ML Research Pipelines

What you'll build

A CIFAR-10 image classification pipeline:

• Dataset node loading CIFAR-10.
• Two Logic augmentation nodes (random crop + horizontal flip).
• A Model node training a small CNN.
• An evaluation node reporting top-1 accuracy.
• Two variants — Baseline and No augmentation — for comparison.

Estimated time: 10–15 minutes.

Step-by-step

1. Create a new project

   From the Graphs Hub, click New graph. Name it "CIFAR-10 ablation". You'll land in the editor with one empty variant ("v1") and a blank canvas.

2. Add the Dataset node

   Click + Node → Dataset. In the Inspector, set source to cifar-10 and split to train. Its single output port produces image data with a label sidecar.

3. Add augmentation

   Add two Logic nodes — random_crop (padding=4) and horizontal_flip (p=0.5). Wire Dataset → random_crop → horizontal_flip. Both nodes pass image through unchanged.

4. Add the Model node

   Click + Node → Model. Choose small_cnn from the model library. The model has two inputs (image, label) and one output (tensor for predictions). Wire augmented images and labels in.

5. Configure training

   Click the Model node. In the Config tab set epochs to 10, lr to 1e-3, batch_size to 128. Leave the optimizer at the default (AdamW).

6. Add evaluation

   Add a Logic → eval/top1_accuracy node. Wire the model's predictions and the labels in. Its output is a float — the metric you care about.

7. Run the baseline

   Click Run. Watch edges animate as data flows. When the model finishes training, click the eval node and check Run Data — you should see something around 78–82% accuracy on a freshly initialized CNN.

8. Create the ablation variant

   Click + Variant. Name it "No augmentation". In the new variant, delete both augmentation nodes and wire Dataset → Model directly.

9. Set the baseline and compare

   Right-click the "v1" tab → Set as baseline. Switch to "No augmentation", click Run. When it finishes, open Compare → v1 (Baseline). You'll see the structural diff and the accuracy delta side by side.

10. Trace the result

    On the baseline, switch to the Path tool and click the accuracy output. Save the path as "Top-1 → augmented input". Anyone opening the project later can click that chip to see exactly what produced the metric.