keep your edits.

   You should see: the version appears in Version history and Next: Run is available.