Validation | Dagnam.AI Docs

When Studio checks your architecture, how to read the status bar, and how to fix the most common problems.

Validation Studio checks your architecture continuously while you build it, and again on the server before it lets you generate code or train. This page covers when validation runs, how to read its results, what it checks, and how to fix the messages you will see most often. When validation runs Validation runs on every edit, after a short pause so it does not fire on every keystroke or drag. The server checks the design again when you click Generate , and when you click Start Training in the Train dialog, so nothing is generated or trained from a design the server rejects. Reading the status bar The bottom status bar reads "Pipeline is valid" or "Pipeline is invalid", with counts for nodes, connections, errors, warnings, and info items. Clicking it expands into Errors, Warnings, and Info lists; clicking an error or warning that names a node highlights that node. Edge errors A connection with a problem is drawn red and dashed with a "!" badge; its tooltip explains the problem. A valid connection is green. What gets checked Validation covers, as separate categories: the graph's structure missing connections, cycles , merges, shape and type compatibility between connected nodes, the output objective, sequence length and vocabulary consistency, the dataset source, dataset versus model warnings, blocks, parameter ranges, and general tips. Parameter range messages name the node and the field, for example "Learning Rate must be at most 1 currently 2 ". Common fixes Message pattern Fix ------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------ "Node '<name>' has no input connections" Connect it to an upstream node, or delete it if you do not need it. "'<name>' input '<handle>' has N incoming connections" Add a Merge node, or remove the duplicate connections. "'<A>' outputs 4D data but '<B>' requires 1D input" Add a Flatten or a Global Pooling layer between them. "'<A>' outputs N units but Output expects M" Set the last Dense layer's units to match, or change Output's Number of Classes. "Softmax final activation is used but loss is '<loss>'" Use CrossEntropyLoss for multi-class classification or BCELoss/BCEWithLogitsLoss with a Sigmoid final activation for binary . "Dataset vocab size ... doesn't match Embedding vocab size ... " Set the Dataset node's Vocabulary Size and the Embedding Layer's Vocabulary Size to the same value. "Dataset node is not configured; select a dataset" Drag a dataset card from the Data tab onto the canvas, connect it in place of the empty Dataset node, then delete the empty one. Why Generate and Train are locked Generate and Train appear once your design is saved as a project, and are enabled only while it is valid with no unsaved changes. Hover a disabled button to see why; for Train it reads "Fix validation errors before training" or "Save the model before training". An invalid design still saves; it just cannot generate code or train until you fix it. When a field in the node configuration panel holds an invalid value, the panel says "Please fix the validation errors before continuing." and that value is not applied until you correct it. See Save a version /docs/studio/versions-and-sharing save-a-version and Before you generate /docs/studio/generate-code before-you-generate .